Release Notes

Latest version: v0.7.1 (2026-08-09).

v0.7.1 (2026-08-09)

Highlights

New capabilities: serve docling OCR with the English recogniser
Reliability and operations: meter executed GLiNER token window

Features

server: serve docling OCR with the English recogniser

Bug Fixes

server: meter executed GLiNER token window

v0.7.0 (2026-08-08)

Highlights

Breaking change: aws-control-plane no longer accepts component_contract_sha256 or injects its environment marker.
New capabilities: add fail-closed estate-bootstrap saved-plan inspection; add frozen assembly apply seam; add minimal promotion eligibility guard; admit estates by declared capability and add the sandbox desired state; allow benchmark campaign self-review; assert apply-free and same-scope second convergence runs
Reliability and operations: align catalog smoke timeout budget; apply the checkpoint whitelist to record, harden the git guard; bind the published price table to the active billing authority; harden document field fixture validation; harden generation response handling

⚠ BREAKING CHANGES

terraform: aws-control-plane no longer accepts component_contract_sha256 or injects its environment marker.

Features

cloud: add fail-closed estate-bootstrap saved-plan inspection
cloud: add frozen assembly apply seam
cloud: add minimal promotion eligibility guard
cloud: admit estates by declared capability and add the sandbox desired state
cloud: allow benchmark campaign self-review
cloud: assert apply-free and same-scope second convergence runs
cloud: batch the priced rate-grade cells behind one approval
cloud: bill audio_ms at a 1,150 ms minimum duration
cloud: bind the inspection verdict to the exact plan digest
cloud: declare the terraform backend kind per estate cell
cloud: deepen the read-only estate bootstrap preflight
cloud: default tier points to min(floor x 2.0, ceiling x 0.85)
cloud: freeze infrastructure values in assemblies
cloud: hydrate terraform outputs for remote-backend cells
cloud: implement the #2441 pricing-policy decisions (PR #3024)
cloud: local-only session auth for the L3 verification stack
cloud: make the sandbox cell serving
cloud: permanently allow benchmark campaign self-review
cloud: persist deliverable metadata in s3
cloud: publish the live rate book to a public price artifact
cloud: record and verify estate bootstrap approvals
cloud: record convergence evidence and compare two runs
cloud: warn loudly when a —live preflight run proves nothing
control-plane: serve-image capability manifest and operator-opened tasks
model-ops: add the docling English OCR artifact publish tool

Bug Fixes

ci: refuse an unusable assets_path before the gate
ci: replace the batch ledger atomically and unwrap the summary prose
clients: address terminal response review
cloud: accept padded whisper transcripts
cloud: accept real names, escape paths, close the commondir fail-open
cloud: address deployment review findings
cloud: admit provisional task measurement pairs
cloud: admit same-apply IAM unknowns by provenance, not by fiat
cloud: align catalog canary model contracts
cloud: align catalog smoke timeout budget
cloud: align document field canary semantics
cloud: align phase 3 exit criteria and make skip tests prove their cause
cloud: align PII acceptance taxonomy
cloud: allow task candidates across routes
cloud: allowlist read-only init by its full argument set
cloud: apply the checkpoint whitelist to record, harden the git guard
cloud: assert the full run built something and order the pair
cloud: bind every resource to the verified provider, not only data sources
cloud: bind read-only terraform access to the resolved cell’s root
cloud: bind the published price table to the active billing authority
cloud: bound Modal image build cache
cloud: bound provenance to policy CONTENT, not just mentioned resources
cloud: close checkpoint leak, env-independent repo guard, strict schema
cloud: close fail-open branch chains in the estate preflight
cloud: close fail-open paths in estate plan inspection
cloud: close preflight fail-opens found by the mandated self-review
cloud: close the fail-open paths found in review threads
cloud: close the fail-opens the third self-review found
cloud: close the holes the fourth self-review found in this round’s own permissive paths
cloud: close the plan-inspection bypasses from the adversarial review
cloud: close the two backend-classifier evasions from verification
cloud: commit job results before settlement
cloud: confine catalog subsets to sandbox
cloud: constrain checkpoint field values, not just field names
cloud: constrain screenshot bbox axes
cloud: correct the apply-check rationale and scope-check claim
cloud: correct the offload-smoke minted tuple annotation and notice an ignored —account
cloud: correct the provider-key rule against ground truth, never echo after values
cloud: declare the revoke 404 contract and pin its no-oracle body
cloud: decouple availability from measurements
cloud: defer gateway replica probe until serve
cloud: derive the declared commercial margin from the price multiplier
cloud: enforce the D1 boundary on the terraform primitive
cloud: escape the last raw path, refuse an unreadable HEAD
cloud: fail closed on an unreadable desired-state spec
cloud: freeze deployment check results
cloud: fullmatch the rate-book version before it becomes a filename
cloud: give every checkpoint field a format, not a length cap
cloud: guard the path parameters, pin refusals, drop the bare except
cloud: harden document field fixture validation
cloud: harden generation response handling
cloud: harden the estate preflight probes per security review
cloud: harden the point default and the artifact publish tool
cloud: harden the preflight per the CodeRabbit review threads
cloud: hold the D1 remote-backend boundary in live preflight too
cloud: hold the priced generation window’s prefill mix to a review
cloud: isolate deployment plan inputs
cloud: keep production catalog topology strict
cloud: keep the pricing-authority resolver import-safe off a checkout
cloud: make shape handling total and stop rejecting the honest plan
cloud: make the unpriced-storage zero loud
cloud: parse s3 backend blocks brace-aware so nested blocks are read
cloud: pin the AWS-edge coupling and refuse malformed input everywhere
cloud: read not-owned terraform roots through a narrow allowlist
cloud: read trust_remote_code from both adapter scopes everywhere
cloud: recover externally cancelled live gateway candidate
cloud: refuse impossible instants and stop the scope string overclaiming
cloud: register migration 0079 and name the book on a merged sweep
cloud: reject a run paired with itself and fix the stale exit bullet
cloud: reject lineage cycles and parented full runs
cloud: reject unusable SSM pagination tokens; pin cross-estate isolation
cloud: repair chat catalog canary
cloud: repair document field canary semantics
cloud: require exact screenshot labels
cloud: require statement constancy per security-relevant field
cloud: restore cross-estate rejection on excluded entries, refuse composition
cloud: restrict the instant pattern to ASCII digits
cloud: retry transient job reads
cloud: scan every leaf for foreign-estate references, not just name keys
cloud: scope key mutations to an account the caller names
cloud: snapshot untracked development roots
cloud: stabilize catalog contract canary
cloud: stabilize screenshot acceptance fixture
cloud: stabilize screenshot contract canary
cloud: stabilize whisper catalog canary
cloud: stop arming gateway proxy auth without a public edge
cloud: stop publishing SIE’s markup to customers
cloud: tighten promotion evidence validation
cloud: tolerate EC2 worker visibility lag
cloud: track %{} template-directive spans in the backend classifier
cloud: track template interpolations in the backend classifier
cloud: unpin gateway path dependency
cloud: validate mode, harden provider-key parsing, make two tests honest
cloud: validate read entries and close self-review fail-opens
cloud: validate the estate id where it enters, tighten approver
cloud: verify materialized candidate bytes
deps: provide the dependencies the bundle and the eval already declare
model-ops: pin trusted digests and refuse unrecorded provenance
release: unpin sidecar path dependency
sdk: refresh stale job result refs
server: echo queued extract item ids
server: normalize Grounding DINO instructions
server: preserve pinned models in serve config
server: validate pinned model selection
tools: allow SDK 0.7 releases

Code Refactoring

terraform: drop control-plane contract marker

v0.6.30 (2026-08-07)

Highlights

New capabilities: compile the deployment matrix from the estate registry
Reliability and operations: drop the invalid queue key from the staging concurrency blocks
Performance: accelerate Qwen thinking inference

Features

cloud: compile the deployment matrix from the estate registry

Bug Fixes

ci: drop the invalid queue key from the staging concurrency blocks

Performance Improvements

models: accelerate Qwen thinking inference

v0.6.29 (2026-08-07)

Highlights

New capabilities: bind evidence and acceptance to estate identity; optimize Gemma thinking profiles
Reliability and operations: preserve release tag during public sync verification; verify the acceptance OIDC identity token

Features

cloud: bind evidence and acceptance to estate identity
server: optimize Gemma thinking profiles

Bug Fixes

ci: preserve release tag during public sync verification
cloud: verify the acceptance OIDC identity token

v0.6.28 (2026-08-07)

Highlights

New capabilities: derive resource names and paths from estate coordinates; gate public edge and vendor identity on estate capability; split managed gates into tier and estate capability; optimize non-thinking generation profiles
Reliability and operations: classify entity canary failures; gate Stripe preflight before any provider request; keep vendor provider access tier-gated; refresh siglip diagnostic projection; remove stale console gateway URL and reject tampered cells

Features

cloud: derive resource names and paths from estate coordinates
cloud: gate public edge and vendor identity on estate capability
cloud: split managed gates into tier and estate capability
server: optimize non-thinking generation profiles

Bug Fixes

cloud: classify entity canary failures
cloud: gate Stripe preflight before any provider request
cloud: keep vendor provider access tier-gated
cloud: refresh siglip diagnostic projection
cloud: remove stale console gateway URL and reject tampered cells
cloud: scope Modal denial floor to serving estates
cloud: update gateway profile fixtures
quality: extend PubTabNet task budget
release: extract refresh artifact by id
release: keep diagnostic suites out of git
release: preserve refresh artifact across retries
server: pin speculative draft launches
server: stage drafts for local models

v0.6.27 (2026-08-06)

Highlights

New capabilities: add staging comparison diagnostics; pin benchmark campaigns to a verified main revision; add connector plan controls; expose recovery-bound connector repair; activate atomic connector run; activate explicit connector execute
Reliability and operations: authenticate quality gate attempt markers; authenticate req14 quality gate markers; pin void authorization Python; bind retained recovery to candidate tasks; bind vendor mutations to estate authority

Features

bench: add staging comparison diagnostics
ci: pin benchmark campaigns to a verified main revision
clients: add connector plan controls
clients: expose recovery-bound connector repair
cloud: activate atomic connector run
cloud: activate explicit connector execute
cloud: activate governed postgres connectors
cloud: activate recovery-bound connector repair
cloud: add atomic connector run admission
cloud: add bounded connector recovery probes
cloud: add bounded postgres source planning
cloud: add component convergence contracts
cloud: add connector execution journal client
cloud: add connector plan guardrails
cloud: add connector plan schema policy
cloud: add crash-safe connector outer runner
cloud: add direct connector inference
cloud: admit executable connector plans
cloud: attest connector workers before payload
cloud: band the size-class tier book on (operation, parameters)
cloud: band the tier book on (operation, parameters)
cloud: bind connector executor dispatch
cloud: bind connector plans to worker identity
cloud: bind direct connector executions
cloud: bound postgres connector execution context
cloud: bridge durable connector execution dispatch
cloud: commit the control-plane OpenAPI contract with a diff gate
cloud: compile estate vendor coordinates
cloud: compile optional estate vendor authority
cloud: compose connector recovery journal
cloud: compose dormant connector execution
cloud: enable native console previews
cloud: enable selective component deployment
cloud: expose connector execution transitions
cloud: expose durable connector plans
cloud: fail closed on unconfigured vendor authority
cloud: fence postgres connector publication
cloud: govern connector plan manifests
cloud: isolate SigLIP rate-grade diagnostics
cloud: journal connector recovery phases
cloud: land governed Postgres connector runtime
cloud: persist connector batch lane authority
cloud: persist connector execution authority
cloud: persist durable connector plans
cloud: preflight connector worker discovery
cloud: prepare and fence connector execution
cloud: prove executable postgres preflight
cloud: provision connector runtime authorities
cloud: provision isolated console previews
cloud: reconcile connector publication recovery
cloud: record the owner’s acceptance of the declared exposure
cloud: replace catalog DSL with task implementations
cloud: replay governed postgres manifests
cloud: resolve estate coordinates at the provisioner boundary
cloud: resolve terraform coordinates from the estate registry
cloud: route generation through local ingest
cloud: route managed generation through local ingest
cloud: scaffold additional staging estate
cloud: scaffold sandbox foundation
cloud: settle connector staging before publication
cloud: unify Modal environment authority in the estate registry
gateway: expose immutable execution evidence
generation: add 256k profile candidates
generation: add hardware-specific context profiles
generation: add strict Responses API parity
model-ops: allowlist lightonai and OpenSearch-AI as trusted publishers
model-ops: capture nested multimodal config scalars in onboarding evidence
model-ops: enable multimodal image-text and document-retrieval discovery families
model-ops: enable sparse-retrieval discovery family
model-ops: re-task late-interaction discovery onto SCIDOCS retrieval evidence
model-ops: trust JinaAI discovery publisher
models: onboard lightonai/mLateOn
models: onboard lightonai/mLateOn (scaffolded)
models: seed mLateOn quality floors from governed measurement
server: add bounded local-ingest generation client
server: complete generative model serving
sidecar: stream generation over local ingest
terraform: register shared Scalr modules

Bug Fixes

bench: validate release evidence ingress early
build: relock every standalone project that depends on sie-server
ci: align standalone validation artifacts
ci: authenticate quality gate attempt markers
ci: authenticate req14 quality gate markers
ci: clarify invalid model gate error
ci: exempt registry library modules from the estate root list
ci: fail closed for unlisted registry module roots
ci: format campaign source tests and state the lowercase SHA rule
ci: isolate connector verifier dependencies
ci: pin void authorization Python
ci: trust quality gate attempt cap source
ci: validate quality gate void reasons
clients: send connector repair idempotency header
cloud: accept launch catalog schema v2
cloud: accept operation-valid usage supersets
cloud: address connector runtime review findings
cloud: address Gateway identity review
cloud: address staging rollback review
cloud: align estate module contracts
cloud: align preview posture lint config
cloud: allow console preview bootstrap
cloud: allow verified gateway final reuse
cloud: batch the token-capped encode lanes and price two more launch cells
cloud: bind connector intent to canonical worker
cloud: bind every priced-window factor to non-caller evidence
cloud: bind generation lanes to catalog hash
cloud: bind retained benchmark owners
cloud: bind retained recovery to candidate tasks
cloud: bind task evidence to reviewed cases
cloud: bind the billed report period to the priced window
cloud: bind vendor mutations to estate authority
cloud: bound Modal history reads separately
cloud: bound Modal inventory reads separately
cloud: close connector activation review gaps
cloud: close connector review findings
cloud: close connector review gaps
cloud: close two vacuous guards in the tier banding logic
cloud: declare the below-cost classes instead of hiding them
cloud: default estate_id inside resolve_estate
cloud: deny unused staging scheduler authority
cloud: derive allocation shares from measured work, not fiat
cloud: derive offered load from measured lane capacity
cloud: disable staging cold-start reporting
cloud: drop the exporter’s path workaround, annotate the document type
cloud: enforce console preview posture
cloud: fence selective Gateway dispatch by contract
cloud: finish connector protocol review sweep
cloud: fold the 2026-08-04 lane sweeps into the offered-load plan
cloud: forward selective deployment mode
cloud: gate connector activation on desired state
cloud: handle local generation stream failures
cloud: harden connector activation boundaries
cloud: harden deployment mode forwarding
cloud: harden edge backend ordering, cell parity, bucket names
cloud: harden managed generation streaming
cloud: harden Modal inventory validation
cloud: harden preview OIDC validation
cloud: harden registry identity validation and lock estate identity
cloud: harden selective deployment reuse
cloud: harden staging estate isolation
cloud: harden staging foundation identity
cloud: ignore local source caches
cloud: isolate console preview git config
cloud: isolate sidecar subprocess secrets
cloud: isolate staging estate authorities
cloud: make Vercel readback authoritative
cloud: measure the enlarged fixtures and price 18 more launch cells
cloud: merge main and add band-6 rerank capacities
cloud: normalize component source modes
cloud: omit read-only webhook rules
cloud: omit response-only webhook rules
cloud: order config image local layers
cloud: pin gateway plan evidence
cloud: pin sandbox Scalr modules at their first healthy versions
cloud: preserve connector execution topology
cloud: preserve positive sandbox activation target
cloud: preserve retained worker provenance in release gates
cloud: preserve valid rate book evidence
cloud: price the measured window, not the billed report hour
cloud: re-render siglip diagnostics lock after server source change
cloud: reconcile connector activation with main
cloud: record the measured decode length on the generation cells
cloud: recover native production gateway
cloud: recover retained candidates after newer merges
cloud: refresh combined topology evidence
cloud: refresh siglip diagnostic lock
cloud: refresh task contract test fixtures
cloud: refuse an unpublished class name instead of raising KeyError
cloud: reject empty connector inference text
cloud: reject JSON Terraform escapes
cloud: remove unused type suppression
cloud: report rejected reuse evidence
cloud: require integer estate schema version
cloud: reserve governed page ceilings
cloud: restrict promotion authority to full estates
cloud: retain legacy standalone tag gate
cloud: scope promotion authority to estate cells
cloud: stabilize component binary identity
cloud: stabilize named-entity canary fixture
cloud: state the audio_ms clip-length assumption in the quote
cloud: tighten connector manifest cleanup
cloud: tighten task contract evidence gates
cloud: unblock reproducible Modal convergence
cloud: use explicit connector protocol stubs
cloud: use non-ambiguous visual distractor
cloud: use representative vision smoke fixture
cloud: validate retained migration without activation
cloud: validate reused worker runtime CAS
cloud: validate selected gateway in catalog smoke
cloud: validate settled acceptance usage
cloud: verify Modal service-user access
colbert: honor transformers-5 rope_parameters when serving on the 4.x bundle
generation: align direct backend output limits
generation: defer resolved profile output limits
generation: fit Gemma 256k KV cache on H100
generation: handle family-specific reasoning boundaries
generation: hide split Gemma thinking boundary
generation: pair long-context qwen image bounds
generation: preserve Gemma reasoning boundaries
generation: preserve qwen vision bounds
generation: reject overflowing runtime numerics
generation: reserve Gemma long-context workspace
generation: unblock cold model startup
generation: use compatible Qwen 27B grammar backend
generation: use compatible Qwen 35B grammar backend
generation: validate high-context thinking profiles
modal: isolate overlay FlashInfer cache
modal: keep image yaml loading local
modal: support strict GPU allocation
modal: validate strict GPU selectors
model-ops: an unmeasured expected_dim never confirms a multivector smoke
model-ops: batch discovery evidence revision listing via one pinned tree fetch
model-ops: bind trusted terminal outcomes
model-ops: bound authority listing by model directories, not raw entries
model-ops: calibrate late-interaction and doc-retrieval scanner blocks
model-ops: close scaffold authorization gaps
model-ops: complete scaffolds before Stage 4
model-ops: exempt tooling stops from attempt cap
model-ops: fail closed on duplicate search errors
model-ops: harden scaffold validation boundaries
model-ops: ignore fenced tooling markers
model-ops: insert matrix rows once per models node so alias blocks ride their anchor
model-ops: isolate scaffold candidate tests
model-ops: namespace the config projection_dim claim and cap consumer-side nested ints
model-ops: order attempt comments deterministically
model-ops: preserve adapter auto authorization
model-ops: scope duplicate onboarding guard
model-ops: teach the Stage-4 smoke to probe multivector models
models: complete mLateOn serving recipe
quality-eval: harden runner repair
quality-eval: quarantine unregistered runners
quality-eval: shorten wake grace and map sglang bundles
quality-eval: shorten wake grace and support sglang logical bundles
quality-eval: use prepared scorer environment
release: bound commit history batches
release: deduplicate generated changelog notes
release: finalize source before diagnostics
release: refresh topology-bound diagnostics
release: reject repeated release headings
sdk: retry pre-execution SSE capacity errors
server: align Gemma xgrammar dependency
server: await standalone smoke teardown
server: bound Qwen visual context
server: constrain SGLang compatibility paths
server: enforce Qwen fast image bounds
server: keep Qwen3-VL reranker batches working on transformers 5.3
server: pin docling to its artifact revision and stop OCR gating load
server: preprocess native extract images
server: preserve LightOn SGLang compatibility
server: replace stale root playground with redirect to Swagger docs
sidecar: harden local generation cancellation
sidecar: reserve cancel request bytes
terraform: address internal module review
terraform: bound internal module lifecycle
terraform: exclude worktree caches from remote config uploads
terraform: harden Scalr module releases
terraform: package internal Scalr module archive root
terraform: package private Scalr registry modules
terraform: package Scalr module archive envelopes
terraform: register Scalr modules from plain monorepo roots
terraform: reject null module registry provider
terraform: relocate Scalr registry modules under deploy tree
terraform: validate sandbox module with OpenTofu

v0.6.26 (2026-08-02)

Highlights

New capabilities: add H100 FP8 generation profiles; governed first-baseline mode + 16 vanilla floors for Qwen3-Embedding-8B and nomic v2 MoE
Reliability and operations: add protected retained recovery dispatch; bind modal env before retained recovery; bind retained gateway recovery to failed release; bound retained recovery runtime; clarify retained gateway recovery

Features

model-catalog: add H100 FP8 generation profiles
quality-eval: governed first-baseline mode + 16 vanilla floors for Qwen3-Embedding-8B and nomic v2 MoE

Bug Fixes

bench: initialize buildx for worker image
bench: quiesce gc during timed campaigns
bench: validate non-root worker runtime
cloud: add protected retained recovery dispatch
cloud: address provision review findings
cloud: align rollback preflight diagnostic
cloud: allow full cold dispatch preflight budget
cloud: audit extracted provision secrets
cloud: bind modal env before retained recovery
cloud: bind retained gateway recovery to failed release
cloud: bind retained gateway to failed release
cloud: bound retained recovery runtime
cloud: clarify retained gateway recovery
cloud: close provision review gaps
cloud: converge config before cold gateway
cloud: fence initial gateway absence recovery
cloud: govern reconciliation cadence in releases
cloud: guard bootstrap and benchmark worker runtime
cloud: guard current-schema bootstrap seed
cloud: harden connector smoke recovery
cloud: harden extracted provision edges
cloud: harden retained recovery diagnostics
cloud: honor generation smoke cold starts
cloud: isolate config app runtime import
cloud: isolate connector smoke environments
cloud: isolate sibling smoke environments
cloud: keep connector diagnostic out of MVP release gate
cloud: keep prod generation warm and scope spans
cloud: make generation smoke single shot
cloud: narrow connector loopback scope
cloud: preserve accepted bundle promotion compatibility
cloud: preserve recovery fence diagnostics
cloud: recover first gateway to absence
cloud: recover retained legacy gateway
cloud: recover retained staging gateway to absence
cloud: recover staging gateway to absence
cloud: reject incomplete cold gateway plans
cloud: retain production deployment evidence
cloud: retain zero-floor generation attestations
gateway: preserve grammar-safe FP8 profiles
loadtest: prepare pinned perf-lab source before AWS
model-ops: bind scaffold AUTO to trusted evidence
model-ops: mirror task-equivalent standard blocks
preserve cold-start reporter membership history
quality-eval: preserve missing metric diagnostics
quality: run comparison through locked project
release: identify expired Modal attestation
release: keep Modal diagnostic trim ASCII-only
release: reconcile lagging Modal attestations
release: tolerate completed Modal exec targets
release: wait for refreshed PR head
validate initial reporter lease atomically

v0.6.25 (2026-07-30)

