# Screenshot mining: fixed multi-app frame

The page compares recorded responses on 100 fictional screenshots. Kanboard,
OrangeHRM, Moodle, osTicket and Uptime Kuma contribute 20 screens each. There are
36 application-qualified template labels and one shared fictional construction
cluster. These descriptive results do not establish population accuracy,
equivalence or independent replication. The archived 25-screen Firefly study
remains separate in [its original notice](../fixed25/SOURCES.md).

## The declared chart metric

The chart shows **Field agreement**, the equal-weight mean of per-screen
normalized agreement on the same 1,645 retained checks across all 100 screens.
The [nine-field visibility audit](https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/18e8c9ae848f55f5a5b2039bd79065b80b9c00ef/screenshot-request-correction/2026-10-14/audit/field-exclusions.json) removes four
truncated ticket subjects and five numeric inputs hidden by a Save overlay,
identically for every arm. The complete original gold has 1,654 checks. Extra-row
penalties remain in the stock score and are reported separately in this fixed-field
sensitivity, so every arm uses the same denominator.

This is literal field agreement, not semantic accuracy. Readable logo, label and
field-boundary differences remain. NFKC, dash and ellipsis substitutions,
digit-comma removal, whitespace removal and casefold come from the unchanged
registered scorer; diacritics are not removed. Valid-request invalid JSON is zero
on the full retained denominator. No answer is repaired.

A separate [exploratory prompt/format-boundary audit](https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/18e8c9ae848f55f5a5b2039bd79065b80b9c00ef/screenshot-request-correction/2026-10-14/audit/ambiguity-exclusions.json)
removes 85 total checks, including the original nine. Some are clear prompt
conflicts; others are plausible interpretations where the original gold can also
be defended. UK17 has no remaining checks, so this analysis uses 99 common
screens and 1,569 checks for every arm. It is not the chart metric. The audit
began with model misses and was not blinded; it is not a fresh independent
annotation of every semantic target. Both masks and every complete-arm result
were independently inspected and recomputed.

[Method and runnable scorer](https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/18e8c9ae848f55f5a5b2039bd79065b80b9c00ef/screenshot-request-correction/2026-10-14/audit/README.md), [declared chart results](https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/18e8c9ae848f55f5a5b2039bd79065b80b9c00ef/screenshot-request-correction/2026-10-14/audit/visibility-summary.json),
[all per-screen stock and visibility results](https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/18e8c9ae848f55f5a5b2039bd79065b80b9c00ef/screenshot-request-correction/2026-10-14/audit/visibility-sensitivity.json),
and [exploratory results](https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/18e8c9ae848f55f5a5b2039bd79065b80b9c00ef/screenshot-request-correction/2026-10-14/audit/ambiguity-summary.json) preserve the distinctions.

| Setting | Original stock, all 100 | Chart field agreement, all 100 | Exploratory agreement, 99 | USD per 1,000 screenshots |
| --- | ---: | ---: | ---: | ---: |
| Gemma 4 31B | 94.147016969% | 94.718445541% | 97.719095719% | $1.873180 |
| GPT-6.1 Sol, medium | 96.197025313% | 96.715543831% | 99.984217172% | $6.070240 |
| GPT-6 Luna, medium | 94.748095794% | 95.266614312% | 98.957741590% | $0.375847 |
| GPT-6 Luna, high | 95.068827802% | 95.587346321% | 99.262074088% | $0.414432 |
| Claude Sonnet 5.5, high | 95.838784291% | 96.556133219% | 100.000000000% | $7.688420 |
| Claude Haiku 5.5, medium | 91.491495079% | 92.008589096% | 94.643011310% | $0.411251 |
| Claude Haiku 5.5, high | 88.803668091% | 89.322186610% | 92.525252525% | $0.480826 |

Each model is charted at its best measured primary-quality setting. Luna high
scores above medium; both appear so the cheaper setting is visible. Haiku medium
is both cheaper and higher-scoring than high. Sonnet high is its sole corrected
setting. Sol medium and SIE retain their original complete 100-screen recordings.
No quality share is claimed in a heading.

## Corrected requests and complete outcomes

The [plan and exact runner](https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/e0eb84babb14493fae96d6d14876f94c9ecdb4e8/screenshot-request-correction/2026-10-14/prereg/PREREGISTRATION.md) were public before any new
calls. Original [PNGs, instructions, schemas and gold](https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/77a6f4402054447580e55cd60c82ddfcd9813497/screenshot/2026-10-11/fixed-public-multi-app-fixtures-v2/FRAME.json) stayed
fixed. All five corrected settings attempted all 100 screens once, with no retry,
a 32,768-token output allowance and public provider APIs from one US East client.

Luna used Responses with strict original schemas and medium/high reasoning.
Sonnet used adaptive high effort; Haiku used adaptive medium/high effort. Claude
received the unchanged instruction plus the complete schema as plain JSON
prompting because its strict compiler rejects schemas above its union limit.
Every resulting answer is checked against the original schema. This is a supported
request alternative, not a dropped-field or repaired-answer comparison.

All 500 requests returned HTTP success. Luna medium/high and Sonnet each produced
100 schema-valid JSON objects. Haiku medium produced 99 JSON objects, 97 schema
valid; Haiku high produced 97 JSON objects, 93 schema valid. Their parse and schema
failures remain in the results. [Complete outcomes and usage](https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/b3f9e0b87b4bb99dca93424f890891dc8fb8dd55/screenshot-request-correction/2026-10-14/providers/corrected-summary.json),
[unchanged stock scoring](https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/b3f9e0b87b4bb99dca93424f890891dc8fb8dd55/screenshot-request-correction/2026-10-14/providers/corrected-scores.json), and [all request/response
hashes and declared identity removals](https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/b3f9e0b87b4bb99dca93424f890891dc8fb8dd55/screenshot-request-correction/2026-10-14/providers/manifest.json) are public.
Readable answers, exact requests and usage are unchanged by sanitization.

The original partial provider run is preserved in its [historical ledger](https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/bb61a3e5c41628f028a95cde5235bc32de28b14c/evaluations/2026-10-11/screenshot-multiapp-v2-20261011-v1/results/final/records.jsonl).
It had 85 Luna rejected configurations; Sonnet had 8 errors and 14 unsent calls;
Haiku had 3 errors and 19 unsent calls. Unsupported/unsent requests are not
model-quality failures and do not enter the corrected chart. Failure billing is
unknown in that historical packet, rather than assumed zero.

## Prices and conservative claims

The pricing module applies the recorded profile's Serverless rates, $0.60 input
and $2.00 output per million tokens, to 125,670 prompt tokens and 55,958 completion
tokens across 100 saved SIE calls. Prompt tokens include 12,421 cached tokens;
all are priced at the input rate. Completion already includes reasoning. This
is $1.87318 per 1,000 screenshots. The [original native study](https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/f11bfb25f6e244141df364d943d9e1d663592086/screenshot-multiapp-v2-native-sie/2026-10-12/results/records.jsonl)
and its [earlier quote correction](https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/eff7242e0a7a0242f299756ce37c92eff1c70303/screenshot-multiapp-v2-native-sie/2026-10-12/CORRECTION-2026-10-12.md)
remain unchanged. The page does not use the study's older $0.51/$1.36 quote.

Provider cost uses each response's reported tokens, including caching, at the
registered list rates. The same 100 requests define cost and agreement. Printed
prices round SIE upward and rivals downward to cents: $1.88 versus Sol $6.07
and Sonnet $7.68. Flooring savings from those printed prices gives 69% and 75%.
The cheaper Luna and Haiku alternatives stay visible. No latency claim is made
from this correction; original Rust-sidecar timing remains separate evidence.

## Two fixed pain cases

The preregistered rule chooses the first lexicographic case for the selected
model with a real scalar error, omitted visible field or wrong association that
persists at every corrected setting. Ambiguous targets, format-only differences,
rejected requests and failures repaired by stronger settings are excluded.
[Selection and response hashes](https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/18e8c9ae848f55f5a5b2039bd79065b80b9c00ef/screenshot-request-correction/2026-10-14/audit/pain-selection.json) record the review.

- HR17: both Luna medium and high return `Mia L` for the visible `Mía L`.
  This is literal name fidelity: an exact-string lookup differs. It does not
  demonstrate a different person or an actual downstream incident.
- MD13: both Haiku medium and high return null for the three requested visible
  reference IDs, `REF-612/OI`, `REF-712/A6`, `REF-812/ZŁ`.