Highlights

New capabilities: activate bootstrap-v2 pricing in prod; add —skip for recorded vendor-convergence skips; add cold-start live acceptance proof; add the mechanical launch-rate candidate promotion pipeline; add the size-class tier commercial policy to the rate book; admit rate-grade evidence in the launch catalog and coverage audit
Reliability and operations: admit an image-witnessed authoritative zero at settlement; align runner authorization with smoke protocol; async-surface reservation parity — video guard, byte cap, tokenizer allowance; bound gateway cold-start recovery; cap identity convergence retry seam
Performance: index the rotation-successor probe on api_keys

Features

cloud: activate bootstrap-v2 pricing in prod
cloud: add —skip for recorded vendor-convergence skips
cloud: add cold-start live acceptance proof
cloud: add the mechanical launch-rate candidate promotion pipeline
cloud: add the size-class tier commercial policy to the rate book
cloud: admit rate-grade evidence in the launch catalog and coverage audit
cloud: batch/jobs generation via buffered single-terminal settlement
cloud: bound and measure the sealed cold-start subsidy
cloud: carry an authoritative-dimension claim set on UnitCounts
cloud: carry the settled charge into realtime usage blocks
cloud: commit reviewed launch allocations
cloud: editable per-key spend limits + monthly auto-recharge cap
cloud: editable per-key spend limits + monthly recharge cap
cloud: GA audio on /v1/extract — book-driven whisper pricing
cloud: gate production on a degraded staging acceptance
cloud: generation input-token reservation bound + settle-contract pins
cloud: gpu_second sealed billing regime — custom models become sellable
cloud: install digest-bound generation routes
cloud: install digest-bound governed generation routes
cloud: lease dynamic cold-start reporters
cloud: low-balance threshold email outbox + SES seam
cloud: low-balance threshold emails via SES
cloud: metering GA — waves 1–3
cloud: move the license exclusion verdict to the registry resolution seam
cloud: org-scoped usage export API — CSV + JSON
cloud: persist cold-start subsidy evidence
cloud: POST /v1/estimate dry-run cost endpoint
cloud: POST /v1/estimate dry-run cost endpoint + SDKs
cloud: production-bootstrap-v2 — prefill closed, sealed lanes priced
cloud: promote launch-rate candidates to rate-grade evidence
cloud: rate-grade size-class tier book assembly + operator-gated activation
cloud: response-final settlement for streaming generation
cloud: storage GiB-day accrual + book-driven register quote
cloud: surface credits_charged + rate_book_version in usage
cloud: unfence OpenAI-compat /v1/audio/transcriptions
cloud: unify the settled-charge name across jobs, batches, and SDKs
cloud: weekly/daily reconciliation with paging alerts and operator-gated LIVE mode
cloud: weekly/daily reconciliation, attribution categories and the operator LIVE gate
cloud: widen whole-job settlement to every contract dimension
gateway: add optional governed generation routing seam
model-ops: add ASR discovery family
model-ops: add fail-closed discovery family registry
model-ops: add multimodal discovery families
model-ops: add object detection discovery
model-ops: add OCR discovery families
model-ops: add report-only generation discovery
model-ops: add VQA and KIE discovery
model-ops: discover guardrail candidates
model-ops: discover sparse retrievers
model-ops: discover structured extraction candidates
model-ops: discover text classification candidates
model-ops: discover text rerankers
model-ops: read the executor’s trust clearance from the onboarding artifact
model-ops: register guardrail discovery family
model-ops: register structured extraction discovery
model-ops: register text classification discovery
model-ops: scan guardrail model candidates
model-ops: scan structured extraction candidates
model-ops: scan text classification candidates
model-ops: support report-only family scans
model-ops: wire the trust_remote_code scan into the AUTO gate
models: onboard Qwen/Qwen3-Embedding-8B (scaffolded)
sdk: estimate() on both SDKs for the /v1/estimate dry run
server: bill sampled video frames as images
telemetry: export the sealed cold-start metric remotely and chart it

Bug Fixes

benchmark: validate sorted deployment evidence
catalog: bound GLiNER sequence lengths
ci: reseal staging catalog evidence
cloud: accept client-scoped WorkOS sessions
cloud: accept Terraform canonical false in bootstrap guard
cloud: accept WorkOS environment issuers
cloud: address billing bootstrap review
cloud: address CodeRabbit review — SES client bounds, send pacing, re-arm audit
cloud: address CodeRabbit review on the /v1/estimate dry run
cloud: address CodeRabbit review on the key-limit + recharge-cap PR
cloud: address CodeRabbit review on the settled-charge surfaces
cloud: address CodeRabbit review on the usage export
cloud: address prerequisite gate review
cloud: address the CodeRabbit review on the batch/jobs generation PR
cloud: address the CodeRabbit round on the metering GA integration
cloud: address the second CodeRabbit round on /v1/estimate
cloud: admit an image-witnessed authoritative zero at settlement
cloud: admit exact task-role tag bootstrap
cloud: admit observed Modal US locations
cloud: align benchmark smoke with rated profile
cloud: align runner authorization with smoke protocol
cloud: allow nested ECR repositories
cloud: async-surface reservation parity — video guard, byte cap, tokenizer allowance
cloud: attest benchmark runtime identity
cloud: bind a rate_grade key to the price and hardware its envelope measured
cloud: bind nightly health checks to staging
cloud: bind paused forwarder candidate before prod cutover
cloud: bind the charge-slot backstop to the stream, not the disconnect grace
cloud: block the generation in/out split on a named measurement
cloud: bound gateway cold-start recovery
cloud: bound nightly evidence execution
cloud: bound the sealed resolver cache and cap proxy response bodies
cloud: cap identity convergence retry seam
cloud: claim the hold before releasing it, and fold case like the registry
cloud: classify benchmark evidence failures
cloud: close cold recovery review gaps
cloud: close cold-start evidence failure modes
cloud: close final metering review faults
cloud: close meter-only review gaps
cloud: close metering GA review findings
cloud: close rate evidence review gaps
cloud: close reporter lease invariants
cloud: close six gate findings on the /v1/estimate dry run
cloud: close the batch-2a gate findings on editable limits + recharge cap
cloud: close the CodeRabbit round on the candidate-promotion branch
cloud: close the CodeRabbit round on the integration head
cloud: close the CodeRabbit round on the reconciliation branch
cloud: close the CodeRabbit round on the waves-1-3 integration head
cloud: close the final CodeRabbit round on /v1/estimate
cloud: close the gate findings on the resolution-seam verdict
cloud: close the promotion gate’s operator-assertion escapes
cloud: close the second CodeRabbit round on the credit surfaces
cloud: CodeRabbit review — extract params must be explicit per line
cloud: CodeRabbit review fixes — bounded sealed-class cache, doc reconciliation
cloud: CodeRabbit review fixes — quote authority degradation, single-model projection
cloud: CodeRabbit review fixes — sealed posture floor, generator provenance
cloud: CodeRabbit round 2 — evidence doc, strict decode, single query walk
cloud: CodeRabbit round 3 — exhaustive ceiling match, honest parse row
cloud: CodeRabbit round on the control-plane prerequisite gate
cloud: CodeRabbit round-2 — redact resolver errors, tighten class guard
cloud: CodeRabbit round-6 — quote survives a dead pool, pin e2e clock
cloud: CodeRabbit round-7 — atomic cache admission, sweep resilience
cloud: compile exact generation customer routes
cloud: compile exact generation routes
cloud: correct console wallet path to /console/usage — verified against sie-web
cloud: correct the race() return annotation in the snapshot test
cloud: cover regular fleet cold starts
cloud: cover regular Modal fleet cold starts
cloud: decide license exclusion at the registry resolution seam
cloud: declare generation input_tokens in the launch catalog
cloud: default customer charging to disabled
cloud: disable recharge schedule without teardown
cloud: distinguish authoritative zero from missing count at settlement
cloud: enforce the image witness on both sides of the rating boundary
cloud: expose collector test configs to non-root
cloud: fail a tier book whose price point has no exact Metronome decimal
cloud: fail closed on incomplete preflight reads
cloud: fail closed when a declared skip cannot be recorded
cloud: fail loudly when test prerequisites vanish
cloud: gate batch usage on settlement evidence, not a running counter
cloud: gate estimates on resolved model identity
cloud: gate fixes for the usage export — day-sliced scan, pool isolation, uniform masking
cloud: harden cold-start acceptance proof
cloud: harden dynamic cold-start membership
cloud: harden staging release verification
cloud: harden the multipart licensing gate and lock the route wiring
cloud: index count-gated SES resources via one()/splat
cloud: isolate benchmark evidence runtime
cloud: log truncated usage-export streams; pin legacy-row null passthrough
cloud: make extract canonical for Docling billing
cloud: make generation rate evidence prefix-cache safe
cloud: make test prerequisites and tracing deterministic
cloud: make the jobs loopback estimate allowance-aware and prove the guard end to end
cloud: make the jobs planner uphold the invariant settlement now enforces
cloud: make the reconciliation bars statements the cost basis supports
cloud: make the response cap operative, unbilled, and honest
cloud: make the witnessed zero reach the wire and cover both rails
cloud: merge candidate promotion and close #2628 review
cloud: mirror the widened model canonicalization in the control plane
cloud: observe the sealed cold-start metric; scope it to the prometheus path
cloud: parse the multipart model under the route’s real body limit
cloud: pin benchmark credential versions
cloud: pin the sealed tariff evidence; harden the split and the posture tests
cloud: pin the storage eligibility cutoff to UTC
cloud: preflight input token rate bounds
cloud: preserve cold-start evidence gaps
cloud: preserve inactive prod rollback tasks
cloud: price sealed-lane memory at the Sandbox rate, not standard
cloud: propagate connector request identity
cloud: publish a committed charge even when settlement then faults
cloud: pull the authoritative invoice inside the single-flight lock
cloud: recipient chain, re-arm edge, staleness bound, per-row commit
cloud: reconcile what the batchgen merge hid from git
cloud: record an admission outcome on every licensing rejection
cloud: redact benchmark operator preflight
cloud: reject a negative storage-accrual lookback
cloud: reject an explicit storage-accrual day that is not closed
cloud: reject expired recovery evidence
cloud: reject null WorkOS client claims
cloud: release the hold when the gateway cannot deliver the result
cloud: renumber monthly-cap migration to 0040 — 0039 taken by sibling W3 branch
cloud: renumber rotation-successor index migration to 0043 — 0042 taken by #2580
cloud: report diagnostic cleanup failure
cloud: report every live-preflight failure in one pass
cloud: report malformed release pointers
cloud: reseal release catalog dependents
cloud: restore the spend-limit error contract on the W3 base
cloud: retain exact benchmark deployment evidence
cloud: retain exact benchmark deployment evidence\n\nRefs #2573
cloud: retry benchmark identity convergence
cloud: round-8 minors — env var name, authority restore, day-sweep guard
cloud: route expected no-charge stream releases off the fault alerts
cloud: route readiness failures through the gate without widening the opt-out
cloud: route the async reservation through the shared video and token seams
cloud: scope authoritative zero evidence
cloud: scope rate preflight to exact authority
cloud: scope the generation input bound to books that price input
cloud: seal generation campaign artifact paths
cloud: tell the truth about the video byte rail, and bound the tests that prove it
cloud: tighten console origin validation, pace off the previous row
cloud: tighten evidence diagnostic contract
cloud: treat an explicit null usage as an empty block, not a refusal
cloud: treat an explicitly null video as absent, not as a video
cloud: type-gate the chat_template_kwargs allowlist
cloud: unblock billing bootstrap migrations
cloud: unseal console-free release subsets and offload-smoke
cloud: unseal sealed release subsets, report every preflight failure, and gate degraded acceptance
cloud: valid IAM tag charset + keep bootstrap secret-gate count at 2
cloud: validate tier scheme inputs canonically
cloud: wire the meter-only operator gate
gateway: cap embeddings requests at 256 inputs
gateway: harden governed generation dispatch
gateway: preserve self-host grammar routing
loadtest: use job token for GHCR pulls
model-ops: accept GLiNER discovery models
model-ops: address authorization review feedback
model-ops: address evidence review feedback
model-ops: authenticate the evidence producer and bind the identity to execution
model-ops: bind CodeRabbit review completion
model-ops: bind crash fallback repository
model-ops: bind executor authorization
model-ops: bind trusted completion status
model-ops: close final trust gate review gaps
model-ops: close malformed trust evidence gap
model-ops: close three bypasses in the trust_remote_code gate
model-ops: dispatch the Stage 4-5 gate completion explicitly
model-ops: distinguish disabled discovery families
model-ops: exclude OCR-only KIE models
model-ops: explicitly dispatch discovery handoffs
model-ops: gate readiness on bot reviews
model-ops: harden bot review handoff
model-ops: harden cold Modal image builds
model-ops: harden completion status fallbacks
model-ops: harden report-only scanning
model-ops: harden trust gate provenance
model-ops: keep a failed gate origin recoverable
model-ops: narrow document OCR compatibility
model-ops: never log gh’s stderr, which carries GH_TOKEN by construction
model-ops: preserve completed review mutations
model-ops: re-derive the clearance in matrix-block, bind evidence to the revision
model-ops: re-grant contents:read to the dispatch job
model-ops: recognize Copilot bot aliases
model-ops: reconcile gate recovery reviews
model-ops: refuse an unset RUNNER_TEMP instead of falling back into the workspace
model-ops: refuse instead of crashing when the evidence fetch faults
model-ops: refuse unusable evidence instead of falling back to blockers
model-ops: reject ambiguous family matches
model-ops: reject failed Copilot reviews
model-ops: repair gate recovery replay
model-ops: retry transient Modal build downloads
model-ops: scan all loader-visible code
model-ops: scan zero-shot classification models
model-ops: suppress dynamic artifact failure logging
model-ops: tolerate pending status retries
model-ops: validate gate head before status writes
model-ops: validate review completion evidence
model-ops: wake the adapter executor from an agentless dispatch job
model-ops: wake the adapter-exec executor by explicit dispatch
quality-eval: retain models pending vanilla floors
quality-eval: surface unfloored matrix rows
release: bootstrap a fresh production database
release: compare fresh proof types exactly
release: retry malformed Modal attestations
release: retry transient Modal attestation output
release: verify empty replay target digest
server: address CodeRabbit review on the video meter
server: bill SigLIP padded text work
server: bound isolation fan-out and close the NaN decode escapes

Performance Improvements

cloud: index the rotation-successor probe on api_keys

v0.6.24 (2026-07-26)

Highlights

New capabilities: add agent shell GPU fallbacks; add agent shell placement controls; add launch catalog contract canary; add Modal-native placement primitives; add native placement primitives; allow snapshotless native lanes
Reliability and operations: align rate-book recovery checks; harden dispatch preflight gate; harden final generation gates; harden retained release recovery; preserve capacity diagnostics on timeout races
Performance: cache MUVERA projection state

Features

cloud: add agent shell GPU fallbacks
cloud: add agent shell placement controls
cloud: add launch catalog contract canary
cloud: add Modal-native placement primitives
cloud: add native placement primitives
cloud: allow snapshotless native lanes
cloud: attest realized placement identity
cloud: establish broad-native release baseline
cloud: govern prod legacy billing cutover
cloud: preflight offline tokenizer dependencies
server: pinnable LoRA adapter revisions in lora_paths

Bug Fixes

cloud: accept scheduled orphan-hold sweeps
cloud: address native placement review
cloud: address release review feedback
cloud: align rate-book recovery checks
cloud: align tokenizer snapshot dependencies
cloud: allow absent generation in final smoke
cloud: allow bounded prod platform bootstrap
cloud: attest final gateway replacement
cloud: bound snapshot lifecycle boots
cloud: bound worker identity collection
cloud: canonicalize generation GPU profiles
cloud: close bounded prod bootstrap plan
cloud: close launch review findings
cloud: declare bootstrap alert container
cloud: default orphan hold sweep ttl
cloud: derive exact rate topology from broad release
cloud: derive gateway gpu admission from plan
cloud: diagnose bounded worker capacity waits
cloud: drain legacy generation revisions
cloud: enforce global canary request identity
cloud: fail closed on tokenizer pin drift
cloud: finalize gateway after generation deploy
cloud: frame reusable candidate outputs
cloud: gate billable release commit safely
cloud: gate every gateway replica on dispatch readiness
cloud: guard broad placement pending realized identity
cloud: harden dispatch preflight gate
cloud: harden final generation gates
cloud: harden retained release recovery
cloud: integrate launch remediation release gates
cloud: keep slow batch leaders warm
cloud: make Vercel deploy creation transport-safe
cloud: package excluded gateway target
cloud: package gateway from crate target
cloud: preflight gateway generation dispatch
cloud: preflight route and worker hash parity
cloud: preserve capacity diagnostics on timeout races
cloud: preserve launch route errors
cloud: preserve snapshot cancellation cleanup
cloud: reconcile launch remediation release gates
cloud: recover ambiguous Vercel deploy creates
cloud: recover bounded prod bootstrap outputs
cloud: recover rate-book releases drain-first
cloud: redact readiness diagnostics
cloud: refresh staging catalog evidence digests
cloud: rehearse immediate launch accounts
cloud: reject partial readiness placement
cloud: reserve token inputs from governed bounds
cloud: respect Better Stack token scope
cloud: restore authenticated gateway rollback health
cloud: restrict rehearsal recovery fields
cloud: serialize bootstrap state reads
eval: make run-eval-group.sh and its tests portable to macOS
eval: reject symlink escapes from the /tmp guard; reap the requirements temp dir
loadtest: use valid AWS IAM purpose tag
models: declare tokenizer dependencies
quality: address remaining review findings
quality: close nightly model gaps and add exact quarantine
quality: gate bounded NVIDIA MUVERA checks
quality: isolate parameterized FiNER evals
quality: preserve floor metric contracts
quality: record bounded concurrency-one targets
quality: record faithful CMedQA floors
quality: record faithful quality floors
quality: reject non-finite qrel scores
quality: restore bounded FiNER coverage
quality: restore Mixedbread query expansion
quality: stabilize scoped PR quality runs
release: close prod bootstrap output contract
release: redact bootstrap state read failures
smoke: restore blocking-probe assertions and cold-touch retry coverage
worker: copy sie_telemetry crate into worker image builds

Performance Improvements

server: cache MUVERA projection state

v0.6.23 (2026-07-24)

Highlights

New capabilities: accept native multimodal generation; support routed media measurement campaigns; add production bootstrap rate book; add pre-public task measurement candidates; bind launch catalog to commercial rate book; bind launch rates to commercial book
Reliability and operations: harden generation quality sharding; harden multimodal batch validation; harden queue readiness for generation defaults; harden shard failure capture; restore governed retrieval floors
Performance: bind bounded extraction evidence; bind current nonvision launch evidence; recover granite guardian queue evidence; reuse siglip preprocessing

Features

api: accept native multimodal generation
bench: support routed media measurement campaigns
billing: add production bootstrap rate book
catalog: add pre-public task measurement candidates
catalog: bind launch catalog to commercial rate book
catalog: bind launch rates to commercial book
catalog: prepare launch serving evidence
catalog: prepare non-public launch catalog checkpoint
catalog: record current nonvision evidence
catalog: record Snowflake and R3 evidence
catalog: validate native multimodal recipes
catalog: verify document vision recipes
catalog: verify existing launch recipes
catalog: verify GLiNER2 large launch evidence
catalog: verify R3 pipeline on L4
catalog: verify redaction task evidence
catalog: verify Snowflake on preferred L4
ci: complete req14 gates after lane runs
ci: remove runner-held waits from official-recipe gates
cloud: activate accounts and route launch events
cloud: add bounded prod platform bootstrap
cloud: add exact launch rate campaign suite
cloud: add immutable managed release pipeline
cloud: add launch account credit flow
cloud: add launch rate campaign suite
cloud: add Qwen task evidence campaign
cloud: add secure staging catalog evidence dispatch
cloud: add secure staging evidence dispatch
cloud: archive Modal measurement releases
cloud: assemble commercial rates from exact evidence
cloud: authenticate Modal rate evidence
cloud: bracket generation rate window
cloud: compile isolated measurement topology
cloud: derive commercial rates from exact resource seconds
cloud: derive commercial rates from Modal billing
cloud: execute document vision recipes
cloud: execute remaining reviewed recipes
cloud: route signup notifications directly to Slack
cloud: verify BGE and GTE launch evidence
deploy: add scalr-bootstrap root for internal terraform state
deploy: run scalr-bootstrap under opentofu with live scalr backend
deploy: scalr-held terraform state for internal roots
deploy: split scalr workspaces into cli- and vcs-driven
generation: type structured output contracts
model-ops: split feasibility PROCEED-WITH-ADAPTER-AUTO vs human
model-ops: trust_remote_code security pre-gate
quality-eval: in-pipeline wall-clock stop with partial report
sdk: add typed Responses client