Sonnet has no eligible unambiguous persistent factual error under this rule.
Its label and target-boundary differences are not presented as invented facts.
Display crops retain adjacent source context; full original PNGs remain linked.
Neither crop replaces a model input, and SIE output is absent from pain.

## Native SDK request and response

The API illustration is a separate, fixed KB01 HTTP recording using released
Python SDK 0.10.1, with [exact request](https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/c718fe42a8e914e122ffe8e5e20ac6f5a81bd6f9/screenshot-request-correction/2026-10-14/native/KB01-request.json), [response](https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/c718fe42a8e914e122ffe8e5e20ac6f5a81bd6f9/screenshot-request-correction/2026-10-14/native/KB01-response.json),
[validation](https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/c718fe42a8e914e122ffe8e5e20ac6f5a81bd6f9/screenshot-request-correction/2026-10-14/native/validation.json) and [terminal cleanup](https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/c718fe42a8e914e122ffe8e5e20ac6f5a81bd6f9/screenshot-request-correction/2026-10-14/native/native-terminal-verification.json).
One text warmup and one KB01 request ran on a single H200; the app is stopped.
No timing claim or aggregate quality claim comes from this illustration.

It uses `google/gemma-4-31B-it:h100-96k-hires-thinking-budget512-no-spec`,
checkpoint `842da3794eaa0b77d5f08bae87a17459d91ff475`, the released 0.10.0 CUDA13
image with the public server overlay at `ab4c9962d5c0a411b7832f0b7b50cc8ff8d57566`.
The original PNG and instruction are unchanged. The full strict schema is in
`response_format` only, with temperature 0, presence penalty 0 and output cap
32,768. All 18 root fields and ordered subtasks are retained; all 26 checks pass.

The actual generated Python was executed offline through SDK 0.10.1 with a mock
HTTP transport. Model/profile, exact instruction, image bytes, full strict schema,
sampling and output cap match the recording. Copy emits `max_completion_tokens`;
the recording uses `max_tokens`. The [pinned server alias handling](https://github.com/superlinked/sie/blob/ab4c9962d5c0a411b7832f0b7b50cc8ff8d57566/packages/sie_server/src/sie_server/api/openai_local.py#L762-L778)
normalizes both to the same cap. This is transport-equivalent, not byte-identical.
The page binds the recorded KB01 preset for Run and preserves the full request
in Copy. The earlier [local-ingest recipe](kb01-recipe.json) remains an archive,
not the HTTP request now illustrated.

## Rights and attribution

All names, organizations, monetary records and domains were authored as fictional fixtures. No private or customer data is included. [RIGHTS](https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/77a6f4402054447580e55cd60c82ddfcd9813497/screenshot/2026-10-11/fixed-public-multi-app-fixtures-v2/RIGHTS.json), [attributions](https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/77a6f4402054447580e55cd60c82ddfcd9813497/screenshot/2026-10-11/fixed-public-multi-app-fixtures-v2/ATTRIBUTIONS.md), [application revisions and notices](https://huggingface.co/datasets/superlinked/sie-task-evidence/resolve/77a6f4402054447580e55cd60c82ddfcd9813497/screenshot/2026-10-11/fixed-public-multi-app-fixtures-v2/APPLICATIONS.json) and the complete source packet preserve each upstream license. Project names identify source software without endorsement or additional trademark rights.

The shown Kanboard UI is copyright Frédéric Guillot and contributors, under [MIT](licenses/kanboard-MIT.txt). Font Awesome by Dave Gandy and contributors retains its [MIT code notice](licenses/font-awesome-3.2.1-license-code.less), [font and documentation terms](licenses/font-awesome-3.2.1-license.html.txt), [brand-icon notice](licenses/font-awesome-3.2.1-brand-license.html.txt) and [upstream README](licenses/font-awesome-3.2.1-README.md). Fixture records, prompts, schemas and generic helpers are [Apache-2.0](licenses/benchmark-Apache-2.0.txt). Corresponding application sources and modifications remain linked in the public packet.


OrangeHRM and Moodle display excerpts are covered by their application notices
and corresponding sources in the pinned source packet above. Fixture authorship
does not replace the upstream application licenses.