Bug Fixes

1841: stamp sealed metric units on the Modal collector, and make the CP filters cover what the suite reads
adapters: serialize output_hidden_states forwards — close the #2144 recorder-race class
adapters: serialize output_hidden_states forwards to close the #2144 recorder-race class
api: address multimodal contract review
api: close multimodal parity gaps
bench: add faithful PyLate official recipe
bench: address launch evidence review
bench: align Any2Any dispatch semantics
bench: align ColBERT reference validation
bench: align pinned dataset metadata
bench: bind generation evidence to runtime protocol
bench: bind Modal measurement provenance
bench: bind Qwen3.6 launch evidence to runtime identity
bench: canonicalize campaign route identities
bench: clean up failed server startups
bench: close operation evidence review gaps
bench: discard partial server results
bench: emit strict quality JSON
bench: fail closed on benchmark operation identity
bench: forward only encode/score to EvalRunner in quality path
bench: harden generation quality sharding
bench: harden multimodal batch validation
bench: harden queue readiness for generation defaults
bench: harden shard failure capture
bench: keep image types out of runtime imports
bench: pin ragged PyLate reranking fix
bench: pin trusted dataset loaders
bench: preserve canonical FewRel dataset identity
bench: preserve generation evidence imports
bench: preserve invalid campaign archives
bench: preserve profile benchmark identity
bench: preserve safe Modal shard failures
bench: reject extraction item errors
bench: reject incompatible reference operations early
bench: reject incomplete shard assembly
bench: reject partial performance measurements
bench: remove stale generation import
bench: report empty perf evidence explicitly, not as ambiguous
bench: restore governed retrieval floors
bench: route profiled performance requests
bench: satisfy review static analysis
bench: start named profile routes
bench: stop plural options poisoning choice extraction
bench: type bounded image dispatch
bench: validate multimodal encode batches
bench: version routed rate candidates
billing: bind Modal tags to pretraffic contract
billing: bind zero-page job evidence
billing: preserve zero-page job evidence
billing: project observed units through reservation plans
billing: restrict zero terminals to pages
billing: verify multimodal fixture units
catalog: bind named output evidence
catalog: disable unsafe qwen speculative default
catalog: harden native acceptance validation
catalog: raise GTE Muvera repetitions
catalog: route mxbai edge through native attention
catalog: use pinned GTE query length
catalog: validate named profile evidence
ci: authenticate quality dependency sync
ci: diff quality evals from current PR base
ci: fail closed on mixed Modal output
ci: isolate privileged req14 completion
ci: isolate quality eval evidence
ci: isolate quality runner credentials
ci: let scalr-bootstrap track the opentofu pin
ci: repair Qwen quality shard workflow
ci: verify trusted req14 callback checkout
cloud: accept canonical Modal volume mount
cloud: accept Modal GCP provider identity
cloud: address release workflow review follow-ups
cloud: align launch evidence contracts
cloud: align release lock namespace
cloud: align release probes with scoped listing
cloud: allow release image manifest readback
cloud: allow runtime secret namespace
cloud: allow scoped release key discovery
cloud: allowlist release child environments
cloud: approve permissive launch licenses
cloud: attest LightOnOCR image compatibility
cloud: attest Modal snapshot memory limits
cloud: bind billing to executing Modal app
cloud: bind billing to Modal environment
cloud: bind commercial evidence to routed candidates
cloud: bind ECR login to OIDC identity
cloud: bind rate probes to archived model pins
cloud: bind score probes to campaign batches
cloud: bind topology plan semantics
cloud: bootstrap purpose-bound Slack sinks
cloud: bound and isolate release preflight reads
cloud: classify ECS start failures first
cloud: complete managed release operability follow-ups
cloud: complete release operability audit
cloud: dispatch exact catalog profiles
cloud: dispatch quality shards in Modal batch mode
cloud: fail closed in candidate image lookup
cloud: fail closed on ECR lookup errors
cloud: finish release child isolation
cloud: finish release preflight hardening
cloud: guard acceptance projection freshness
cloud: guard catalog projection freshness
cloud: harden commercial evidence joins
cloud: harden launch identity and approval invariants
cloud: harden launch rate campaign probes
cloud: harden managed release promotion
cloud: harden notification launch paths
cloud: harden screenshot acceptance
cloud: harden vision acceptance bounds
cloud: include smart chat task evidence
cloud: isolate launch credit approvals
cloud: keep LightOn preflight CPU-safe
cloud: make topology imports typecheck
cloud: map static lanes to catalog pools
cloud: measure both SigLIP towers
cloud: pin batch worker execution identity
cloud: pin dev worker identity
cloud: pin launch measurement resources
cloud: pin native worker execution identity
cloud: pin staging batch memory limit
cloud: preflight live release dependencies
cloud: preflight Modal SGLang runtime
cloud: preflight release task arguments
cloud: preserve concurrent topology output
cloud: preserve governed release tool paths
cloud: preserve sidecar image contract
cloud: prime catalog before gateway rollout
cloud: probe exact campaign request batches
cloud: reconcile commercial rates with launch profiles
cloud: reconcile measurement topology with launch catalog
cloud: refresh staging evidence pins
cloud: reject duplicate generation shards
cloud: reject malformed release metadata
cloud: reject mismatched catalog evidence
cloud: reject undecodable evidence
cloud: reject unversioned S3 evidence
cloud: require complete pii redaction proof
cloud: reseal licensed launch release locks
cloud: reseal production rate release locks
cloud: restore CUDA toolkit to SGLang sidecars
cloud: retain published measurement evidence
cloud: retire legacy environment ECR state
cloud: scope Better Stack credentials by product
cloud: separate guardrail policy fixtures
cloud: use lightweight generation evidence contract
cloud: validate launch evidence integrity
cloud: verify control-plane runtime dependencies
cloud: verify Modal markers from files
cloud: verify redact pii acceptance
cloud: verify redact PII postprocessing
deploy: pin internal scalr workspaces to opentofu
gateway: validate schema keywords by context
generation: close structured grammar review gaps
grammar: bound pointer index parsing
grammar: harden schema normalization
loadtest: manage github-token in terraform + sync via ESO, pin sie-perf-lab ref
loadtest: persist github-token in terraform, sync via ESO, pin perf-lab ref
loadtest: remove illegal variable interpolation from output description
model-ops: close review gaps in the AUTO verdict contract
model-ops: require an explicit matrix_task key in auto_authoring
profiles: honor routed outputs and retrieval recipes
quality-eval: address Copilot review on the wall-clock stop
quality-eval: remove shadow tree on exception paths, not just returns
quality-eval: select governed official metric
quality-eval: serve non-default smoke bundles from a pip —target shadow so the overlay reaches the bundle floor
quality-eval: validate wall_budget_minutes inside run_pipeline
release: authenticate gateway rollback probe
release: keep sie-audio-prep internal to builds
sdk: accept nullable grammar metadata
sdk: keep grammar types on supported module path
security: include aws edge in standing review
security: pin hf_revision on the 10 unpinned trust_remote_code models
server: apply trained Rotary ColBERT projection
server: defer LightOn OCR compatibility patch
server: hide unsupported visual muvera profiles
server: mask ColBERT expansion keys in flash attention
server: meter SigLIP tokens from attention masks
server: pin launch guardrail policies
server: preserve stream helper compatibility
server: preserve trusted code revision on fallback
server: restore ColBERTv2 retrieval recipe
server: restore Jina ColBERT retrieval recipe
server: serialize remaining tokenizer access
sie_server: register remote-code snapshot dir on sys.path for ST-Router models
sie-bench: isolate PyLate reference evaluation
sie-bench: preserve expansion token attention
smoke: fail loud when a bundle overlay undershoots its declared pins
terraform: manage github-token container only, drop placeholder version
tooling: derive overlay cache version from lock
tooling: gate Modal quality shard fanout
tooling: harden launch failure handling
tooling: isolate Modal shell worktree environment
tooling: make empty Modal status succeed
tooling: mount tracked playbook fixtures on Modal
tooling: namespace Modal JIT caches by ABI
tooling: preserve Modal status errors
tooling: preserve multiline Modal commands
tooling: read Modal lock from mounted workspace
tooling: resolve modal from task path
tooling: scope Modal inventory conflicts
tooling: validate Modal description types
tooling: version Modal datasets cache
vision: harden launch evidence
vision: pad fused detection batches

Performance Improvements

catalog: bind bounded extraction evidence
catalog: bind current nonvision launch evidence
generation: recover granite guardian queue evidence
vision: reuse siglip preprocessing

v0.6.22 (2026-07-22)

Highlights

New capabilities: bind exact rate measurement windows; shard OCR quality by page; support package revision evidence; attest worker execution identity; fan out quality shards on Modal; pin immutable Docling artifacts
Reliability and operations: harden OCR shard provenance; harden package revision evidence; fail fast on managed gateway lock drift; attest OCR shard execution identity; clarify complete rate evidence
Performance: concurrent OCR page dispatch so server batching actually engages

Features

bench: bind exact rate measurement windows
bench: shard OCR quality by page
bench: support package revision evidence
cloud: attest worker execution identity
cloud: fan out quality shards on Modal
server: pin immutable Docling artifacts
server: support immutable Docling model artifacts
tooling: attach named Modal shell secrets

Bug Fixes

bench: attest OCR shard execution identity
bench: clarify complete rate evidence
bench: clarify rate-window mismatch errors
bench: harden OCR shard provenance
bench: harden package revision evidence
bench: mark serving basis non-commercial
bench: require exact rate resource evidence
ci: install sdk deps for catalog tests
ci: queue release rerun when main moves
ci: regenerate launch catalog on release PR
ci: regenerate release PR when main moves
cloud: bind benchmark evidence to deployed workers
cloud: bind benchmark model identity
cloud: bind quiesced ECS candidates before scaling
cloud: bind reviewed catalog recipe executors
cloud: bind worker resource attestation
cloud: fail fast on managed gateway lock drift
cloud: include compiler version dependency
cloud: isolate catalog compiler dependencies
cloud: meter detection catalog smoke images
cloud: preserve deployment validation ordering
cloud: preserve identity across billing handoff
cloud: reject noninteger GPU identity counts
cloud: remove member key spend ceiling
cloud: require clean governed release sources
cloud: validate integral worker identity fields
cloud: validate managed gateway lock before release
eval: constrain temporary artifact root
eval: serve canonical profile routes
tooling: bind release locks from clean head
tooling: close Modal fanout lifecycle races
tooling: fail closed on Modal volume sync
tooling: isolate Modal uv cache
tooling: isolate quality shards on Modal v2
tooling: keep active modal shells alive
tooling: preserve Modal fanout failures
tooling: reject symlink fanout outputs
tooling: resolve modal in detached shells

Performance Improvements

sie-bench: concurrent OCR page dispatch so server batching actually engages

v0.6.21 (2026-07-21)

Highlights

New capabilities: add rate-grade evidence contract; add reusable local queue stack; add upstream GLiNER NER reference; govern vision launch evidence; seed load-test request selection; SIESparseImageWrapper for served visual sparse retrieval
Reliability and operations: capture generation timeout evidence; harden local queue stack lifecycle; drop stale metrics registry dependency; finalize billing money hardening; harden account lifecycle rehearsal
Performance: record SPLADE 512-token run; record extraction L4 measurements; compact sparse output in native precision; fuse packed SPLADE pooling

Features

bench: add rate-grade evidence contract
bench: add reusable local queue stack
bench: add upstream GLiNER NER reference
bench: govern vision launch evidence
bench: seed load-test request selection
bench: SIESparseImageWrapper for served visual sparse retrieval
billing: meter multimodal catalog units
candle: add and optimize native SPLADE sparse encoding
candle: add native SPLADE sparse encoding
cloud: add exact qwen35 perf probe
cloud: add fail-closed catalog acceptance gate #1943
cloud: add fail-closed catalog acceptance runner
cloud: add fast chat catalog acceptance
cloud: add managed-edge abuse and tenant gates
cloud: add OCR and speech acceptance probes
cloud: add operator key administration
cloud: add operator key administration and keyed usage
cloud: add OWLv2 acceptance executor
cloud: add OWLv2 catalog acceptance
cloud: add staging account lifecycle rehearsal
cloud: add staging rehearsal rate authority
cloud: audit measured launch rate-book coverage
cloud: audit measured rate book coverage
cloud: bind verified DINO and Whisper evidence
cloud: declare exact lane placement resources
cloud: declare exact lane resources
cloud: enforce conserved usage rating
cloud: enforce exact managed catalog authority
cloud: execute text acceptance recipes
cloud: exercise DINO catalog acceptance
cloud: expose operator billing health
cloud: gate catalog tasks on evidence
cloud: govern launch catalog topology
cloud: govern vision launch catalog
cloud: integrate control-plane launch readiness
cloud: integrate exact request billing authority
cloud: notify operators of new signups
cloud: pack Whisper into shared launch pool
cloud: record generation launch evidence
cloud: record Qwen3.5 launch evidence
cloud: rotate staging rehearsal rate book
cloud: serve release-locked model catalog
cloud: serve release-locked model catalog #1944
cloud: settle exact pair and generation usage
cloud: ship proprietary vision launch catalog
cloud: validate extraction catalog recipes
cloud: validate structured output acceptance
cloud: verify qwen35 A100 model evidence
cloud: wire privacy-safe console RUM
cluster: opt-in bake cache write via SIE_BAKE_CACHE_WRITE
core: expose authoritative request usage
core: report authoritative score pair usage
evidence: propagate worker execution identity
extract: add pinned PII quality gate
gateway: add strict Cohere rerank compatibility
model-ops: feasibility accepts transformers-5-line bundle choice
model-ops: scaffold re-renders cloud release locks with the model YAML
models: onboard Snowflake/snowflake-arctic-embed-s (scaffolded)
models: raise qwen36 launch context to 8k
playbooks: upstream SparseEncoder reference arm for sparse-visual models
quality-eval: wire the signature-DB loop — adjudicated-skip + verdict write-back
reranking: aggregate launch runtime and release locks
reranking: harden Qwen score runtime
sdk: expose terminal billing metadata
sdk: preserve terminal request metadata on errors
serve visual-SPLADE via transformers5 (sparse-vision adapter + measurement path)
server: bind package artifact identity
server: bind package-backed model artifacts
server: transformers514 bundle + SparseEncoder vision sparse adapter
sie-bench: add generation quality sharding
sie-bench: add generic generation quality sharding
telemetry: centralize observability on OTLP
vision: harden OSS runtime contracts
vision: ship OSS launch runtime and evidence

Bug Fixes

audio: sandbox image builds the audio wheel opportunistically
bench: align local generation lane routing
bench: attest generation execution revision
bench: bind broker artifact ARN
bench: capture generation timeout evidence
bench: clarify direct revision semantics
bench: close execution identity provenance gaps
bench: exercise generation queue sidecar
bench: harden local queue stack lifecycle
bench: isolate generation smoke stack lifecycle
bench: isolate queue smoke ports
bench: make fake stack no-op explicit
bench: preserve ChemProt entity offsets
bench: preserve exact reranker task identity
bench: preserve NER task parameters
bench: reject coerced execution identity
bench: remove failure log content
bench: render generation smoke help
bench: resolve canonical AG News dataset
bench: route canonical model profiles
bench: select encode output families
bench: separate model and machine profiles
bench: serialize image rerank inputs
bench: use SPIA field labels for PII eval
bench: verify broker artifact bucket before reads
bench: warm generation at measured concurrency
billing: preserve legacy outbox drain
billing: reject unexpected metronome rates
billing: verify vendor-bound raw usage
candle: bound CUDA compiler parallelism
candle: include CUDA math constants for SPLADE
candle: preserve tiny SPLADE log1p weights
ci: bind cloud gateway cache to OSS source
ci: install catalog compiler dependency
ci: load sidecar image for release smoke tests
ci: make cloud gateway checks deterministic
ci: supersede obsolete pull request runs
ci: validate release guard wait settings
ci: wait for release check reruns
cloud: accept numeric whisper fixture transcript
cloud: address billing review findings
cloud: address rate authority review
cloud: align acceptance with native labels
cloud: align rehearsal rates with measured audit
cloud: benchmark qwen35 through native generate
cloud: bind acceptance hash to source artifact
cloud: bind exact lane cpu limits
cloud: bind generation grammar pool
cloud: bind structured schema semantics
cloud: bound outbox poison isolation
cloud: clarify idempotent deletion tests
cloud: clean up metered rehearsal credits
cloud: close console release preflight gaps
cloud: close launch integration review gaps
cloud: close launch review gaps
cloud: close operator key launch gaps
cloud: converge managed catalog exactly
cloud: cover multimodal rehearsal units
cloud: declare redact PII labels
cloud: derive migration log stream
cloud: drop stale metrics registry dependency
cloud: enforce billing activation invariants
cloud: finalize billing money hardening
cloud: govern legacy billing cutover
cloud: harden account lifecycle rehearsal
cloud: harden billing money flows
cloud: harden generation acceptance evidence
cloud: harden launch rate evidence
cloud: harden lifecycle rehearsal cleanup
cloud: harden signup alert routing
cloud: harden Stripe clawbacks and usage delivery
cloud: include registry compiler dependency
cloud: keep cleanup accounts suspended
cloud: keep live secret values out of state
cloud: keep proxy dry-run provenance current
cloud: keep qwen35 measurements pending
cloud: keep qwen35 performance pending
cloud: match Better Stack canonical alert fields
cloud: pin exact console release commits
cloud: preserve charged structured failures
cloud: preserve committed key recovery
cloud: preserve custom catalog evidence
cloud: preserve proxy token shape in dry-run
cloud: rebase catalog acceptance gate
cloud: reconcile exact metronome rating
cloud: reconcile native vision label bindings
cloud: refresh release lock artifact pins
cloud: reject duplicate billing bindings
cloud: reject empty key scopes
cloud: reject malformed catalog exports
cloud: reject unknown extraction recipes
cloud: reject unsupported lane providers
cloud: report gross Stripe funding
cloud: require console release domain
cloud: require ordered whisper transcript
cloud: restore Metronome billing authority
cloud: restore telemetry branch CI
cloud: retain charged acceptance failures
cloud: tighten billing terminal invariants
cloud: tolerate inaccessible gateway candidates
cloud: validate acceptance readiness reasons
cloud: validate edge origin before secret sync
cloud: validate lazy snapshot lanes
cloud: validate modal cloud placement
cloud: validate native label binding kinds
cloud: validate rate tariff shape
colpali: serialize model forwards — transformers output-recorder race
colpali: stop GPU memory accumulation on Vidore evals
control-plane: make wallet authority explicit
core: preserve unit count wire order
deploy: bound generation readiness process
deploy: harden generation readiness gate
deploy: honor local weight source precedence
deploy: prove exact generation revision readiness
deploy: resolve model staging weight sources
eval: add faithful Qwen3-VL reference
eval: bound multimodal reference scoring
eval: bound reranker candidates by corpus
eval: isolate VL reference dependencies
eval: keep gate plan aligned with dispatch
eval: preserve VL text coverage
eval: reject empty reranker qrels
evidence: bind immutable model identity
evidence: close worker identity contract gaps
evidence: harden worker identity boundaries
extract: close review validation gaps
extract: filter unconstrained relation endpoints
keda: keep gateway coordination under autoscaling
keda: simplify forward OTLP upgrade
keda: support forward-only OTLP upgrade
keda: verify lean forward upgrade
keep audio prep release pin in sync
loadtest: skip doomed bench-only scenarios when no release deployed
metering: reject invalid score counts
model-ops: align exact vision gate evidence
model-ops: gate parses the trigger’s ‘Triggered by #N’ issue reference
model-ops: void gate attempts that provably began no spend
model: restore SPLADE quality sequence cap
models: pin hf_revision for gte-Qwen2-7B remote-code load
models: restore SPLADE 512-token capacity
models: serve gte-Qwen2-7B via its own bidirectional forward
playbooks: v2-native task extraction + timeout contract for the sparse arm
quality-eval: identity-based signature dedup + branch hash (#2096 review)
quality-eval: make runner wake parallel, fail-loud, and demand-accurate
quality-eval: reset stale WakeRequestedAt against launch time
quality-eval: tolerate YAML date-typed fields in signature identity (#2096 review)
quality: tolerate missing gpu identity tooling
refresh managed release locks
release: refresh deployment locks during release regeneration
release: run deployment lock refresh in project env
reranking: address exact-head review
reranking: emit authoritative fake usage
reranking: emit fake score usage
reranking: reject blank compatibility inputs
server: empty bundle model_filter must advertise zero models, not all
server: remove unused fake adapter import
sie-bench: harden generation sharding evidence
sie-tools: keep session failures auth-scoped
sie-tools: preserve account lifecycle errors
smoke: align OTel collector contract checks
smoke: align OTLP metric assertions
smoke: restore kind smoke on main
splade: align telemetry and model docs
splade: harden profile and sparse defaults
tooling: fail fast before Modal dataplane runs
tooling: stop modal shells by app id
tools: preserve baked mise in modal shells
vision: address review edge cases
worker: propagate effective sequence caps
worker: release OOM tracebacks in the no-recovery branch too (CodeRabbit)

Performance Improvements

benchmarks: record SPLADE 512-token run
bench: record extraction L4 measurements
candle: compact sparse output in native precision
candle: fuse packed SPLADE pooling
candle: pack SPLADE BERT inference
vision: bulk-convert detection outputs

v0.6.20 (2026-07-18)

Highlights

New capabilities: add GLiGuard and fix extraction catalog behavior; add GLiGuard and fix LightOnOCR and GLiREL extraction; add cloud benchmark campaigns; add native R3-Skill evaluation; add staging cloud smoke path; add two-client staging rehearsal
Reliability and operations: fence broker rollback recovery; harden cloud campaign evidence; harden cloud campaign readiness; harden live campaign preflight; harden managed campaign control plane
Performance: cache exact bf16 gelu values; fuse ModernBERT exact GeGLU; optimize ModernBERT CUDA hot path; vectorize bf16 gelu lookup gate

Features

add GLiGuard and fix extraction catalog behavior
add GLiGuard and fix LightOnOCR and GLiREL extraction
bench: add cloud benchmark campaigns
bench: add native R3-Skill evaluation
bench: add staging cloud smoke path
bench: add two-client staging rehearsal
bench: complete managed cloud campaigns
bench: cover cloud task modalities
bench: inspect diagnostic campaign archives
bench: operationalize cloud measurement campaigns
bench: operationalize managed SIE Cloud measurement campaigns
ci: fake-stack-test job — Fake Engine regression suite + Cloud dispatcher loopback
ci: fake-stack-test job + five Fake Engine regression tests + Cloud loopback
cloud-gateway: serve job chunk results via signed capability URLs
cloud-gateway: serve offload job results via signed capability URLs
cloud: apex-domain cutover + console deploy model
cloud: control-plane audit-read + org-settings + residency-read endpoints (console Tier-2)
cloud: export worker-lane OTLP traces to the collector (lane-only, #1740)
cloud: gateway OTLP trace+metric export — clean end-to-end telemetry spine
cloud: govern Modal deployments with typed specs
cloud: per-org alert-threshold toggles (0006 alert_prefs) gating _fire_alerts
cloud: per-org console service keys via /internal/console-keys
cloud: per-org console service keys via POST /internal/console-keys
cloud: relocate the schedule runner in-VPC (EventBridge -> /internal/schedules/run-due)
cloud: resolve fresh keys on-demand to kill the ~5-10s gateway 401 window
cloud: surface the saved card’s brand/last4/exp (Stripe payment-method retrieve)
cloud: WP4 billing-depth control-plane endpoints (invoices, usage, alert toggles, payment method)
cloud: WP4 billing-depth endpoints — invoices, usage series, alert-config toggles
control-plane: make auto-recharge fire end-to-end
mac-smoke: additive fake phase 0 — weightless surface contract
model-ops: add stage-5 quality+perf onboarding gate
model-ops: add-model trigger — labeled issue starts the onboarding agent
model-ops: stage-5 quality+perf onboarding gate — measure vs target, bounded fix loop, evidence packet
models: add Tencent R3-Skill support
models: onboard intfloat/multilingual-e5-small (scaffolded)
quality-eval: #1810 experiment runner — parallel Modal sessions, GPU selection, budget backstop
quality-eval: autonomous experiment runner — parallel Modal sessions, GPU selection, budget backstop (Req14 v2 3/4)
quality-eval: end-to-end pipeline wiring + dedup/budget semantics (Req14 v2 4/4)
quality-eval: wire triage→planner→runner→autofix into a bounded pipeline
server: add deterministic weightless fake adapter family
server: failure injection for the fake adapter family
server: fake adapter family — deterministic weightless sie-fake models
server: synthetic memory-tracker mode — deterministic pressure/eviction
server: synthetic memory-tracker mode for deterministic eviction
support Iso ModernColBERT on Candle
tests: lost-result 120s characterization on the full queue topology
tests: lost-result characterization on the full queue topology

Bug Fixes

bench: accept canonical UTF-8 broker artifacts
bench: accept omitted optional request refs
bench: align broker contract boundary nesting
bench: align broker log retention
bench: align campaign runtime dependencies
bench: avoid imagebuilder template collision
bench: bind ami runtime kernel evidence
bench: bound R3 quality regression gate
bench: bound staging smoke below service knee
bench: close campaign review and ci gaps
bench: close campaign review edge cases
bench: expose safe broker failure diagnostics
bench: externalize broker deployment settings
bench: fence broker rollback recovery
bench: harden cloud campaign evidence
bench: harden cloud campaign readiness
bench: harden live campaign preflight
bench: harden managed campaign control plane
bench: make SSM dispatch at most once
bench: pass authorized request to orchestrator
bench: pin runnable worker manifests
bench: prepare campaign clients before start
bench: preserve broker terminal state
bench: preserve cloud campaign timing evidence
bench: preserve Fargate scratch permissions
bench: probe exact start control key
bench: publish worker image to owned ecr path
bench: reject malformed canonical artifacts
bench: relocate runtime caches to scratch
bench: remove stale quality import
bench: retain ami assertion diagnostics
bench: retry transient ECS role propagation
bench: retry transient R3 score timeouts
bench: roll task revisions without downtime
bench: secure ephemeral identity creation
bench: secure staging cloud validation
bench: skip absent profile deletion
bench: support repeated maintenance pauses
bench: trust scheduler group for worker expiry
bench: validate R3 evaluation inputs
candle: close ModernBERT audit gaps
ci: disable implicit cloud tool installs
ci: pin nested mise runtime
ci: verify cloud IaC toolchain
cloud-gateway: complete loopback class-fix via one canonical helper; align Python mirror + docs (#1906 final review)
cloud-gateway: IP-literal loopback check + recover stuck payload-volume reload (CodeRabbit #1906 r2)
cloud-gateway: reject non-ASCII signature hex without panicking
cloud-gateway: validate capability-URL origins, bound TTL, and fix object-store minting (CodeRabbit #1906)
cloud: address CodeRabbit — same-item batch id pairing + metrics protocol env + /metrics deny log
cloud: address CodeRabbit — serialize env-mutating metrics-protocol test with a module-local ENV_LOCK
cloud: address follow-up review
cloud: address Modal deployment review
cloud: alert on zero-metering invoice drift; reject ambiguous usage_events apportionment
cloud: align dev backend benchmark
cloud: align vendor billing with ledger debits
cloud: batch/offload jobs retry retryable per-item MODEL_LOADING instead of failing the chunk
cloud: bound on-miss key resolver’s single-flight map + harden lookup
cloud: CodeRabbit CP review — alert-configs PUT 404s, invoice-payload 502 guard, ty-ignore format
cloud: console SIE_GATEWAY_URL is the public edge, not the raw Modal URL
cloud: customer-scope the saved-card lookup — close cross-tenant card-PII leak
cloud: deploy otel-collector + merge OTLP endpoint before lane snapshot so lane tracing activates
cloud: derive outbox credits from the ledger debit and alert on reconciliation drift
cloud: diagnose stale-checkout CP destroy + bridge VERCEL_API_TOKEN
cloud: diagnose stale-checkout CP destroy + bridge VERCEL_API_TOKEN in cloud-provision
cloud: don’t persist raw billing_email in audit_log detail (PII)
cloud: final WP4 review — strict invoice currency, atomic alert toggle, UTC-naive parse, card guard
cloud: gateway OTLP export — blocking reqwest client for span batch processor + HTTP/TLS metric transport
cloud: gateway OTLP export — Tokio runtime for span batch processor + HTTP/TLS metric exporter
cloud: harden console-key lifecycle
cloud: harden whole-estate finalization
cloud: hide console keys from /me/usage + /me revoke (security review)
cloud: preserve billing alert and meter attribution
cloud: refresh recovery handle on batch chunk re-spawn (avoid orphaned container + 409 spam)
cloud: release the scheduler lock_conn on any pre-hand-off failure
cloud: stage the gen-lane model in model-stage and env-qualify the unstaged-weights hint
cloud: stop deploy from dirtying satellite uv.locks
cloud: stop deploy from dirtying satellite uv.locks (clean-tree guard)
cloud: WP4 review follow-ups — usage upper bound, invoice-fallback diagnosability, card exp contract
control-plane: hand auto-recharge sweep off the EventBridge request path
control-plane: harden auto-recharge collection
control-plane: make recharge retries lease-safe
fake-engine: address two-axis review findings across the stack
fake-engine: CI loopback interop + CodeRabbit review findings
gateway: attest deployed execution revision
gateway: include unanimous error_code in all_items_failed 500 logs
gateway: reject malformed queue-request bodies with a typed 400
harden classification quality verification
model-ops: apply CodeRabbit review findings on the stage-5 gate
model-ops: apply review findings on the friction PR
model-ops: deterministic escalation when the agent itself crashes
model-ops: filter vanilla floor material to fleet metric families
model-ops: harden add-model trigger per review
model-ops: raise directly in _read_json per code-quality finding
model-ops: read vanilla numbers from the verdict’s evidence field
model-ops: rehearsal follow-ups — agent guardrail, record valves, auto-review, log-tail noise
model-ops: rehearsal follow-ups — guardrails, record valves, log-tail noise
model-ops: scaffolder crash on present-but-null sibling fields
model-ops: seed-floor emits fleet metric families, not the raw mteb vector
model-ops: stage-5 gate read vanilla numbers from verdict evidence, not probe measurements
model-ops: tighten the agent-crash fallback per review
ocr: address review lifecycle findings
preserve parameterized quality gate targets
quality-eval: accurate backstop messages for CPU playbooks (review S1)
quality-eval: guard non-dict cluster signals in the pipeline (review)
quality-eval: harden experiment runner per PR #1883 review (traversal, stale-verdict, rollback, build-clock)
quality-eval: harden pipeline per PR #1891 review (report/verdict/workflow)
quality-eval: reap process groups in rollback to avoid zombies (re-review)
quality-eval: reject non-object verdict JSON in escalate_verdict (re-review)
quality: address R3 review feedback
quality: complete R3 faithful gate workflow
resolve file-backed adapter impacts
resolve onboarding bundle metadata
resolve query-qualified quality targets
reuse resolved uv in vanilla probe
server: allow dependency-free model selections
server: recommend bundle for full selection
server: support cloud catalogs in bundle checks
server: validate LightOnOCR bundle runtime

Performance Improvements

candle: cache exact bf16 gelu values
candle: fuse ModernBERT exact GeGLU
candle: optimize ModernBERT CUDA hot path
candle: vectorize bf16 gelu lookup gate
cloud: index usage_events_outbox (account_id, occurred_at) for the usage read (0007)
ocr: serve generative models with SGLang
ocr: serve generative OCR with SGLang continuous batching

v0.6.19 (2026-07-14)

Highlights

New capabilities: add catalog-smoke step — per-model serve_i6pn load verification; add env-specific Modal deploy provisioning; align prod-us manifest to latest i6pn fast-path topology; commit generated price book so the provisioner locates it; deploy-both IaC hardening — env-derived Modal env + env-targeted B6 smoke; event-driven fast-lane promotion + credit ramp
Reliability and operations: drop the speculative mise install retry chain; harden deploy tooling — modal —json ANSI, async snapshot-commit race, Vercel sensitive-var upsert; harden lane_deploy driver — modal —json ANSI + async snapshot-commit race; retry server launch on startup death; scope the port sweep to our serve chain and harden the start guard

Features

cloud: add catalog-smoke step — per-model serve_i6pn load verification
cloud: add env-specific Modal deploy provisioning
cloud: align prod-us manifest to latest i6pn fast-path topology
cloud: commit generated price book so the provisioner locates it
cloud: deploy-both IaC hardening — env-derived Modal env + env-targeted B6 smoke
cloud: event-driven fast-lane promotion + credit ramp
cloud: normalize the OSS terminal failed readiness state to MODEL_LOAD_FAILED on the Modal lane
cloud: productionize the i6pn realtime dispatch fast path end-to-end
cloud: readiness-gated i6pn dispatch — thread loaded_models to the wire, gate cold workers
model-ops: deterministic add-a-model scaffolder — verdict → YAML + matrix + target
model-ops: deterministic add-a-model scaffolder + onboarding runbook
model-ops: structured feasibility verdict — eval-model stage-2 gate
model-ops: structured feasibility verdict for add-a-model stage 2
quality-eval: add fix-proposal planner (tiered Fable-5) + experiment-plan schema
quality-eval: bounded auto-fix — agent PRs for floor re-baselines, human merge gate (Req 14 stage 4)
quality-eval: bounded auto-fix CLI for floor re-baseline
quality-eval: fix-proposal planner (tiered Fable-5) + experiment-plan schema (Req14 v2 2/4)
quality-eval: make the planner reason from signals + consume the LLM’s predicted evidence
quality-eval: onboarding smoke playbook — served-route bring-up on Modal
quality-eval: provenance-gated preflight + vanilla-lane floor template
quality-eval: provenance-gated preflight + vanilla-lane floor template (Req14 v2 1/4)

Bug Fixes

ci: drop the speculative mise install retry chain
ci: give the scheduled quality run its own concurrency group
ci: stop rolling the un-retried rustup ensure in mac-smoke-nightly
cloud: address catalog-smoke review — gate internal token, harden revoke + warm-up bound
cloud: address CodeRabbit round-2 review on #1782
cloud: address env deploy review comments
cloud: address PR review comments
cloud: bound wedged model loads + typed dead-letter on the lazy lane
cloud: bound/adapt slab credit holds so a funded org stops self-wedging at 402
cloud: catalog-smoke presents SIE_API_KEY when the gateway enforces managed keys
cloud: don’t send type on Vercel env update (sensitive-var 400)
cloud: drop candle-only bge-large-en from the default L4 lane’s models_served
cloud: gateway must not 401 valid keys during cold-replica key-snapshot warmup
cloud: gen-lane warm floor + gateway-config secret merge in deploy-edge
cloud: harden deploy tooling — modal —json ANSI, async snapshot-commit race, Vercel sensitive-var upsert
cloud: harden lane_deploy driver — modal —json ANSI + async snapshot-commit race
cloud: heartbeat TTL for i6pn discovery keys (corpse expiry on hard stop)
cloud: narrow B6 local gateway bundle to the deployed models_served (#1771 hash parity)
cloud: pg-connector smoke builds ModalLaneExecutor from SIE_LANE_APP_MAP
cloud: pg-connector smoke builds ModalLaneExecutor from SIE_LANE_APP_MAP (#1782 missed caller)
cloud: scope _modal_cli TERM=dumb to captured calls only (CodeRabbit #1807)
cloud: skip B6 offload-batch in external-CP mode (unshareable payload store)
cloud: stage refs/main alias so offline lanes resolve pinned-SHA models
cloud: stop provisioning unbacked Stripe metered prices
cloud: stop provisioning unbacked Stripe metered prices (prepaid-credits + Metronome rating is the model)
cloud: tighten edge deploy conflict resolution
cloud: Vercel env update sends value only (sensitive-var target 400)
cloud: warm managed key before catalog-smoke sweep to beat the key-propagation race
cloud: wire the default gen lane into the gateway census + app map
gateway: bump yanked spin 0.10.0 to 0.10.1 in Cargo.lock
mac-smoke nightly score-path contract + unknown-model 500s
mac-smoke: call /v1/score with the slash-form model id
mac-smoke: kill stale server by port between phases
mac-smoke: retry server launch on startup death
mac-smoke: scope the port sweep to our serve chain and harden the start guard
model-ops: harden feasibility validator per review
model-ops: harden scaffolder per review
poc: pass explicit github_token to claude-code-action
quality-eval: charset-gate vanilla_vector metric keys (review C1)
quality-eval: classify malformed encode responses as REFUTED, not a crash
quality-eval: controlled exit on planner output-write failure (CodeRabbit)
quality-eval: harden autofix per CodeRabbit review round 2
quality-eval: harden autofix preflight against output injection (review round 1)
quality-eval: require string cell fields before charset validation
server: colbert-small multivector encode returns one vector-set per input, worker-item order
server: Florence-2 processor compat with transformers 4.57
server: guard cross_encoder predict tokenization against the concurrent-metering race (class-fix for #1800)
server: load tokenizer by pinned revision so offline chat-template rendering doesn’t need refs/main
server: return 404 instead of 500 for unknown models on inference endpoints
server: terminal Failed readiness + Florence-2 processor compat + pin launch-set revisions
server: terminal Failed readiness state so the sidecar dead-letters permanently-failed model loads

v0.6.18 (2026-07-12)

Highlights

New capabilities: add flat→per-account Files-store migration; co-located AWS us-east-1 reverse-proxy edge for api.superlinked.com; consolidate data-plane deploy + manifest-drive proxy-auth/i6pn; namespace Files-store objects by account; opt-in Modal proxy-auth layer for config/lane/OTLP ingress; parameterize Modal environment + prod-us tf root by SIE env (§B2)
Reliability and operations: fail fast on terraform plan errors before the destroy-check; bake SIE_MODAL_PROXY_AUTH into image env so the proxy-auth knob doesn’t diverge local↔remote; make proxy_auth_env authoritative in both directions; pin npm for TypeScript release publishing; aws-edge S3 backend uses native state locking (use_lockfile)

Features

cloud: add flat→per-account Files-store migration
cloud: co-located AWS us-east-1 reverse-proxy edge for api.superlinked.com
cloud: consolidate data-plane deploy + manifest-drive proxy-auth/i6pn
cloud: namespace Files-store objects by account
cloud: opt-in Modal proxy-auth layer for config/lane/OTLP ingress
cloud: parameterize Modal environment + prod-us tf root by SIE env (§B2)
cloud: per-account Files-store namespacing + migration
cloud: prod bring-up terraform prerequisites
cloud: prod bring-up terraform prerequisites — prod-us auth-token outputs, aws-edge prod state
cloud: prod enablement — aws-edge provisioner step, prod-us root, Modal-env parameterization
cloud: prod-us terraform root — env-isolated control plane, separate state bucket
cloud: reconcile OTLP transport + proxy-auth sender headers
cloud: stand up the AWS edge via cloud-provision
dispatcher: full IP-pin for object-store + postgres verify-full egress
dispatcher: org-ownership assertion on upload:// ends — defense in depth behind the gateway check
dispatcher: resolve-then-pin connector egress — SSRF TOCTOU backstop

Bug Fixes

ci: fail fast on terraform plan errors before the destroy-check
ci: pin npm for TypeScript release publishing
cloud: address CodeRabbit review on the prod-enablement PR
cloud: aws-edge S3 backend uses native state locking (use_lockfile)
cloud: bake SIE_MODAL_PROXY_AUTH into image env so the proxy-auth knob doesn’t diverge local↔remote
cloud: collision-aware Files migration move
cloud: correct console callback route + guard the CP terraform apply against silent destroys
cloud: correct scheduler/slo-tuner DSN-secret comments + enforce non-ingress secret refs in deploy-posture audit
cloud: default Metronome credit type to fiat USD-cents so low-balance alerts and grants provision
cloud: gateway payload store must be a local path, not the s3 payload_store_url
cloud: gateway payload store must be a local path, not the s3 payload_store_url (#1743 follow-through)
cloud: gateway ships the cloud-storage backend; provisioner wires SIE_CONFIG_SERVICE_URL
cloud: guard config-URL resolver against a malformed manifest
cloud: harden Phase-3 CI + sibling-sweep per adversarial review (Refs #1757 #1758)
cloud: make proxy_auth_env authoritative in both directions
cloud: parameterize the deploy-env slug in managed Modal app names
cloud: parameterize the OTLP collector app name with the deploy-env slug
cloud: provision §7.6 prepaid credit-pack price + STRIPE_PRICE_ID
cloud: redact all secret-bearing Config fields in Debug
cloud: redact ModalProxyToken Debug + cover /gen handshake forwarding
cloud: resolve gateway app name from gateway_app.APP_NAME, not ctx.env
dispatcher: address CodeRabbit review on the #1742 egress pins
dispatcher: harden connector egress pin against CodeRabbit findings
stabilize tilt e2e regressions
use pinned cargo-zigbuild in tilt builds

v0.6.17 (2026-07-09)

Highlights

New capabilities: add GTE RoPE embedding kernel; CLI WorkOS device-flow auth — retire the shared-token shortcut for customer commands; SIE Cloud managed service — AWS control plane, Modal data plane, provisioning + billing kernel; OSS amendments from the SIE Cloud build — engine, SDKs, gateway/sidecar seams; add deterministic regression-ledger detector for scheduled quality runs; add floor-archaeology playbook (CPU-only)
Reliability and operations: retry model-dependent calls on MODEL_LOADING; harden regression-ledger live mode and audit input validation; harden multi-gpu worker routing; compile fused gated activation op; import CUDA pointer trait

Features

candle: add GTE RoPE embedding kernel
cloud: CLI WorkOS device-flow auth — retire the shared-token shortcut for customer commands
cloud: SIE Cloud managed service — AWS control plane, Modal data plane, provisioning + billing kernel
oss: OSS amendments from the SIE Cloud build — engine, SDKs, gateway/sidecar seams
quality-eval: add deterministic regression-ledger detector for scheduled quality runs
quality-eval: add floor-archaeology playbook (CPU-only)
quality-eval: add Modal parity-probe playbook (diagnostic-only)
quality-eval: add official-recipe, fp32-ab, config-sweep, determinism playbooks
quality-eval: add report-only triage classifier + workflow for scheduled quality runs
quality-eval: add validation-playbook harness + hypothesis/verdict contract
quality-eval: report-only triage classifier + gated workflow for scheduled quality runs
quality-eval: seed triage signature DB and short-circuit known clusters
quality-eval: spread runner fleet across capacity pools with always-on floors
quality-eval: validation playbooks — adversarial repro harness (Req 14 Proj 1, stage 3)
server: support multi-gpu worker placement
worker: route by child queue pressure
worker: route multi-gpu sidecar children
worker: support multi-gpu sidecar children

Bug Fixes

candle: compile fused gated activation op
candle: import CUDA pointer trait
mac-smoke: retry model-dependent calls on MODEL_LOADING
quality-eval: address CodeRabbit + code-quality review comments
quality-eval: address CodeRabbit review findings
quality-eval: address triage review findings
quality-eval: detect EBS cache volume via lsblk model string
quality-eval: enforce budget cap on all eval playbooks + budget-stop artifacts
quality-eval: fp32 ST arm via throwaway -st model, not a profile
quality-eval: harden playbook verdict logic (review round 1)
quality-eval: harden regression-ledger live mode and audit input validation
quality-eval: harden triage input handling per CodeRabbit review
quality-eval: keep register-runners.sh bash-3.2 compatible
quality-eval: tolerate stop/start race in wake instance-stopped wait
quality-triage: share data-bearing baseline discovery with triage live mode
quality-triage: skip data-less scheduled runs in ledger baseline auto-discovery
req6: align candle chart after rebase
req6: harden multi-gpu worker review fixes
req6: harden multi-gpu worker routing
req6: resolve main rebase fallout
req6: wire pressure-aware multi-gpu routing
server: keep ipc ping alive on gpu health errors
sidecar: bound queue admission through scheduler completion
sidecar: refresh ready slots on cancel fanout failure
sidecar: release cancelled scheduler pressure
sidecar: wire local ingest through adapter pool

v0.6.16 (2026-07-07)

Highlights

New capabilities: opt-in template_as_prompt for ST prompt-length-aware templates; use_model_encode — delegate to checkpoint-native encode(); add Candle OOM pressure recovery; add Candle residency eviction; per-pool gpu_driver option + confidential H100 tracing example
Reliability and operations: harden splade flash gate + add flash-vs-native parity test; splade-v3 floors were cosine-epoch artifacts — re-baseline to dot self-measures + harden flash gate; restore executable bit on mac-smoke mise task; faithful ModernBERT forward + hardened 1_Dense loading; restore e5 query/passage templates on the multilingual-e5-large sentence_transformer profile

Features

adapters: opt-in template_as_prompt for ST prompt-length-aware templates
adapters: use_model_encode — delegate to checkpoint-native encode()
add Candle OOM pressure recovery
add Candle residency eviction
terraform/azure: per-pool gpu_driver option + confidential H100 tracing example

Bug Fixes

adapters: harden splade flash gate + add flash-vs-native parity test
adapters: honor caller instruction in use_model_encode without a template
address candle catalog review comments
address Candle residency review comments
align Candle idle eviction lifecycle
align candle workers with catalog lane config
bench: cap docling pubtabnet-html at 2000 pages to fit the job budget
bench: splade-v3 floors were cosine-epoch artifacts — re-baseline to dot self-measures + harden flash gate
bundles: pin flash-attn-4 prerelease so sglang 0.5.10.post1 resolves
ci: restore executable bit on mac-smoke mise task
colbert_modernbert_flash: apply trained ColBERT Dense projection (+ guard, tokenizer, re-baseline)
colbert_modernbert_flash: faithful ModernBERT forward + hardened 1_Dense loading
colbert: GTE muvera also at simhash 6 - CMedQA/MMarco evals OOM at dim 81920
colbert: mxbai muvera at simhash 6 (FDE dim 20480) - runner dies at 81920
colbert: retune GTE/mxbai muvera for the Dense-projected representation
colbert: serve pylate Dense chains and align fallback with flash
deploy: cap tester l4 lanes
deploy: zero and cap tester L4 lanes
deploy: zero AWS worker warm floor
deploy: zero tester worker warm lanes
gateway: fan out gpu-agnostic demand across pool profiles for scale-from-zero
gateway: normalize explicit GPU demand labels
helm: wake a deterministic owner lane on gpu-agnostic demand for multi-profile bundles
models: NV-Embed-v2 — serve via checkpoint-native encode() (latent-attention recipe), NFCorpus 0.3715→0.4479
models: restore e5 query/passage templates on the multilingual-e5-large sentence_transformer profile
models: restore e5 templates on multilingual-e5-large ST profile
models: serve NV-Embed-v2 via its native latent-attention ST path
preserve Candle idle evictor failure logs
quality-eval: fail fast when sie-server dies during readiness wait
quality-eval: unbreak 7B sglang lanes (flash-attn-4 resolution) + fail-fast readiness + docling job budget
restore kind smoke candle routing

v0.6.15 (2026-07-03)

Highlights

New capabilities: add Candle profile routing and catalog models; add candle runtime diagnostics; add candle xlm-roberta flash attention; add dev AMI bake script; add dev ec2 cleanup workflow; add dev EC2 launch script
Reliability and operations: harden native worker routing; harden native worker runtime; restore sidecar model readiness; delete Python NATS pull loop from worker hot path; harden SIE dev AMI workflow
Performance: add GEMM diagnostics switch; fuse xlm-r qkv projection; optimize bge-m3 xlm-r runtime path; enable candle reduced precision gemms

Features

add Candle profile routing and catalog models
add candle runtime diagnostics
add candle xlm-roberta flash attention
add dev AMI bake script
add dev ec2 cleanup workflow
add dev EC2 launch script
add SIE API key rotation tooling
candle: add native adapter runtime
dev: add personal SIE CUDA/Rust AMI workflow
document ssm ssh dev launcher
gateway: count dropped stale NAKs from abandoned attempts
helm: add opt-in OTLP tracing wiring + bundled collector to sie-cluster
helm: add optional bundled Tempo tracing backend
helm: add SIE tracing dashboard and tempo datasource uid
helm: opt-in OTLP tracing wiring + bundled collector in sie-cluster
helm: optional bundled Tempo tracing backend + Grafana datasource for sie-cluster
improve candle replica concurrency
improve dev AMI cache scalability
observability: bound tracing shutdown + trim env inputs across all four runtimes
observability: bound tracing shutdown and trim env inputs across all four runtimes
observability: gate Rust tracing on SIE_TRACING_ENABLED
observability: unify tracing enable switch across gateway, sidecar, and worker
quality-eval: gate score-on-retrieval candidate cells once floor committed
reattach dev cache volume
server: land Apple Silicon as a primary device onto main
server: land Apple Silicon as a primary device onto main (re-land of #1455)
sie_gateway: emit gateway.publish span on non-streaming queue-publish path
sie_server_rust: add OTLP distributed tracing
sie_server_rust: add OTLP distributed tracing to the Rust native worker
terraform: dedicated single-AZ node group for stateful observability
tilt: enable tracing with bundled Tempo
tracing: emit sidecar.dispatch span from sie_server_sidecar
tracing: inject queue trace context for non-generation endpoints
tracing: instrument sie_server_sidecar with OTLP (sidecar.dispatch span)
tracing: propagate trace context through the non-streaming worker loop
tune candle forward concurrency
tune dev EC2 launch defaults
wire dev AMI cache disk helpers

Bug Fixes

#1340: stella per-task instruction + re-baseline stale dense/colbert floors (blast radius was #1432)
adapters: guard empty-document MaxSim batches; strengthen single-doc test
address API key review feedback
address API key rollback review
address rust worker smoke review comments
address sidecar PR review feedback
align rust worker kind smoke wiring
align rust worker warm smoke behavior
align worker pool config hashes
back up API key secret before apply
candle: align cublaslt with candle cuda stream
candle: align readiness with loaded model state
candle: copy vendored cuda dependency in image
candle: harden native worker routing
candle: harden native worker runtime
candle: import cublaslt cuda extension
candle: keep catalog additions candle-only
candle: publish loaded models in health
candle: repair layer norm cuda lifetimes
candle: restore sidecar model readiness
candle: run bge-m3 profile in bf16
candle: vendor cublaslt for cuda build
candle: vendor layer norm cuda dependency
clean stale Python queue ownership drift
complete empty preformed worker requests
compress dev ami user data
config: defer cancellation after config commit starts
config: handle idempotency in-flight races
config: shield idempotency failure cleanup
copy rust worker vendor deps before docker cook
dashboard: harden run_status backfill per review feedback
dashboard: keep cancelled nightlies off the Quality on main widget
delete Python NATS pull loop from worker hot path
dev: address dev AMI review hardening
dev: harden SIE dev AMI workflow
dev: persist Codex state on dev cache disk
docs: drop otel.name from span-attribute table; tighten tracing dashboard test
export profile bundle targets
gateway: append forwarded embeddings headers to preserve multi-value tracestate
gateway: cancel abandoned direct batch fallback work
gateway: don’t report a succeeded SSE generation as inter_chunk_timeout on consumer lag
gateway: drop stale NAKs from abandoned attempts in handle_nak
gateway: forward traceparent/tracestate through /v1/embeddings rewrite
gateway: harden batch direct fallback cancellation
gateway: ignore pre-latch stale NAKs after republish
gateway: ignore quick-xml <0.41 RustSec advisories in cargo-deny
gateway: keep signalling SSE incompleteness on consumer lag; skip redundant cancel
gateway: keep signalling SSE incompleteness on lag; only skip cancel when done
gateway: record demand on backpressure so KEDA scales the lane
gateway: route engine-pin errors through endpoint_error_response
gateway: separate queue timeouts from model loading
gateway: use real RUSTSEC id for second quick-xml advisory
harden candle bge-m3 runtime bring-up
harden dev EC2 launcher dry-run
harden dev profile home fallback
harden SIE API key tooling
helm: make bundled collector downstream OTLP TLS configurable
helm: set OTEL_EXPORTER_OTLP_PROTOCOL=grpc on traced pods
helm: suppress bundled collector when explicit tracing endpoint set
improve candle embedding runtime throughput
install gateway pool config hashes
install gpg agent for dev AMI bake
keep dev AMI user data under limit
log candle batch failure context
models: load stock tokenizer for gte-Qwen2-7B-instruct
models: restore per-task MTEB instruction for stella_en_v5
models: serve embeddinggemma-300m via SentenceTransformerDenseAdapter
models: serve gte-Qwen2-1.5B via faithful sentence_transformer adapter
models: serve gte-Qwen2-1.5B via sentence_transformer adapter
muvera: center tokens before SimHash to fix answerai @muvera Blackwell collapse
muvera: drop same-dim count-sketch + retune colbert muvera configs
muvera: drop same-dim count-sketch, retune colbert configs + floors
persist dev cache volume
persist dev kubeconfig
playground: keyboard-accessible tabs + associated form labels
polish rollback backup examples
preserve candle compute cap diagnostics
preserve control-plane bundle hash in sidecar
quiet dev AMI tool installs
remove candle replica fallback path
remove Python worker pull-loop path
repair candle ci routing checks
repair candle rebase fallout
repeat dev AMI bake markers
require explicit API key rollback path
resolve latest dev AMI base
respect hf cache env in rust worker
restore default candle cuda stream
route kind rust worker build through docker task
sdk: raise actionable error on encode result-count desync
select rustls provider at binary startup
select rustls provider for rust services
server: /v1/embeddings maps OOM to 503 RESOURCE_EXHAUSTED, not 500
server: address CI + review comments on the Apple Silicon re-land
server: apply profile runtime options + append EOS on cluster encode path
server: apply profile runtime options and append EOS on cluster encode path
server: bound continuous-batch drain so a busy LoRA can’t starve siblings
server: bound the continuous-batching drain so a busy LoRA can’t starve siblings
server: cap LoRA drain to snapshot backlog so post-snapshot arrivals don’t overshoot
server: make SGLang load eviction headroom-aware
server: move load headroom behind adapter contract
server: serialize hot reload registry mutations
server: skip EOS on empty input; cover non-default profile on queue path
server: stop mislabelling uint8 embeddings as “binary” on the queue path
server: stop mislabelling uint8 embeddings as binary on the queue path
server: unregister from MemoryManager even if adapter.unload() raises
server: unregister model from MemoryManager even if adapter.unload() raises
set home during dev AMI bake
sidecar: bound scheduler bulk enqueue waves
sidecar: keep only config hash wiring
sidecar: start batch cancel before direct pulls
sie_server_rust: only log OTLP init success when exporter built
stabilize rust worker kind smoke
terraform: address CodeRabbit review on obs node group
terraform: auto-derive observability node group AMI from instance arch
terraform: make observability node group ami_type/disk_type configurable
tighten rust worker smoke diagnostics
tilt: use single string for candle-helm-smoke cmd (Starlark has no implicit concat)
tilt: wait for complete trace before asserting span chain
tolerate missing zstd devel package
tracing: assert items/metadata alignment before zip in scheduler batch
tracing: attach inbound trace context for non-generation publishes
tracing: parent sidecar.dispatch on first valid inbound context
verify dev AMI bake via ssm

Performance Improvements

candle: add GEMM diagnostics switch
candle: fuse xlm-r qkv projection
candle: optimize bge-m3 xlm-r runtime path
enable candle reduced precision gemms
fuse candle xlm-roberta qkv projection
gateway: build the HRW ring from the by_bundle index, not a whole-cluster scan
improve candle worker concurrency and build cache
restore candle xlm-roberta qkv layout
server: vectorize bge_m3_flash CLS gather to drop N per-encode GPU syncs
sidecar: bulk enqueue scheduler batches

Reverts

restore gateway sidecar rustls defaults

v0.6.14 (2026-06-26)

Highlights

Reliability and operations: unblock ColBERT score-on-retrieval — score_pairs 500 + muvera dense-encode + gate co-serve; include sie-bench in image context

Bug Fixes

#1430: unblock ColBERT score-on-retrieval — score_pairs 500 + muvera dense-encode + gate co-serve
docker: include sie-bench in image context

v0.6.13 (2026-06-25)

Highlights

New capabilities: publish sie-bench image to GHCR on release; per-tool GPU routing and env-overridable GLiNER models
Reliability and operations: extend tester structured generation timeouts; retry failed payload cleanup deletes; align sie-test qwen27 lane routing; build sie-bench image CPU-only with headless opencv; tolerate cold-start provisioning errors in scale-from-zero loadtests

Features

ci: publish sie-bench image to GHCR on release
sie-mcp: per-tool GPU routing and env-overridable GLiNER models

Bug Fixes

align sie-test qwen27 lane routing
bench: address CodeRabbit review (non-root image, clearer comment, checkout creds)
bench: build sie-bench image CPU-only with headless opencv
bench: tolerate cold-start provisioning errors in scale-from-zero loadtests
extend tester structured generation timeouts
gateway: dedupe offloaded payload cleanup keys
gateway: retry failed payload cleanup deletes
gateway: track exact offloaded payload keys
loadtest: tolerate cold-start provisioning errors + publish sie-bench to GHCR
persist sie-test qwen warm lanes
scale sie-test spot lanes to zero
server: drain removed profile variants during config apply
server: preflight profile variant config updates
server: preserve profile variants during config reconcile
server: serialize async config registry updates
server: tighten profile variant reconciliation
sie-mcp: chunk GLiNER extract/redact for large docs; flag the other 4B

v0.6.12 (2026-06-24)

Highlights

New capabilities: enable gateway-API model pinning on sie-test; add per-profile quality floors for bge-m3 sparse/multivector
Reliability and operations: exclude per-profile target.<profile>.json floors from result loaders; run docker task in project env; skip payload-store cleanup DELETEs for inline requests; guard label-window overflow + defer LEDGAR for v1.0 uni-encoders; gate subset-parameterized tasks; keep skipping candidate-gen dirs
Performance: release MaxSim per-batch intermediates to cap peak memory

Features

deploy: enable gateway-API model pinning on sie-test
quality-eval: add per-profile quality floors for bge-m3 sparse/multivector

Bug Fixes

bench: exclude per-profile target.<profile>.json floors from result loaders
ci: run docker task in project env
deploy: address CodeRabbit review on sie-test pinning
gateway: record offload before the store write (PR review)
gateway: skip payload-store cleanup DELETEs for inline requests
gliclass: guard label-window overflow + defer LEDGAR for v1.0 uni-encoders
quality-eval: gate subset-parameterized tasks; keep skipping candidate-gen dirs
sie_bench: batch-local query padding in MaxSim to fix #1461 OOM

Performance Improvements

sie_bench: release MaxSim per-batch intermediates to cap peak memory

v0.6.11 (2026-06-23)

Highlights

New capabilities: floor Florence-2-base on COCO detection (AP 0.2785); floor Florence-2-base on olmOCR-bench (accuracy 0.0824); fail fast when a bundle is placed on the wrong platform; add Gemma 4 generation models (E2B, E4B, 26B-A4B); first-class cuda13 platform for the gemma bundle; first-class cuda13 platform for the gemma bundle (build/CI/release/deploy)
Reliability and operations: raise uv HTTP timeout for the gemma cu130 wheel downloads; harden logical pool queue routing; harden worker-direct batch fallback; return retryable model loading for generation; gate mteb task instruction to instruction-following models

Features

bench: floor Florence-2-base on COCO detection (AP 0.2785)
bench: floor Florence-2-base on olmOCR-bench (accuracy 0.0824)
deploy: fail fast when a bundle is placed on the wrong platform
server: add Gemma 4 generation models (E2B, E4B, 26B-A4B)
server: first-class cuda13 platform for the gemma bundle
server: first-class cuda13 platform for the gemma bundle (build/CI/release/deploy)
server: keep pinned models loaded and exclude them from LRU eviction
support logical pools backed by queue pools
terraform: add marci-dev Azure A10 single-node smoke-test example
terraform: Azure marci-dev A10 smoke-test example + node-subnet NSG LB fix

Bug Fixes

align retryable generation review feedback
bench: gate mteb task instruction to instruction-following models
bench: gate MTEB task instructions to instruction-following models
bench: resolve profile output_types from adapter_options.runtime for bge-m3 sparse/multivector evals
ci: raise uv HTTP timeout for the gemma cu130 wheel downloads
ci: run the gemma cuda13 build smoke + cover it in release verify/stamp/warm
disable payload store in kind smoke
handle sidecar progress ack failures
harden logical pool queue routing
harden worker-direct batch fallback
loadtest: disable payload store in nightly deploy
merge runtime pool capacity changes
preserve queue delivery budget
return retryable model loading for generation
server: correct stale cuda13/gemma platform references
server: make docker bake/verify platform-aware (cuda13/gemma in mixed builds)
server: make transformers version bundle-controlled
terraform: allow public LoadBalancer/ingress inbound on AKS node subnet NSG
terraform: clear internal mise references from public Azure module
terraform: clear internal mise/tooling references from public AWS and GCP modules
terraform: drop internal references from public Azure module README and variables

v0.6.10 (2026-06-22)

Highlights

New capabilities: floor measure-first dark models, retire donut-rvlcdip; provision and enable the payload store by default; fix <OD> bbox conversion, re-task base-ft/large to COCO detection
Reliability and operations: align gliner label inputs; align gliner v2.5 conll protocol; improve generation failure observability; coalesce whitespace-only instruction to default; pool post-RMSNorm last_hidden_state

Features

bench: floor measure-first dark models, retire donut-rvlcdip
deploy: provision and enable the payload store by default
florence2: fix <OD> bbox conversion, re-task base-ft/large to COCO detection

Bug Fixes

bench: align gliner label inputs
bench: align gliner v2.5 conll protocol
deploy: address payload-store review feedback
improve generation failure observability
qwen3_vl_embedding: coalesce whitespace-only instruction to default
qwen3_vl_embedding: pool post-RMSNorm last_hidden_state
sie_bench: coerce ClassLabel int names to str in classification loader
sie_bench: repoint extract-kie FUNSD to live nielsr/funsd dataset
sie-cluster: make the self-host deploy skill robust to its own runtime
sie-server: make transformers5 bundle reachable via pip
sync generation visibility api contract

v0.6.9 (2026-06-19)

Highlights

New capabilities: add per-pool pinned-model set to the pool config API; per-pool pinned-model set in the pool config API; support profile-qualified ids in per-pool pinned-model set; enable S3 payload store for >1MB work items
Reliability and operations: bound pinned-model metric labels and add handler validation tests; scope worker preload models by lane; alias PEFT LoRA adapter names; expose model_cache_bucket_url as a root output

Features

gateway: add per-pool pinned-model set to the pool config API
gateway: per-pool pinned-model set in the pool config API
gateway: support profile-qualified ids in per-pool pinned-model set
tester-cluster: enable S3 payload store for >1MB work items

Bug Fixes

gateway: bound pinned-model metric labels and add handler validation tests
scope worker preload models by lane
server: alias PEFT LoRA adapter names
tester-cluster: expose model_cache_bucket_url as a root output

v0.6.8 (2026-06-16)

Highlights

Reliability and operations: harden release artifact workflows; keep score traffic from idle-evicting rerankers; serialize worker use with unload state; share score media batch cost

Bug Fixes

ci: harden release artifact workflows
server: keep score traffic from idle-evicting rerankers
server: serialize worker use with unload state
server: share score media batch cost

v0.6.7 (2026-06-16)

Highlights

New capabilities: answer_questions transient-QA tool (Req 12 #1309)
Reliability and operations: guard model ready timeout against liveness budget; wire model ready timeout through helm; add Qwen3.6 27B 32K RTX serving profile; route grammar requests to a non-speculative profile (NEXTN bypasses Outlines FSM); make worker config reconciliation no-op safe
Performance: extract shared vision patch-embed rebind helper; batch generate() across pages to close #601 L4 throughput gap

Features

mcp: answer_questions transient-QA tool (Req 12 #1309)

Bug Fixes

add Qwen3.6 27B 32K RTX serving profile
address qwen 27b review feedback
generate: route grammar requests to a non-speculative profile (NEXTN bypasses Outlines FSM)
guard model ready timeout against liveness budget
make worker config reconciliation no-op safe
route qwen 27b fp8 cold starts
wire model ready timeout through helm

Performance Improvements

adapters: extract shared vision patch-embed rebind helper
lighton_ocr: batch generate() across pages to close #601 L4 throughput gap

v0.6.6 (2026-06-14)

Highlights

Reliability and operations: align pool-scoped bundle hashes; avoid sticky missing bundle hashes; clarify missing profile inheritance; fail closed on missing bundle metadata; stabilize keda all-marker e2e

Bug Fixes

config: align pool-scoped bundle hashes
config: avoid sticky missing bundle hashes
config: clarify missing profile inheritance
config: fail closed on missing bundle metadata
tilt: stabilize keda all-marker e2e

v0.6.5 (2026-06-13)

Highlights

New capabilities: demand-side token-reduction benchmark (Req 12, #1311); add describe_image tool (caption + zero-shot tags); add describe_image tool (caption + zero-shot tags) — Req 12 #1310; cap describe_image payload size before cluster calls; claude.ai connector surface — OAuth bridge + skill ZIP (Req 12 #1312); sie_mcp edge with docs_to_markdown tool (Req 12 #1306)
Reliability and operations: harden replace snapshot IPC; return retryable OpenAI provisioning errors; bind OAuth authorization codes to client_id; doctor classifies probe read-timeouts as cold, not unreachable; align GPU memory pressure defaults
Performance: skip tag embedding when top_k <= 0

Features

bench: demand-side token-reduction benchmark (Req 12, #1311)
mcp: add describe_image tool (caption + zero-shot tags)
mcp: add describe_image tool (caption + zero-shot tags) — Req 12 #1310
mcp: cap describe_image payload size before cluster calls
mcp: claude.ai connector surface — OAuth bridge + skill ZIP (Req 12 #1312)
mcp: sie_mcp edge with docs_to_markdown tool (Req 12 #1306)
mcp: structured extraction + structured generation tools (Req 12 #1308)
mcp: wire measured token-reduction figures into savings metadata
tools: add sie doctor — per-capability cluster diagnostics
tools: Florence-2 fallback for image OCR
tools: sie_tools — Claude Code context-offload client for managed clusters

Bug Fixes

align GPU memory pressure defaults
config: detect bundle config hash drift
config: fingerprint model pool ownership
config: harden replace snapshot IPC
config: replace drifted export snapshots
deps: bump sidecar prometheus for protobuf advisory
gateway: address provisioning review feedback
gateway: align provisioning contract docs
gateway: decode native media JSON bytes
gateway: dereference structured output schema refs
gateway: make provisioning non-2xx universally
gateway: preserve ref sibling schema semantics
gateway: return retryable OpenAI provisioning errors
mcp: address review feedback on structured tools
mcp: bind OAuth authorization codes to client_id
mcp: blank-env fallback for model ids; honor SIE_MCP_IMAGE_TOP_K=0
mcp: deep-copy committed token-reduction figures in build_metadata
mcp: validate embedding shapes in _top_k_tags
sdk: normalize score image payloads for wire transport
sdk: normalize score images for wire transport
server: guard readiness for removed configs
server: honor pool-aware model configs
server: render qwen3 vl reranker document images in user prompt
server: render Qwen3-VL reranker document images in user prompt
sie-cluster: add spot toleration to AKS worker pool
tools: address doctor review feedback
tools: doctor classifies probe read-timeouts as cold, not unreachable
worker: keep SGLang loads off event loop

Performance Improvements

mcp: skip tag embedding when top_k <= 0

v0.6.4 (2026-06-11)

Highlights

New capabilities: add grant_admin_to_creator (opt-in AAD-RBAC for caller); lock model-cache storage account to cluster VNet by default; install kubelogin and convert kubeconfig after AKS get-credentials; wire Azure provider tooling; add azure (AKS) terraform module; ship values-aks.yaml AKS overlay with the Azure module
Reliability and operations: harden storage_allowed_ip_ranges CIDR validation; harden release guarded merge checks; resubscribe stale NATS health stream; emit az aks get-credentials --overwrite-existing; drop unreachable final_registry guard so ACR path can fire

Features

azure-terraform: add grant_admin_to_creator (opt-in AAD-RBAC for caller)
azure-terraform: lock model-cache storage account to cluster VNet by default
cluster: install kubelogin and convert kubeconfig after AKS get-credentials
cluster: wire Azure provider tooling
deploy: add azure (AKS) terraform module
helm: ship values-aks.yaml AKS overlay with the Azure module

Bug Fixes

azure-terraform: emit az aks get-credentials --overwrite-existing
azure-terraform: harden storage_allowed_ip_ranges CIDR validation
ci: harden release guarded merge checks
cluster: address review feedback on Azure provider wiring
cluster: drop unreachable final_registry guard so ACR path can fire
cluster: set TF_VAR_* on Azure destroy path (same as create)
deploy: revert system pool default to Standard_D4s_v3 (zoned everywhere)
gateway: resubscribe stale NATS health stream
sidecar: preserve msgpack work item payloads

v0.6.3 (2026-06-10)

Highlights

New capabilities: add azure blob payload store support; add server-side copy fast path for cloud weight sync; informational generation eval CI gate over committed floors; add vision (image) input to generate(); preserve text/image content-part ordering; vision (image) input for generate()
Reliability and operations: harden cloud cache sync paths; clear HIGH Dependabot alerts (docling, rustls-webpki); ensure cloud weight sync creates local parents; evict stale gateway workers on shutdown; fall back to relay on S3/GCS server-side copy failure
Performance: engage conformant image preprocessing for v1; engage conformant image preprocessing for v1 (1.8x)

Features

add azure blob payload store support
add server-side copy fast path for cloud weight sync
bench: informational generation eval CI gate over committed floors
generate: add vision (image) input to generate()
generate: preserve text/image content-part ordering
generate: vision (image) input for generate()
support azure blob cluster cache
tester-cluster: rtx6000 g7e.4xlarge + sglang preload + hf-token wiring

Bug Fixes

address azure cache review feedback
address final cloud storage review issues
bench: harden generation eval gate per review
deps: clear HIGH Dependabot alerts (docling, rustls-webpki)
ensure cloud weight sync creates local parents
evict stale gateway workers on shutdown
fall back to relay on S3/GCS server-side copy failure
generate: address CodeRabbit review on vision input
generate: address huronat review on vision input (F2-F8)
generate: image-free content_parts field must not shadow layout
generate: reject both-present image-bearing content layouts
harden cloud cache sync paths
loadtest-ci: self-heal orphaned cluster + stale lock in preflight
nemo_colembed: trim left-padding rows from v1 conformant doc embeddings
normalize local weight sync destination
quality-adapter: gate v1 Vidore3 on English; finalize ?lang= plumbing
skill: add bash language tag to hfCache —set fenced block (MD040)
skill: move inline comments off shell continuation lines so the helm snippet pastes cleanly
support cloud source weight sync
tester-cluster: update rtx6000-spot machineType doc to g7e.4xlarge to match terraform

Performance Improvements

nemo_colembed: engage conformant image preprocessing for v1
nemo_colembed: engage conformant image preprocessing for v1 (1.8x)

v0.6.2 (2026-06-08)

Highlights

New capabilities: defer sie-config NATS startup and honor log levels; refresh KEDA Tilt local dev branch; M4 dense encoders — mxbai-embed-large-v1, arctic-embed-l-v2.0, modernbert-embed-base; add daily guarded stable releases
Reliability and operations: accept dense dim in qwen3 vl embedding adapter; preserve model query templates in mteb eval; scale single-profile bundles on gpu-agnostic demand; consolidate runtime ninja install; install ninja in cuda runtime

Features

defer sie-config NATS startup and honor log levels
dev: refresh KEDA Tilt local dev branch
models: M4 dense encoders — mxbai-embed-large-v1, arctic-embed-l-v2.0, modernbert-embed-base
release: add daily guarded stable releases

Bug Fixes

accept dense dim in qwen3 vl embedding adapter
bench: preserve model query templates in mteb eval
dev: address KEDA Tilt PR review
helm: scale single-profile bundles on gpu-agnostic demand
server: consolidate runtime ninja install
server: install ninja in cuda runtime
server: install ninja in CUDA SGLang runtime
terraform: deny non-HTTPS access on state and quality-eval S3 buckets

v0.6.1 (2026-06-07)

Highlights

New capabilities: configure GPU disk sizing and generate smoke; support static queue pools
Reliability and operations: fail fast on invalid static pool config; pin kind smoke workers to default queue pool; canonicalize static queue pool names; stabilize GPU disk Terraform test

Features

configure GPU disk sizing and generate smoke
gateway: support static queue pools

Bug Fixes

address GPU disk review comments
ci: pin kind smoke workers to default queue pool
gateway: canonicalize static queue pool names
gateway: fail fast on invalid static pool config
stabilize GPU disk Terraform test

v0.6.0 (2026-06-07)

Highlights

Breaking change: Queue work subjects and pool streams use the new sie.work.{pool}.{machine_profile}.{bundle}.{model} shape only; legacy subject filters are intentionally not preserved.; workers will subscribe to sie.work.*.<poolName> instead of sie.work.*.default. Deployed alone (without the matching gateway/sidecar update that publishes/filters on the new subject) this will break routing on every cluster. To preserve the old shared-queue behavior, set workers.common.queuePool: "default" explicitly.
New capabilities: route work by queue pool lanes; default SIE_POOL to pool name (not “default”)
Reliability and operations: harden queue lane admission; bump vitest 2.1.9 -> 4.1.0 (CVE-2026-47429); align lane defaults and tilt e2e; preserve worker-group queue defaults

⚠ BREAKING CHANGES

gateway: Queue work subjects and pool streams use the new sie.work.{pool}.{machine_profile}.{bundle}.{model} shape only; legacy subject filters are intentionally not preserved.
helm: workers will subscribe to sie.work.*.<poolName> instead of sie.work.*.default. Deployed alone (without the matching gateway/sidecar update that publishes/filters on the new subject) this will break routing on every cluster. To preserve the old shared-queue behavior, set workers.common.queuePool: "default" explicitly.

Features

gateway: route work by queue pool lanes
helm: default SIE_POOL to pool name (not “default”)

Bug Fixes

deps: bump vitest 2.1.9 -> 4.1.0 (CVE-2026-47429)
gateway: harden queue lane admission
helm: align lane defaults and tilt e2e
helm: preserve worker-group queue defaults

v0.5.0 (2026-06-04)

Highlights

Breaking change: workers.pools.<name>.bundle (string), workers.pools.<name>.minReplicas, workers.pools.<name>.maxReplicas, workers.pools.<name>.extraEnv, and workers.pools.<name>.imageBundle are replaced by workers.pools.<name>.bundles.<bundle>.{minReplicas, maxReplicas, extraEnv, imageBundle, enabled}. workers.common.bundle is removed (no longer consumed). StatefulSet, ScaledObject, PDB, and image-prepull DaemonSet names change from worker-{pool} to worker-{pool}-{bundle}, so in-place upgrades require deleting the old resources first.
New capabilities: agent-jobs text-gen readiness — code/SQL/tools/guard evals + Qwen3.6-27B + precision routing; transfer sie-cluster claude skill; P(unsafe) logprob threshold for CHECK POLICY precision; split worker pools into pool × bundles schema; surface code/sql/guard capabilities; resolve job aliases in configs/resolve; add sglang worker pool for generative models
Reliability and operations: expose unauthenticated metrics scrape port; expose unauthenticated metrics scrape port safely for prom; preserve gateway metrics scrape labels; drop unsupported ebnf advertisement + restore guardian a100 guard threshold; fail-fast on missing Spider DBs + order-sensitive SQL exec accuracy

⚠ BREAKING CHANGES

helm: workers.pools.<name>.bundle (string), workers.pools.<name>.minReplicas, workers.pools.<name>.maxReplicas, workers.pools.<name>.extraEnv, and workers.pools.<name>.imageBundle are replaced by workers.pools.<name>.bundles.<bundle>.{minReplicas, maxReplicas, extraEnv, imageBundle, enabled}. workers.common.bundle is removed (no longer consumed). StatefulSet, ScaledObject, PDB, and image-prepull DaemonSet names change from worker-{pool} to worker-{pool}-{bundle}, so in-place upgrades require deleting the old resources first.

Features

agent-jobs text-gen readiness — code/SQL/tools/guard evals + Qwen3.6-27B + precision routing
agents: transfer sie-cluster claude skill
guard: P(unsafe) logprob threshold for CHECK POLICY precision
helm: split worker pools into pool × bundles schema
models: surface code/sql/guard capabilities; resolve job aliases in configs/resolve
tester-cluster: add sglang worker pool for generative models

Bug Fixes

agents: address sie cluster review comments
bench: fail-fast on missing Spider DBs + order-sensitive SQL exec accuracy
gateway: expose unauthenticated metrics scrape port
gateway: expose unauthenticated metrics scrape port safely for prom
guard: reject multi-candidate sampling + keep logprobs consistent on rewrite
guard: robust verdict thresholding, logprob hygiene, decoded-token logprobs
helm: fail-fast on missing/invalid bundle replica bounds
helm: preserve gateway metrics scrape labels
helm: use sidecar binary for image pre-pull
models: drop unsupported ebnf advertisement + restore guardian a100 guard threshold
sie_server: honor params.instruction in Florence-2 extract
tester-cluster: cap rtx6000 default bundle to avoid over-subscription
tools: via-SIE EBNF response_format shape + request/preload model split

v0.4.2 (2026-06-03)

Highlights

New capabilities: 5-domain generation bench + via-sie quality matrix + gateway schema gaps; add e0-02 all-minilm time-share experiment; land coalesce_ms=5 + max_batch_requests=12 as Rust defaults; add —via-sie smoke path (route through sie_server); add min_tokens + system_prompt + temperature for G4 retry; close Qwen3.6-27B gap — min_tokens=10 + max=768 + ctx=4096
Reliability and operations: set verbose=True on SIEServer so launch errors surface; document worker-sidecar metrics wiring; gate sidecar nats reconnect refresh; harden sidecar config recovery; budget loadtest barrier timeouts
Performance: anchor min_batch_cost floor at max_batch_tokens // 4; tighten adaptive wait ceiling + revert gte-multilingual 32k; rebind vision Conv3d patch-embed to F.linear; raise max_batch_tokens 16k → 32k to stop IPC-batch shred

Features

5-domain generation bench + via-sie quality matrix + gateway schema gaps
add e0-02 all-minilm time-share experiment
batch_config: land coalesce_ms=5 + max_batch_requests=12 as Rust defaults
bench-27b: add —via-sie smoke path (route through sie_server)
bench-27b: add min_tokens + system_prompt + temperature for G4 retry
bench-27b: close Qwen3.6-27B gap — min_tokens=10 + max=768 + ctx=4096
bench-27b: launch full SIE stack (NATS+worker+gateway) for —via-sie
bench+model: via-sie 4-task n=300 sweep + NEXTN smaller-draft on 27B
bench: 0.6B via-sie validated; harness + 27B config gains
bench: 5-shot CoT for CaseHOLD (item 5 — close 27B target gap)
bench: fix Qwen3-0.6B GPQA (parrot bug) + 27B diagnostics; final matrix
bench: improve perf eval output handling
docling: accept image input + run on OCR-bench quality path
gateway+worker: chat surface accepts min_tokens + chat_template_kwargs
gateway: strengthen generation isolation guardrails
latency: tighten FetchExpiryController defaults to 2/15/50
model+bench: RTX-PRO-6000 FP8 profile for Qwen3.6-27B + 6000 validation
model: bump Qwen3-0.6B serving context 1024→4096 for prod simple-task use
models: add Marqo/marqo-fashionSigLIP (SigLIP open_clip, fashion image-text)
ocr: docling accepts images + quality eval prefers documents
reconcile live worker config in sidecar
RTX PRO 6000 FP8 profile for Qwen3.6-27B + SIE-on-6000 generative benchmark matrix
scheduler: load-aware pipeline_depth autotune (S14 follow-up)
scheduler: production-parity defaults + serial pipeline (carveout p99 fix)
scheduler: restore SIE_RUST_PIPELINE_DEPTH=2 default (deep-saturation fix)
scheduler: SIE_PULL_QUANTUM_INCLUDE_QUEUE_MS for Py-main parity
scheduler: SIE_RUST_WAVE_CADENCE env toggle (default on)
scheduler: step adaptive controller once per wave (Python parity)
sidecar: add worker config and pool admission reconciliation
sidecar: wire generation direct dispatch
sie_server: add MinerU2.5-Pro-2604-1.2B doc OCR adapter
sie_server: carve out QueueExecutor + IPC types for Rust worker POC
sie_server: integrate MinerU2.5-Pro-2604-1.2B doc OCR adapter
sie_server: UDS msgpack IPC server for Rust worker sidecar
sie_worker_rust: close parity gaps with Python pull loop + smoke test
sie_worker_rust: scaffold Rust worker sidecar crate (Phase 1c)
sie_worker_rust: wire end-to-end NATS -> IPC -> publish loop (Phase 1d)
sie-bench: synchronize loadtest measurement start
worker/rust: IPC connection pool — lift the sidecar’s last serialization bottleneck
worker/rust: narrate the hot path — structured INFO, slow-RPC + heartbeat-streak WARNs, full error chains
worker: introduce InferenceBackend trait + BackendRouter
worker: native Candle BERT backend behind candle feature

Bug Fixes

accept dense_dim in dense adapters
adapters: replace Qwen3-VL vision Conv3d patch-embed with matmul
adapters: route Qwen3-VL VLMs through flash attention (Vidore3 throughput)
address pr review quality issues
bench-27b: drop bundle from SIEServer (sie-server rejects bundle+models combo)
bench-27b: set verbose=True on SIEServer so launch errors surface
bench-27b: skip chat_template_kwargs on via-sie (gateway rejects unsupported field)
bench-27b: wait for sie-server /healthz (not /health)
bench: bump casehold/gpqa max_tokens to 2048 (CoT truncation)
bench: let via-SIE smoke serve a profile-variant model end-to-end
bench: resolve CPU deps for quality server
catalog: include eval-matrix tasks so dispatch filter accepts them
ci: address analyzer findings and stale queue test
ci: avoid nested mise in integration fixture
ci: keep sidecar out of warm cache
ci: refresh gateway openapi contract
correct e0 vm runbook paths
deploy: add sidecar registry resources
deploy: address server sidecar review feedback
deploy: align server sidecar naming
deploy: align server sidecar naming and kind preload smoke
deploy: align tilt sidecar image naming
deploy: document worker-sidecar metrics wiring
deploy: keep sidecar on GHCR by default
deploy: normalize server sidecar naming
deploy: publish server sidecar image
deploy: rename sidecar container to worker-sidecar
deploy: wire SIE server sidecar for kind smoke
deploy: wire worker sidecar image across kind and cloud
gate sidecar nats reconnect refresh
gateway+server: queue is the only mode — kill direct-mode cruft
gateway: suppress H9 first-chunk-fallback on single-worker pools
harden sidecar config recovery
impact-map: keep profiles distinct when adapter_options differ
keep generation machinery off default queue path
loader: wire profile runtime.default_sampling into the adapter
modal: report actual GPU on remote, not stale env-default
model: bump Qwen3.6-27B default/h100 mem_fraction_static 0.85 → 0.92
orchestrator: thread CLI -p profile through to client.extract
preserve worker batch identity and publish image
product: update design audit for topical docs
quality_eval: take results-bearing JSON envelope in load_eval_json
quality: batch3 of CodeQL findings + bench KIE bug
quality: batch3 of CodeQL findings + bench KIE root-cause
quality: batches 1+2 of CodeQL quality findings
quality: close CodeQL quality-tab findings
quality: drop redundant inline imports in donut + registry
quality: repair adapter eval harness regressions
remove e0 preflight httpx dependency
require rust sidecar for queue workers
review: 0.6B ctx test 1024->4096, loader except logs, README gaps resolved
review: recompute 27B target delta_vs_baseline for the 2048 scores
run directory creation
run e0 vm scripts via uv
scheduler: autotune signal — observed_p50/target_p50 ratio
scope bundle config hash cache per registry
security: bump astro to ^6.4.2 for website
security: bump gateway deps to patched versions
security: bump product/gtm Python lockfiles
security: bump product/gtm/content/slides npm transitives
security: bump root pnpm deps + add overrides for transitives
security: bump root Python deps to patched versions
security: bump sie_dashboard npm deps to patched versions
security: bump sie_ts_sdk standalone pnpm transitives
security: bump sst to ^4 to drop vulnerable aws-sdk v2
security: cap vite at ^6 + add Node engines to website
security: close ~190 Dependabot alerts across 9 manifests
security: sanitize one-pager template with DOMPurify
security: use Reflect.construct for WebSocket headers shim
sie_bench: send SIE profile via X-SIE-MACHINE-PROFILE header
sie_bench: use rapidfuzz for OmniDocBench edit distance
sie_server: clear CUDA cache on uncovered VLM paths + drop private sem _value access
sie_server: VLM cache clears on uncovered paths + drop private sem _value access
sie-bench: budget loadtest barrier timeouts
slow sidecar nats consumer reconcile
smoke: launch sie_server worker with -b sglang, not -m <model>
smoke: preload the target model in via-sie worker
test: restore donut helper call contract
worker-sidecar: harden queue carveout contracts
worker/rust: one long-lived pull stream — kill 30s ack_wait stall
worker/rust: re-copy src after cargo chef cook so real build isn’t a stub
worker/rust: set CUDA_COMPUTE_CAP at build time (default 89, L4)
worker/rust: stop shipping the cargo-chef stub binary as the real build
worker: harden Candle backend + align dispatcher error contract
worker: harden payload store + error paths; surface silent success bugs
worker: SGLang adapter accepts min_new_tokens kwarg + 27B via-sie validated

Performance Improvements

adaptive: anchor min_batch_cost floor at max_batch_tokens // 4
batching: tighten adaptive wait ceiling + revert gte-multilingual 32k
glm_ocr: rebind vision Conv3d patch-embed to F.linear
gte-multilingual-base: raise max_batch_tokens 16k → 32k to stop IPC-batch shred
mineru_vl: O(L) incremental no-repeat-ngram for greedy decode
ocr: swap pure-Python Levenshtein DP for rapidfuzz
rope_flash: vectorize CLS/mean pooling, eliminate per-item .item() sync
server: FP16 on GPU, coalesce sized for IPC bursts, starvation self-heal

Reverts

restore adaptive batching defaults to 15/50ms
scheduler: drop depth autotune (signal didn’t pan out in S17)

v0.4.1 (2026-05-28)

Highlights

New capabilities: add Qwen3.6-27B model + migrate to CUDA 12.9
Reliability and operations: isolate generation direct dispatch from shared queues; resolve 18 open CodeQL alerts; use SHA256 (not SHA1) for actor_id log tag; colocate tests under infra/, update sync contract

Features

server: add Qwen3.6-27B model + migrate to CUDA 12.9

Bug Fixes

isolate generation direct dispatch from shared queues
security: resolve 18 open CodeQL alerts
security: use SHA256 (not SHA1) for actor_id log tag
terraform-sync: colocate tests under infra/, update sync contract

Reverts

security: drop advanced CodeQL setup

v0.4.0 (2026-05-27)

Highlights

Breaking change: fail-closed authentication (default-deny)
New capabilities: generation quality-gate scoring core (roadmap §5, trust-critical); generation-quality regression gate over the existing scorers; add cohere measurements for us-east-1; add openai measurements for us-east-1; add voyage measurements for us-east-1; regex/EBNF response_format + developer role (roadmap 1.7)
Reliability and operations: forward provision_timeout_s in SIEImageTextWrapper.encode; raise image-task eval timeouts to fix Flickr30k nightly; exclude favicon + OG image from auth middleware; keep public surfaces vague about what’s behind auth; revert NextAuth function-form, use try/catch on Resource
Performance: warm one Lambda, bump timeout, narrow S3 verdict fetch

⚠ BREAKING CHANGES

gateway: fail-closed authentication (default-deny)

Features

bench: generation quality-gate scoring core (roadmap §5, trust-critical)
bench: generation-quality regression gate over the existing scorers
benchmarks: add cohere measurements for us-east-1
benchmarks: add openai measurements for us-east-1
benchmarks: add voyage measurements for us-east-1
chat: regex/EBNF response_format + developer role (roadmap 1.7)
dashboard: add executive quality summary widget on landing
dashboard: add hover-tooltip on ‘to verify’ explaining WARN
dashboard: brand alignment foundation - palette, fonts, header
dashboard: brand favicon, opengraph image, light-mode hover fix
dashboard: brand foundation - palette, fonts, header logo
dashboard: brand surfaces - retokenise landing page + widget
dashboard: public preview shell on / so Slack unfurls work
dashboard: retokenise badges - brand-elevated neutrals, DM Mono labels
dashboard: retokenise landing surfaces onto brand palette
dashboard: retokenise loadtest pages onto brand surfaces
dashboard: retokenise quality pages onto brand surfaces
dashboard: retokenise shared components and sign-in onto brand palette
dashboard: treat WARN as passing-with-verify, shade gauge amber
gateway,sdk,server: add native generate endpoint with improved admission control and validation
gateway,sdk,server: add OpenAI-compatible chat completions with streaming and sampling extensions
gateway,server: add multi-turn tool-use support with OpenAI-compatible message format
gateway: /v1/completions (legacy OpenAI Completions, raw-prompt)
gateway: /v1/completions streaming (text_completion SSE)
gateway: /v1/generate accepts seed/logprobs/logit_bias/n/best_of/lora_adapter (M8)
gateway: /v1/responses (OpenAI Responses API, MVP)
gateway: /v1/responses structured array input (conversation history)
gateway+worker: per-choice OpenAI streaming for n>1 (H4, H5, M4)
gateway: accept OpenAI multimodal content-parts; reject images (no VL model)
gateway: add routing salt + byte-preserving key mode (M11)
gateway: advertise lora_adapters on /v1/models + pre-validate unknown names
gateway: fail-closed authentication (default-deny)
gateway: meaningful system_fingerprint on chat responses (roadmap 1.3/§5)
gateway: refactor streaming and routing with improved error handling and metrics
gateway: register /v1/moderations as explicit 501 (roadmap 1.8, phase 3)
gateway: serve a rendered API reference at /docs (Redoc)
gateway: unify /v1/embeddings on the OpenAI error envelope (roadmap 1.4)
generation: best_of — over-generate + rank by logprob, return top n
generation: complete M4 req2 generation primitive with streaming, structured outputs, and routing
generation: multi-candidate n>1 (non-streaming) end-to-end (roadmap 1.5)
generation: multi-LoRA serving (one base, N adapters, per-request) (roadmap 6.2)
generation: ship generate() primitive — Qwen3.5-4B + NEXTN/MTP + xgrammar, adapter perf at parity with raw SGLang
generation: streaming n>1 — per-candidate SSE interleave
helm/sie-cluster: bundle cert-manager + trust-manager (opt-in) with self-signed TLS mode
openapi: add tool_calls support to chat completion schema
python-sdk: expose typed params for chat n/logprobs/lora_adapter/etc (M7)
quality-eval: add heartbeat logging and improve long-running process observability
routing: cache-aware (prefix-hash) routing (roadmap §6.3)
sie_bench: add Cohere as a first-class eval source
sie_bench: add Cohere multimodal embeddings
sie_bench: add Cohere rerank backend for native MTEB rerank tasks
sie_bench: add OpenAI Embeddings as a first-class eval source
sie_bench: add Voyage provider source plumbing
sie_bench: implement Voyage text embedding runner
sie_dashboard: add /quality/compare to diff two quality runs
terraform: add default_tags Project=sie/Cluster on all AWS providers
terraform: add on-demand RTX 6000 baseline pool to tester-cluster
terraform: add uniform project=sie label across all GCP clusters
terraform: idle-stop and on-demand wake for quality-eval runner fleet
terraform: scale quality-eval fleet to 5+5, smart-wake, 4h timeout
terraform: wake-runners retries + watchdog queued-jobs backstop
tester-cluster: add on-demand L4 worker pool + capacityType node pins
ts-sdk: handle 202 provisioning in chatCompletions + expose missing fields (H1+M6)
worker: SGLang owns grammar; worker preflight opt-in only (H8, ADR-0002)
worker: wire mixed-pool fairness scheduler into the pull-loop (opt-in)
worker: WorkClassScheduler core for mixed-pool fairness (roadmap §6.1)

Bug Fixes

bench: declare olmocr[bench] dep for OCR-bench quality eval
bench: drop —with-deps from playwright install (no sudo on g7e)
bench: forward provision_timeout_s in SIEImageTextWrapper.encode
bench: install playwright chromium for OCR-bench KaTeX rendering
bench: playwright install —with-deps for OCR-bench chromium
bench: raise image-task eval timeouts to fix Flickr30k nightly
bench: respect similarity() inputs in ColBERT/ColPali wrappers
bench: run olmocr.bench.tests off orchestrator’s asyncio loop
centralize worker_id subject normalization (M5)
chart: pin image-prepull DaemonSets to GPU nodes
ci: copy assets/ into the gateway Docker build (redoc bundle)
dashboard: close remaining ‘SIE Dashboard’ leaks in <title> and og:image:alt
dashboard: drop edge runtime on opengraph-image for OpenNext
dashboard: exclude favicon + OG image from auth middleware
dashboard: keep public surfaces vague about what’s behind auth
dashboard: pick healthy daily via coverage + health gates
dashboard: require >=50 pairs on main-run fallback in nightly picker
dashboard: revert NextAuth function-form, use try/catch on Resource
dashboard: short tab title for authed users, neutral for unauth
dev: set explicit auth opt-in for local gateway launchers (post fail-closed)
gateway,sdk,server: add wire-level validation, improve resource cleanup, and enhance observability across request lifecycle
gateway,sdk,server: prevent metric cardinality DoS and fix non-idempotent retry logic
gateway,sdk,server: strengthen validation and eliminate silent failures across request lifecycle
gateway,sdk,server: validate numeric fields and improve error handling
gateway: add NATS config trusted producers helm override
gateway: document ModelCapabilities in OpenAPI + refresh on profile delta-update
gateway: generation timeouts bypass legacy request-timeout ceiling (H7)
gateway: scope LoRA adapter capabilities per profile (M10)
gateway: strict allow-list + 400 contract on /v1/completions (H3)
gateway: strict allow-list + 400 contract on /v1/responses (H2)
gateway: tighten chat sampler/token-cap + tool-history validation (M1, M13)
gateway: trust chart-rendered sie-config pod name for NATS deltas
generation: cancel tombstone prevents first-chunk fallback double-execution (H9)
generation: LoRA lora_path is a top-level /generate field, not a sampling param
generation: tighten lossy tool-control flags (M14)
grammar: resolve tokenizer adapter for Outlines processor factories and remove anchors from regex patterns
helm/sie-cluster: guard validateTls probe against deployments with nil labels
helm/sie-cluster: guard validateTls probe against nil deployment labels
helm/sie-cluster: include cert-manager mode in presence-check gate
helm/sie-cluster: label-based cert-manager detection + bidirectional runtime check
helm/sie-cluster: make self-signed root-CA namespace configurable
helm/sie-cluster: one-step bundled cert-manager install with self-signed TLS
helm/sie-cluster: probe cert-manager controllers cluster-wide
helm/sie-cluster: regenerate Chart.lock with synced digest
helm: trim and drop empty entries in ingress.hosts
modal: exclude cargo target/ from sandbox image mount
pytorch_embedding: accept and forward revision kwarg
quality_eval: tolerate stdout noise around eval JSON envelope
quality-eval: handle paginated jobs API and align last two filters
release-docker,warm-cache: address review findings
server,gateway: add GPU-aware health probes to detect and recover from wedged CUDA contexts
server,gateway: GPU-aware health probes to detect & recover from wedged CUDA contexts
sie_bench: register OPENAI_SOURCE so —save-targets openai actually saves
sie_dashboard: address compare-page review nits
sie_dashboard: label compare log links with run IDs
sie_server: base64-decode JSON image inputs
sie_server: enforce media bytes contract at every consumer
sie_server: install cv2 system libs for docling extract
streaming: no-silent-drop on chunk-queue backpressure (H6)
terraform: detect in-flight workflow runs via explicit status query
terraform: drop redundant Project overrides in quality-eval-l4
terraform: drop watchdog_idle_minutes default to 5, ignore PIP drift
terraform: grant ec2:DescribeInstanceStatus to wake role
terraform: per-runner idle-stop via GitHub Actions runners API
terraform: require positive activity observation before idle-stop
terraform: scope quality-eval IAM on Role tag instead of Project
terraform: seed GPU node-group desired_size from min_size
terraform: watchdog backstop covers queued-status runs
terraform: watchdog grants HANG_MINUTES grace from LaunchTime
terraform: watchdog ignores LastBusyAt older than LaunchTime
tester-cluster: pin worker pool nodeSelectors to gpu-type as well

Performance Improvements

dashboard: warm one Lambda, bump timeout, narrow S3 verdict fetch

Reverts

release-docker: drop matrix consolidation, keep deps push retry

v0.3.4 (2026-05-14)

Highlights

New capabilities: default payload store to model-cache bucket /payloads; typed InputTooLongError for extract 400 INPUT_TOO_LONG
Reliability and operations: bump dev-g6-spot to g6.2xlarge; bump dev-g6-spot to g6.2xlarge so default worker pool fits; default workers to shared queue pool; pin opencv-python-headless to drop X11 runtime deps; tolerate config conflicts in bootstrap, gate on sie-config

Features

infra: default payload store to model-cache bucket /payloads
sdks: typed InputTooLongError for extract 400 INPUT_TOO_LONG

Bug Fixes

aws-example: bump dev-g6-spot to g6.2xlarge
aws-example: bump dev-g6-spot to g6.2xlarge so default worker pool fits
chart: default workers to shared queue pool
cluster.py,aws.py: address review suggestions 1 & 2
deps: pin opencv-python-headless to drop X11 runtime deps
gateway: tolerate config conflicts in bootstrap, gate on sie-config
gateway: tolerate config conflicts in bootstrap, gate on sie-config ready
sdk: widen sie-sdk requires-python to >=3.12
terraform-aws: default ECR creation off, prefix repo names with project_name
terraform-aws: trim slashes from ecr_repository_prefix
terraform-google-sie: wait for identity pool before binding WI
terraform: relax required_version from ~> 1.14.3 to >= 1.14

v0.3.3 (2026-05-13)

Highlights

New capabilities: add ColQwen3 + Nemotron ColEmbed v2 visual doc retrieval; add text classification task support; add post-download load timeout with stall-based download bounds; add scope-able workflow_dispatch with model/profile/task filters; add new INPUT_TOO_LONG ErrorCode; enforce overflow_policy in gliclass adapter
Reliability and operations: surface empty matrix and add measurement-mode for unbaselined adapters; emit task_class in quality-adapter JSON output; score detection eval predictions from result[“objects”]; annotate empty-diff path that bypasses impact_map; capture real exit code from impact_map in resolve-impact.sh

Features

adapters: add ColQwen3 + Nemotron ColEmbed v2 visual doc retrieval
extraction: add text classification task support
model-loader: add post-download load timeout with stall-based download bounds
quality-adapter: add scope-able workflow_dispatch with model/profile/task filters
server: add new INPUT_TOO_LONG ErrorCode
server: enforce overflow_policy in gliclass adapter
server: route INPUT_TOO_LONG to HTTP 400 in extract API
server: validate overflow_policy in resolve_runtime_options

Bug Fixes

bench: emit task_class in quality-adapter JSON output
bench: score detection eval predictions from result[“objects”]
ci: annotate empty-diff path that bypasses impact_map
ci: capture real exit code from impact_map in resolve-impact.sh
ci: collapse adapter-equivalent profiles in quality-adapter matrix
ci: pin mise to 2026.5.5 in loadtest workflows
ci: surface empty matrix and add measurement-mode for unbaselined adapters
docker: install libspatialindex-c6 in worker images
gliclass: catch IndexError empty-tensor crash as InputTooLongError
gliclass: raise InputTooLongError from argmax-empty backstop
probe-chart: make 3-OLD/2-NEW sample asymmetry explicit in title
quality-adapter: namespace quality_eval tests to avoid conftest collision
quality-adapter: split adapter_paths on commas before —changed-dirs
quality-adapter: split Pair column so pair_key | stops breaking the table
quality-adapter: split Pair column to stop pair_key | breaking the table

v0.3.2 (2026-05-08)

Highlights

New capabilities: default score_pairs() in BaseAdapter + baseline reranking targets; bump cold-start schema to v6 with deserialize/warmup split; per-model perf concurrency defaults for OCR adapters; adapter-triggered quality eval on persistent L4 runner; make destroy conditional on workflow_dispatch input; nightly loadtest pipeline + baseline recorder
Reliability and operations: raise loadtest job timeout to GH Actions ceiling (360 min / 6h); clarify experimental NATS health mode; lang tag on fenced block; fail fast on invalid scenarios; surface error/no_results rows in MD; harden parse_label against unexpected filenames; adapt OCR perf Item shape to model.inputs and fail loudly on errors

Features

adapters: default score_pairs() in BaseAdapter + baseline reranking targets
bench,charts: bump cold-start schema to v6 with deserialize/warmup split
bench: per-model perf concurrency defaults for OCR adapters
ci: adapter-triggered quality eval on persistent L4 runner
ci: make destroy conditional on workflow_dispatch input
ci: nightly loadtest pipeline + baseline recorder
colbert: add score_pairs support and expand model coverage
dashboard: add status and kind filters to runs list
dashboard: introduce run-group concept (run = 3 scenarios)
dashboard: loadtest results dashboard (Next.js + SST + DynamoDB)
dashboard: render every metric in the perf-lab archive
dashboard: scaffold loadtest dashboard (Next.js + SST)
dashboard: status and kind filters on runs list
dashboard: track run_status; gh-API one-time backfill
docling: add ocr profile defaulting do_ocr=true
gateway: expose OpenAPI contract
gateway: unify API errors and align probe contracts
helm: expose probes value trees for worker/gateway/config
helm: tighten startup/readiness probes for faster pod-ready
helm: TLS termination via cert-manager + BYO matrix docs
helm: wire probe templates to values trees
infra: opt-in S3 cluster model cache
ltfr: cache-vs-no-cache compare chart, with-cache run data, and 8 single-mode chart refresh
matrix: add task_class stamping to eval measurements
server,bench: split deserialize/warmup in cold-start instrumentation (v6)
server: cap torch CPU threads at worker startup
sie_server: per-stage timing markers in lifespan for engine_boot attribution
sie_server: split adapter.warmup() out of load() with cold-start log markers
tools: bump cold-start bench to v5 with scenario flag
tools: LTFR per-scenario bench tooling + results (issue #652)
tools: ltfr-bench orchestrator (issue #652)

Bug Fixes

bench: adapt OCR perf Item shape to model.inputs and fail loudly on errors
bench: address CodeRabbit review on PR #779
bench: correctly detect v6 split presence in flattened runs[]
bench: derive emitted gpu_load_s from v6 deserialize+warmup split when available
chart: pass —cluster-cache to sie-server and correct populate command in docs
charts: vertical legend so ‘image pull + container init’ and ‘node prov’ aren’t clipped
ci+terraform: three deterministic root causes for loadtest pipeline
ci: address CodeRabbit findings on quality-adapter PR
ci: address CodeRabbit’s second-pass review on quality-adapter
ci: address CodeRabbit’s third-pass review on quality-adapter
ci: auto-clear stale terraform state lock from prior runner crashes
ci: drop double cuda12 suffix + force codebuild for missing images
ci: forensic dump on argo failure + LB-ENI release before destroy
ci: gate stale-lock clear behind force_unlock input + pass —aws-region to destroy
ci: override registry/gpu-selector/tolerations + Python heredoc
ci: parse markdown bench output → result.json synthesis
ci: pass WORKFLOW env to run_scenarios.sh in loadtest.yml
ci: preflight env-var check in run_scenarios + finalize scripts
ci: provision GH PAT secret + in-cluster github-token before bootstrap
ci: raise loadtest job timeout to GH Actions ceiling (360 min / 6h)
ci: read bench-config from local clone, not raw.githubusercontent.com
ci: right-size bench pod + worker pod resources for cluster shape
cluster: move orphan-LB sweep into cmd_destroy, drop parallel script
cluster: use project_name (not example name) for orphan-LB VPC tag
cluster: use project_name for orphan-LB VPC tag lookup
dashboard,ci: keep run_status consistent between S3 and DynamoDB
dashboard,ci: wire real Prometheus matrix shape + extend headlines
dashboard: drop time-based legacy run grouping (was unsafe)
dashboard: GPU util shown as 0-100 (was being multiplied by 100 again)
dashboard: include duration_seconds in run-meta.json (was DynamoDB-only)
dashboard: normalize array-shaped searchParams before .trim()
deps: bump plotly to >=6.1.1 for kaleido compat
docker: stub bundles/ and models/ in deps stage
docling: cache DocumentConverter per (device, ocr_enabled)
docling: mark adapter unloaded in unload()
docling: thread device through PdfPipelineOptions accelerator_options
gateway: address latest coderabbit contract notes
gateway: address PR review for NATS health mode
gateway: address probe and SDK review findings
gateway: align CreatePoolRequest OpenAPI with runtime validation
gateway: clarify experimental NATS health mode
gateway: close remaining review contract gaps
gateway: preserve embeddings timing headers
gateway: preserve scale-from-zero request path
gateway: reject unsupported embeddings token arrays
helm: validate ACME server and privateKeySecretRef in validateTls
infra: grant kms:Decrypt to workers when model cache uses SSE-KMS
infra: normalize whitespace-only model cache string inputs
infra: treat empty model_cache_kms_key_id as unset
infra: use flat lifecycle key for s3-bucket module v5
loadtest-ci: force-delete orphan elbv2 LBs before terraform destroy
loadtest-ci: force-delete orphan LBs and stop swallowing destroy failures
loadtest-ci: poll workflow phase instead of argo submit —wait
loadtest-ci: poll workflow phase instead of relying on argo —wait
ltfr-bench,notes: lang tag on fenced block; fail fast on invalid scenarios; surface error/no_results rows in MD
ltfr-bench: hoist imports to top; guard payload.results shape
ltfr-bench: mark request_failed rows in scenario-a/b MD tables
ltfr-bench: preserve failure context in aggregated rows; add request_failed status
ltfr-bench: treat no_results cells as failures in exit code
ltfr-charts: strip legend clip-path so labels render full width
ltfr: tighten UID/timestamp guards in capture_image_pull_events
multi_pod_cold_start: raise on ASG terminate fail; UID-filter pull events; isolate scenario-c pod
paddleocr_vl: pass use_cache=True to generate
paddleocr_vl: pass use_cache=True to generate to enable KV-cache
review: tighten score_pairs options handling and query text validation
sie-server: include model and bundle directories in wheel distribution
terraform: detect HF-cache EBS by NVMe model + size, not by Linux name
terraform: set resolve_conflicts_on_* = OVERWRITE on EKS addons
tools: drop module-level docstrings (AGENTS.md rule)
tools: guard fig_per_cell_table aggregation against empty results
tools: guard mean() against empty engine_boot_s in aggregate()
tools: harden parse_label against unexpected filenames
tools: mark cold-start-bench.py executable (EXE001)
tools: remove module docstring from cold_start_charts.py (repo rule)

v0.3.1 (2026-04-29)

Highlights

New capabilities: add OmniDocBench OCR quality loader; support /v1/score with dense/sparse/colbert/hybrid modes; add Marqo/marqo-ecommerce-embeddings-B via open_clip backend
Reliability and operations: add terminal failed state to model registry (sie-test#85)

Features

bench: add OmniDocBench OCR quality loader
bge-m3: support /v1/score with dense/sparse/colbert/hybrid modes
siglip: add Marqo/marqo-ecommerce-embeddings-B via open_clip backend

Bug Fixes

server: add terminal failed state to model registry (sie-test#85)

v0.3.0 (2026-04-29)

Highlights

Breaking change: openapi.json is now a committed artifact that must be regenerated and committed when API changes are made
New capabilities: add GLM-OCR adapter; add Qwen3-VL-Embedding-2B and Qwen3-VL-Reranker-2B multimodal adapters; add GLiNER2 and GLiNER-bi adapters; add Qwen3-Reranker-0.6B and 4B causal LM reranker support; add SigLIP 2 base-patch16-224 vision-language encoder; add minimal cache weights snapshot command for offline deployments
Reliability and operations: retry only transient connection errors under wait_for_capacity; surface unrouteable models loudly and helm-repo-add on pristine hosts; emit identical NATS payload to bundle and _all subjects; surface mixed-profile unrouteable models and keep snapshot consistent on writes; add retry logic for deadsnakes PPA to handle Launchpad outages
Performance: cache JPEG-encoded corpus images across queries; lazily JPEG-encode corpus images on first use; cache SDK version parse, integer audit latency, UUIDv7; cut hot-path allocations, fuse numpy decode, tighten backpressure

⚠ BREAKING CHANGES

openapi: openapi.json is now a committed artifact that must be regenerated and committed when API changes are made

Features

adapters: add GLM-OCR adapter
adapters: add Qwen3-VL-Embedding-2B and Qwen3-VL-Reranker-2B multimodal adapters
add GLiNER2 and GLiNER-bi adapters
add Qwen3-Reranker-0.6B and 4B causal LM reranker support
add SigLIP 2 base-patch16-224 vision-language encoder
admin: add minimal cache weights snapshot command for offline deployments
bench: honor SIE_BENCH_SERVER_READY_TIMEOUT in eval orchestrator
ci: nightly loadtest gate against dedicated EKS cluster
ci: nightly loadtest gate, ephemeral cluster per run
extract: add Docling adapter for PDF/DOCX/HTML extraction
extract: add Docling adapter for PDF/DOCX/HTML parsing
extract: plumb document items and structured data results
observability: add Prometheus metrics to sie-config and expand sie-gateway coverage
observability: Prometheus metrics for sie-config and sie-gateway
oom: implement defensive exception fan-out and improve recovery metrics
oom: improve error semantics and budget exhaustion detection
openapi: add static spec export and validation
router: import Rust gateway source tree
server: add reactive OOM recovery and proactive idle eviction
social: daily social content pipeline with 5-source drafts + engagement
types: add document input modality across SDKs, server, and metadata

Bug Fixes

adapters: add input validation guards for empty/failed visual inputs
adapters: address review findings for Qwen3-VL adapters
adapters: clarify video placeholder, validate token IDs, fix torch_dtype key
add client-side hour filter to search_x_posts (was date-level only)
address CodeRabbit review feedback
address follow-up PR review nits
address remaining CodeRabbit feedback (round 2)
address review findings — negative truncation guard, score() options, constant dedup
bench: show correct unit labels for MP/s throughput in —print-gap
bundles: declare Qwen3-VL adapters in default bundle
ci: use Blacksmith runner in CI
client: retry only transient connection errors under wait_for_capacity
cluster: address PR #701 review comments
cluster: correct kubectl flag combo and reorder LB sweep before helm uninstall
cluster: helm uninstall before terraform destroy to clean up AWS LB leftovers
cluster: unblock end-to-end mise run cluster create --build
config,cluster: surface unrouteable models loudly and helm-repo-add on pristine hosts
config: emit identical NATS payload to bundle and _all subjects
config: surface mixed-profile unrouteable models and keep snapshot consistent on writes
docker: add retry logic for deadsnakes PPA to handle Launchpad outages
docker: propagate failure when all add-apt-repository retries exhausted
docling: per-task converter, hf_revision guard, callable typing (CodeRabbit)
docs: Update packages/sie_server/Dockerfile.cuda11
fail closed on missing/unparsable timestamps in lookback filter
gateway,config,sdk: resiliency, concurrency, and cross-service hash parity
gateway,config: address PR review — 404 for unknown models, 202 on default routing, full YAML propagation
gateway,config: harden auth, trusted NATS producers, and recovery path; drop gateway HA default
gateway,sdk: map upstream timeouts to 503+MODEL_LOADING for SDK retry
gateway: add GET /v1/models/{model} detail route
gateway: address PR #716 review feedback
gateway: align /v1/models error and list shapes
gateway: drop double-counted REQUEST_COUNT / REQUEST_LATENCY emit
gateway: emit X-SIE-Error-Code header on model-loading 503
gateway: keep record_request async to match main’s call shape
gateway: make sie-config single source of truth for bundles with live resync
gateway: normalize model ids in NATS work subjects + docs/tooling/ha cleanup
gateway: pre-instantiate request/demand metric families on startup
gateway: prioritize epoch-rewind branch; harden no-thrash test; correct arch-guide on ephemeral restart
guard score() and score_pairs() against empty input lists
helm: default clusterRouting to “queue” on import-sie-router-rust
helm: enable NATS + JetStream by default to match queue clusterRouting
helm: fail fast when gateway has no bundle source
kind-smoke: add —no-pool-isolation for static clusters + contract-drift fixes
kind-smoke: address bot review feedback
kind-smoke: enable configStore and harden config/gateway tests
kind-smoke: enable JetStream on test NATS and drop duplicate subchart
kind-smoke: wire sie-config image and helm overrides into kind cluster fixture
kind-smoke: wire sie-config image into kind cluster fixture
observability: address PR review blockers on metrics PR
sdk: cluster cache prefix probe uses list, not head
sdk: cluster cache prefix probe uses list, not head (Refs #732, #654)
sdk: has_children filters folder-marker objects (Refs #732, #654)
sdk: preserve caller-supplied document format over inferred (CodeRabbit)
sdk: retry mid-flight transport disconnects, not just timeouts
sdk: retry on connection errors and generic 503s
sie_config: address PR review feedback
terraform/aws: set 100GB root volume on cpu node group to avoid DiskPressure
terraform/gcp: undo router→gateway rename on GCP Cloud Router + NAT
tests: include sie-config in expected missing-image list
tests: restore docker gateway smoke test after router rename
tmux-scripts: improve robustness of session parsing and argument handling
types: adapt to ty 0.0.32 stricter ignore handling
use searchTerms for X tweet-scraper actor (was searchQueries)

Performance Improvements

bench: cache JPEG-encoded corpus images across queries
bench: lazily JPEG-encode corpus images on first use
docker: add —link + move ARG BUNDLE to eliminate cross-bundle layer noise
docker: normalize mtimes so shared venv layer is dedupable
docker: reorder stages for maximum BuildKit cache reuse
docker: split worker venv into shared + bundle-specific layers
gateway: cache SDK version parse, integer audit latency, UUIDv7
gateway: cut hot-path allocations, fuse numpy decode, tighten backpressure
gateway: fuse msgpack_numpy decode into the response path
gateway: move score-endpoint unwrap instead of cloning
gateway: pass msgpack items through as rmpv::Value
gateway: publish work items concurrently + borrow shared fields
gateway: tighten cold-pool backpressure + cheaper QPS counter
gateway: trim per-request work on the inference hot path

v0.2.0 (2026-04-17)

Highlights

Breaking change: Removed --model CLI args from worker startup; use SIE_PRELOAD_MODELS env var or --preload flag instead
New capabilities: add ModernBERT flash dense embedding support with fallback mechanism; add OCR quality benchmarks (olmOCR-bench); add OCR quality benchmarks with olmOCR-bench; add pages/sec throughput metric for OCR perf eval; add perf metrics to OCR eval pipeline; also report query throughput in mpix/s for image queries
Reliability and operations: add missing NATS Helm repo to release workflow; don’t set NODE_AUTH_TOKEN for OIDC npm publishes; harden affinity spill with bounds check, clamp, and debug log; make rejected requests visible to KEDA scaling metrics; remove redundant tokenizer validation and unused template parameter

⚠ BREAKING CHANGES

workers: Removed --model CLI args from worker startup; use SIE_PRELOAD_MODELS env var or --preload flag instead

Features

adapters: add ModernBERT flash dense embedding support with fallback mechanism
bench: add OCR quality benchmarks (olmOCR-bench)
bench: add OCR quality benchmarks with olmOCR-bench
bench: add pages/sec throughput metric for OCR perf eval
bench: add perf metrics to OCR eval pipeline
bench: also report query throughput in mpix/s for image queries
benchmarks: add MTEB NFCorpus evaluation results for ModernBERT-based embedders
bench: report vision corpus throughput in mpix/s
bench: report vision corpus throughput in mpix/s instead of items/s
deps: migrate from pynvml to nvidia-ml-py package
haystack: add haystack_integrations namespace aliases
haystack: add namespace-convention aliases
observability: add anonymous usage telemetry
sdk: add max_concurrency param to SIEAsyncClient to prevent connection pool exhaustion
server: add lightonai/LightOnOCR-2-1B OCR adapter with next bundle
workers: implement model preloading at startup to reduce first-request latency

Bug Fixes

adapters: remove redundant tokenizer validation and unused template parameter
address PR review — panel title, namespace variable
bench: handle unloaded images in pixel count computation
bench: use concurrent async requests for OCR perf eval
bench: validate image entries before computing pixel counts
bench: validate pixel counts before using them for image corpus throughput
build: downgrade dockerfile syntax version to 1 for broader compatibility
ci: add missing NATS Helm repo to release workflow
ci: don’t set NODE_AUTH_TOKEN for OIDC npm publishes
dashboard: queue routing dashboard accuracy and usability
docs: add update date to portfolio header
docs: correct PR reference in reranker reclassification note
docs: populate reranker data and simplify table header
docs: update stale model counts after reranker reclassification
haystack: rename namespace alias to sie
install uv via curl instead of COPY —from ghcr.io
preload smoke test checks model.loaded instead of nonexistent workers field
readme: heading format
release: add LanceDB integrations to release-please config
router: add overflow spill to break model affinity deadlock
router: harden affinity spill with bounds check, clamp, and debug log
router: make rejected requests visible to KEDA scaling metrics
sie-bench: account for in-flight drain in throughput calculation
sie-bench: use union wall-clock for multiprocess throughput merge
tester-cluster: patient KEDA scale-down for worker pools

v0.1.10 (2026-04-09)

Highlights

New capabilities: add async, chunking, and streaming to Weaviate document enricher; improve DLQ routing and score response handling; implement Config Management API with NATS-based distribution and review fixes; add LanceDB integration (Python + TypeScript); queue routing dashboard + NATS exporter + router image tag; queue routing dashboard, NATS prom exporter, router image tag
Reliability and operations: correct cluster routing condition, stream max_age units, and reconnect state ordering; add recreate strategy for router deployment when nats config restore is enabled; restore Chart.yaml deps from main, keep appVersion v-prefix; queue routing dashboard PromQL for NATS wait; configurable NATS fetch budget, Helm-wired queue params
Performance: decouple scanner and SIE batch sizes in enrich_table; stream enrich_table batch-by-batch instead of full materialization; use Lance scanner for column projection in enrich_table; bypass FastAPI for hot proxy paths via raw ASGI middleware

Features

add async, chunking, and streaming to Weaviate document enricher
dlq,pull-loop: improve DLQ routing and score response handling
implement Config Management API with NATS-based distribution and review fixes
integrations: add LanceDB integration (Python + TypeScript)
observability: queue routing dashboard + NATS exporter + router image tag
observability: queue routing dashboard, NATS prom exporter, router image tag
sdk: add get_model() and configure LanceDB release workflows
terraform: add AWS eval-eu EKS cluster with multi-GPU support
terraform: add evaluation cluster setup for AWS with multi-GPU support and updated configurations
terraform: add node labels, adjust pool sizes for tester cluster

Bug Fixes

config,queue,nats: correct cluster routing condition, stream max_age units, and reconnect state ordering
handle BytesIO images in LlamaIndex and validate Weaviate classify config
helm: add recreate strategy for router deployment when nats config restore is enabled
helm: restore Chart.yaml deps from main, keep appVersion v-prefix
helm: use generic release-please updater for appVersion
helm: use generic updater for both Chart.yaml version fields
helm: use l4-spot/rtx6000-spot naming convention for spot profiles
integrations: address CodeRabbit review findings for LanceDB PR
observability: queue routing dashboard PromQL for NATS wait
queue-routing: configurable NATS fetch budget, Helm-wired queue params
queue-routing: resolve bugs, add configurable NATS params, fix score wire format
queue-routing: score response format and DLQ fallback routing key
release: use NPM_TOKEN for initial sie-lancedb publish
router: use “scores” key in queue-mode score responses
terraform: add GPU subnet coverage validation
terraform: relax AZ validation and clarify defaults
terraform: review fixes for tester cluster infra
terraform: switch tester-cluster to us-east-2
terraform: Switch tester-cluster to us-east-2 and update deployment docs
terraform: validate gpu_node_groups for duplicate and reserved names
test: add buildx builder pause recovery and improve build error diagnostics
update adapter tests and address code review feedback
use OCI registry URI for helm chart in README

Performance Improvements

lancedb: decouple scanner and SIE batch sizes in enrich_table
lancedb: stream enrich_table batch-by-batch instead of full materialization
lancedb: use Lance scanner for column projection in enrich_table
router: bypass FastAPI for hot proxy paths via raw ASGI middleware
router: reduce thread pool pressure by inlining small deserialization
router: remove msgpack_numpy global patch and BaseHTTPMiddleware
router: replace stdlib json with orjson for 3-10x faster serialization
sdk+router: lazy msgpack_numpy.patch and pure ASGI middleware

v0.1.9 (2026-04-02)

Highlights

Reliability and operations: increase docker smoke test timeouts and add retry; include $platform in worker image tag format; revert pool names to machine profile names; remove —provenance flag (requires public repo)

Bug Fixes

helm: include $platform in worker image tag format
helm: revert pool names to machine profile names
increase docker smoke test timeouts and add retry
remove —provenance flag (requires public repo)

v0.1.8 (2026-04-01)

Highlights

Reliability and operations: add sie-qdrant and sie-weaviate to release-please config; point sync-terraform default repos to production; remove —provenance from npm publish for private repo; correct image.tag comment to reflect actual format; remove duplicate platform suffix from worker image tag

Bug Fixes

add sie-qdrant and sie-weaviate to release-please config
ci: point sync-terraform default repos to production
ci: remove —provenance from npm publish for private repo
helm: correct image.tag comment to reflect actual format
helm: remove duplicate platform suffix from worker image tag
remove internal-only references from COMPATIBILITY.md

v0.1.7 (2026-04-01)

Highlights

New capabilities: add profiling script for sparse encoding hot path; add GitHub Actions workflow to sync Terraform modules to registry repos; apply QoL improvements from PR #484 review comments; switch default GPU from g5 (A10G) to g6 (L4); add rerank/score support to TEI runner; implement configurable document length limits and custom prefix token registration
Reliability and operations: restore triggering ref for source checkout; restore quality by enabling causal attention and QK-normalization; restore dev-l4-spot zones to us-central1 for GPU availability; check /metrics endpoint in test_prometheus_metrics_exist; add per-attempt timeout to lease renewal fetch
Performance: optimize MoE expert dispatch with sorted-expert routing; batch MaxSim scoring across documents on GPU; batch sparse aggregation with segment_reduce and fuse relu; batch split_embeddings + validate ColBERT performance

Features

adapters: add profiling script for sparse encoding hot path
add GitHub Actions workflow to sync Terraform modules to registry repos
apply QoL improvements from PR #484 review comments
aws: switch default GPU from g5 (A10G) to g6 (L4)
bench: add rerank/score support to TEI runner
colbert: implement configurable document length limits and custom prefix token registration
deploy: move namespace, SA, and HF token secret management to Helm chart
deploy: prepare Terraform modules for public registry publishing
deploy: rewrite example module sources to registry references
deploy: rewrite Helm and internal references for public release
deploy: two-artifact model — GCP Terraform infra-only, batteries-included Helm chart
docker: add —docker-platform flag to docker build task
extend create_pool API/SDK with minimum_worker_count and bundle
helm: add batteries-included sub-chart dependencies to sie-cluster
helm: add image pre-pull DaemonSet for GPU worker pools
helm: add step to build Helm chart dependencies in Kind smoke tests
helm: default router to image-embedded model configs
helm: enable image pre-pull DaemonSet by default
helm: port health gates from Terraform to Helm post-install hooks
helm: remove prometheus alias, bump to v0.2.0, standardize chart
infra: add Modal GPU sandbox for remote benchmark execution
infra: add rollout warning and explicit image_type for GCFS
infra: enable GCFS image streaming on GPU node pools
infra: set min_node_count=1 on L4 spot GPU node pools
integrations: add Qdrant integration
integrations: add Qdrant integration with native sparse vector support
integrations: add Weaviate v4 integration with Go module spec
multiprocess loadtest + SDK aiohttp migration
sdk: add version negotiation headers between SDK and server
sdk: default wait_for_capacity=True and timeout=900s
sdk: version negotiation header (SDK ↔ server)
sie-bench: add dataset/input_type fields for mTEB corpus inputs
sie-bench: built-in multiprocess loadtest mode
skills: add eval-model skill for HF model assessment
skills: add eval-model skill for HF model integration assessment
sync Terraform modules to registry repos
tei-runner: add /embed_sparse support for sparse models
tei: add /embed_sparse support and auto-detect pooling mode
terraform/aws: restore cluster autoscaler helm release to infra module
terraform/aws: strip k8s resources, restructure as infra module with examples
terraform: add cluster name and artifact registry variables; update node pool configuration
terraform: add EBS CSI driver, NVIDIA device plugin, default StorageClass
terraform: strip gcp k8s/ layer; examples use infra-only module
tools: add ColBERT query vs document profiling script
tools: add dense P50 latency profiling script

Bug Fixes

adapters: sort IDF unique_ids to satisfy SparseVector contract
add missing production example to tf validate; fix tempfile leak; remove module docstring
address PR #478 review feedback
address PR review — GPU alert formula, kubectl parsing, CI path filter
address review feedback for npm publish
address review findings - race prevention, cleanup, lighter checkout
alloy: add stage.cri{} before stage.json to unwrap CRI log envelopes
alloy: explicitly set configMap name and key for sub-chart wiring
alloy: scope pod discovery to current node via field selector
bench: complete g5 to g6 migration in AWS eval configs and GPU mapping
benchmarks: use TEI /embed_all for ColBERT multi-vector models
bench: skip loading candidates_model for single-model servers
chart: update home URL and Helm install command in README
CI compatibility and consistent env var usage
CI compatibility for sync-terraform workflow
ci: add contents: read permission to publish-pypi-oidc job
ci: add helm repo add + dep build to kind-smoke workflow
ci: g5 refactored to g6 already
cluster: build concrete helm command in status from infra_outputs
cluster: guard helm/kubectl post-create log when outputs are empty
deploy: clean terraform init artifacts before push
deploy: correct smoke test TypedDict access and helm dry-run args
deploy: correct StatefulSet rollout semantics, PDB scope, and KEDA pause
deploy: remove dangling kubernetes_namespace_v1.sie references from health_gates.tf
deploy: restore triggering ref for source checkout
deploy: update default destination repos for GCP and AWS modules
deploy: use triggering ref for source checkout in sync-terraform
disable LoRA adapter layers after loading to prevent quality corruption
docs: clarify optional image push in AWS and GCP README files
docs: update Helm chart path in AWS and GCP README files
fix integration test
helm,hook: deploy/helm/sie-cluster/templates/hooks/prometheus-ready-test.yaml
helm: add before-hook-creation to Job delete policies; document count==0 expectation
helm: address coderabbit findings on health gate hooks
helm: address non-blocking review findings from PR #336
helm: address review findings in batteries-included sub-chart PR
helm: address reviewer suggestions for health gate hooks
helm: address second-pass review findings
helm: aggregate buckets by le in p95 latency alert
helm: bump chart version to 0.1.1 (patch, not minor)
helm: clarify prometheusAddress comment — ignored when sub-chart is installed
helm: correct kube-prometheus-stack semver constraint and remove hardcoded grafana password
helm: correct misleading validation comment in router-deployment.yaml
helm: don’t emit ScaledObject CRDs unless KEDA is confirmed present
helm: downscope KEDA RBAC to Role/RoleBinding; remove runtime apk installs
helm: fix loki service URL and extract alloy config to file
helm: fix three blocking review issues in sub-chart dependencies
helm: improve temporary values file handling in helm_template function
helm: improve, simplify, and modularize sie-cluster chart
helm: move ‘app.kubernetes.io/part-of’ label to selector labels for consistency
helm: remove autoscaling.enabled from values-aws.yaml
helm: render KEDA ScaledObjects via post-install hook to avoid CRD chicken-and-egg
helm: replace hardcoded namespace in provisioning alert rules
helm: require non-empty hfToken.value when hfToken.create is true
helm: sub-chart naming, Loki compactor, event exporter ECR, Grafana folders
helm: use autoscaling.prometheusAddress in prometheus hook; remove stub health_gates.tf
helm: use full FQDN for Prometheus service in KEDA and health gates
helm: use router.service.port in NOTES.txt instead of hardcoded 8080
infra: update min_node_count default in top-level GCP module
normalize SDK version warned-set key to major.minor
pool error types, add pool/progress test coverage
profiling: add flash variant registry, device validation, top-level import
profiling: sync GPU before tensor timing, move script to tools/
profiling: use in-place relu_ to match production code path
qwen3: restore quality by enabling causal attention and QK-normalization
readme: correct helm chart path
readme: correct helm install command
release: track all package versions via release-please extra-files
release: track TS SDK version.ts via release-please
replace corrupted bge-m3 NanoFiQA2018 target + set bfloat16 precision
review items
router: increase pool lease TTL to survive rolling upgrades
router: resolve default pool GPU for scale-up when gpu/pool omitted
router: use effective_pool instead of pool_name for default pool GPU extraction
sdk: defer aiohttp session creation to fix “no running event loop” in SIEAsyncClient
sie_bench: improve —print-gap report accuracy and readability
tei-runner: validate /embed_all returns per-token embeddings
tei-runner: validate output_type in TEIRunner init
terraform/aws: add full -backend-config flags to production init command
terraform/aws: add precondition asserting >=2 GPU-capable AZs exist
terraform/aws: address review findings post-restructure
terraform/aws: correct helm chart path in dev-g5-spot example comment
terraform/aws: filter VPC AZs to only zones offering the GPU instance type
terraform/aws: fix invalid splat on instance type offerings locations
terraform/aws: remove provider aws block from child module
terraform/aws: use var.project_name in VPC subnet cluster tags
terraform: add validation for GPU node pool zones to ensure they match the configured region
terraform: restore dev-l4-spot zones to us-central1 for GPU availability
terraform: update GPU instance type description for clarity and add dev-g6-spot example
terraform: update stale k8s module references in comments
terraform: upgrade AWS modules and fix deprecations
test: check /metrics endpoint in test_prometheus_metrics_exist
test: update EKS tests from g5 to g6 after GPU instance type change
ts-sdk: add per-attempt timeout to lease renewal fetch
use prepack instead of prepublishOnly
validate minimum_worker_count input and soften docstrings

Performance Improvements

adapter: optimize MoE expert dispatch with sorted-expert routing
adapters: batch MaxSim scoring across documents on GPU
adapters: batch sparse aggregation with segment_reduce and fuse relu
adapters: batch split_embeddings + validate ColBERT performance
adapters: batch split_embeddings in ColBERT adapters
adapters: eliminate GPU overhead from IDF query encode path
florence2: greedy decoding for OCR (-23% P50)
florence2: switch OCR configs from beam search to greedy decoding
server,bench: add batch coalescing, query warmup, and benchmark stability improvements
server: dispatch immediately when worker is idle
server: optimize BertFlashAdapter inference path (+35% corpus throughput)
server: reduce batch wait timeout 10ms \u2192 2ms for lower Doc P50

v0.1.6 (2026-03-12)

Highlights

Breaking change: remove florence2 and gliner standalone bundles — extraction adapters (gliner, glirel, gliclass) are now included in the default bundle
New capabilities: add native MTEB reranking task support with MRR metric; encode-dense matrix eval — 3 models × 8 tasks; add date-prefixed versioning for chronological filename ordering; reorganize and expand model size lookup table with alphabetical ordering; add detailed perf metrics, metric filter, and threshold selector; add marimo benchmark dashboard notebook
Reliability and operations: restore /var/cache/apt mounts, keep /var/lib/apt removed; add HF_TOKEN auth and config kwargs and fix stella models; add dense projection support to Qwen2FlashAdapter; apply query_template from runtime options in SentenceTransformerDenseAdapter; replace pip install with uv add in docker error messages
Performance: vectorize GTE sparse encode path; vectorize tokenization and packing for gte-multilingual-base; vectorize tokenization and packing to reduce throughput gap; switch Qwen/GTE models to flash attention adapter

Breaking Changes

bundles: remove florence2 and gliner standalone bundles — extraction adapters (gliner, glirel, gliclass) are now included in the default bundle

Features

bench: add native MTEB reranking task support with MRR metric
bench: encode-dense matrix eval — 3 models × 8 tasks
benchmarks: add date-prefixed versioning for chronological filename ordering
benchmarks: reorganize and expand model size lookup table with alphabetical ordering
benchview: add detailed perf metrics, metric filter, and threshold selector
benchview: add marimo benchmark dashboard notebook
benchview: add perf metric selector to Model Size tab
benchview: detailed perf metrics, metric filter, threshold selector
router,bench,sdk: improve throughput with inflight tracking, batching, and connection pooling
server: typed request parsing with msgspec
sie_server: add gliner, glirel, and gliclass extraction dependencies to the default bundle

Bug Fixes

adapter: add dense projection support to Qwen2FlashAdapter
adapter: apply query_template from runtime options in SentenceTransformerDenseAdapter
apply CodeRabbit auto-fixes
bench: replace pip install with uv add in docker error messages
benchview: add missing statistics import and use _median helper
bundles: include sglang bundle in default cluster and eval-matrix configs
client: update websocket header parameter name from extra_headers to additional_headers
colbert: enable native mode fallback for non-CUDA devices and add Matryoshka truncation
deps: cap timm upper bound and fix lazy handler init
docker: clear stale apt lists before update to prevent 404s
docker: remove all apt cache mounts from Dockerfiles
docker: remove no-op /var/lib/apt cache mount from apt RUN blocks
docker: restore /var/cache/apt mounts, keep /var/lib/apt removed
helm: increase CPU worker pool memory limits for expanded default bundle
model: add missing query_template to stella_en_400M_v5
models: switch all-MiniLM-L6-v2 to SentenceTransformerDenseAdapter
multilingual-e5-large-instruct: use instruct query template, NFCorpus 0.3521 → 0.3567
replace invalid HTML entities in SVG with XML numeric entities
rope_flash: clear cached _rope_dummy on unload and use torch.cat for packing
router: also resolve pool-derived GPU names to spot variants
router: resolve bare GPU types to spot variants for KEDA scaling
sdk: resolve sync/async client inconsistencies in score() and encode()
server: centralize request validation to prevent 500s from malformed items
server: use BertFlashAdapter for e5-small-v2, resolve e5 perf anomalies
server: use BertFlashAdapter for intfloat/e5-small-v2 and remove stale benchmarks
set compute_precision to bfloat16 for stella_en_1.5B_v5
sie_server: add HF_TOKEN auth and config kwargs and fix stella models
sie_server: resolve BGE-M3 linear weights loading for HF model IDs and fix test fixtures
sie_server: support NV-Embed-v2 with PyTorch embedding adapter
splade: align special token filtering and guard empty batches
test: always rebuild Docker images to pick up code changes
typecheck: move ty type checker from mise tool to uv dependency

Performance Improvements

adapters: vectorize GTE sparse encode path
rope_flash: vectorize tokenization and packing for gte-multilingual-base
rope_flash: vectorize tokenization and packing to reduce throughput gap
server: switch Qwen/GTE models to flash attention adapter
splade: vectorize tokenization and sparse aggregation (1.5x throughput)
splade: vectorize tokenization and sparse aggregation in SPLADEFlashAdapter

v0.1.5 (2026-02-27)

Highlights

New capabilities: add GLiNER v2.5 model configs; stream request bodies through proxy instead of buffering; add classification model configs for GLiClass-large and cross-encoder NLI
Reliability and operations: release pipeline cache collision and smoke test timeout; strip content-length header from streamed proxy responses
Performance: stream response body to eliminate bytes.join bottleneck

Features

models: add GLiNER v2.5 model configs
router: stream request bodies through proxy instead of buffering
sie_server: add classification model configs for GLiClass-large and cross-encoder NLI

Bug Fixes

release pipeline cache collision and smoke test timeout
router: strip content-length header from streamed proxy responses

Performance Improvements

router: stream response body to eliminate bytes.join bottleneck

v0.1.4 (2026-02-27)

Highlights

Reliability and operations: revert sharing=locked, add cache-read-only for build step

Bug Fixes

revert sharing=locked, add cache-read-only for build step

v0.1.3 (2026-02-26)

Highlights

Reliability and operations: revert token lifetime extension, re-auth before push instead; revert token lifetime, re-auth before push

Bug Fixes

revert token lifetime extension, re-auth before push instead
revert token lifetime, re-auth before push

v0.1.2 (2026-02-26)

Highlights

Reliability and operations: release image builds failing from GCP token expiry

Bug Fixes

release image builds failing from GCP token expiry

v0.1.1 (2026-02-26)

Highlights

Reliability and operations: update bundle definitions to replace legacy and gte-qwen2 with gliner

Bug Fixes

bundles: update bundle definitions to replace legacy and gte-qwen2 with gliner

v0.1.0 (2026-02-26)

Highlights

Breaking change: HTTP 409 dependency conflict responses are removed from all API endpoints; the DEPENDENCY_CONFLICT error code no longer exists; .beads/ issue tracking data removed from repository
New capabilities: add X-SIE-Worker response header for per-worker metrics tracking; add encode-image-text measurements to benchmarks dir; add encode-multivector perf measurements; add encode-multivector performance measurements; add encode-visual-document perf measurements; add encode-visual-document performance measurements
Reliability and operations: increase helm install timeout from 10m to 15m; add trailing empty line to gitignore; align release-images workflow with docker task flags; register GLiClass and DeBERTa models in bundles; build and deploy gliner bundle in Kind smoke tests
Performance: add connection pooling load test results (Feb 24); pool httpx client and add X-SIE-Worker header in router proxy; pool httpx client in router proxy to eliminate per-request TCP overhead; move transformers imports to module level

⚠ BREAKING CHANGES

deps: HTTP 409 dependency conflict responses are removed from all API endpoints; the DEPENDENCY_CONFLICT error code no longer exists
.beads/ issue tracking data removed from repository
deps: model config files no longer support the dependencies field

Features

add X-SIE-Worker response header for per-worker metrics tracking
benchmarks: add encode-image-text measurements to benchmarks dir
benchmarks: add encode-multivector perf measurements
benchmarks: add encode-multivector performance measurements
benchmarks: add encode-visual-document perf measurements
benchmarks: add encode-visual-document performance measurements
benchmarks: add extract-detection L4-SPOT performance measurements
benchmarks: add extract-kie-docvqa measurements to benchmarks dir
benchmarks: add extract-relation L4-SPOT performance measurement
benchmarks: add score-colbert perf measurements
benchmarks: add score-colbert performance measurements
models: add encode-image-text measurements
models: add extract-detection measurements
models: add extract-kie-docvqa measurements
models: add extract-relation measurements
router: add structured audit logging for API requests

Bug Fixes

.claude: add trailing empty line to gitignore
align release-images workflow with docker task flags
bundles: register GLiClass and DeBERTa models in bundles
ci: build and deploy gliner bundle in Kind smoke tests
colbert: remove CUDA requirement and improve device compatibility
eval: read ‘sie_id’ instead of ‘name’ from model configs in runner
extract: use dict access for Entity TypedDict in sort
gliner: relax stale transformers<4.52 pin
increase helm install timeout from 10m to 15m
reduce cpu-gliner resource requests for Kind CI
router: read ‘sie_id’ instead of ‘name’ from model configs
server: migrate NLI adapter to classifications and improve API consistency
server: migrate nli_classification adapter and improve type annotations
server: populate classifications instead of entities in GLiClass adapter
use manifest mode for release-please and reset to v0.0.0
use nested .gitignore for .claude/ directory

Performance Improvements

add connection pooling load test results (Feb 24)
pool httpx client and add X-SIE-Worker header in router proxy
pool httpx client in router proxy to eliminate per-request TCP overhead
pytorch-embedding: move transformers imports to module level
server: use uvloop as default event loop for uvicorn

Reverts

keep CONTRIBUTING.md clone URLs pointing to sie.git

Miscellaneous Chores

remove beads, agent prompts, mypy refs; consolidate ty config

Code Refactoring

deps: move adapter dependencies from per-adapter pyproject.toml to bundle YAML
deps: remove model-level dependencies feature