Samples / docs /ACCURACY_REVIEW.md
TidorP's picture
Review selected tasks, correct cyber reference cache identity, and clarify reproducibility setup
ddf04fa verified
|
Raw History Blame Contribute Delete
6.23 kB

Selected-task accuracy review — November 18, 2026

Scope: a targeted consistency review of the three GDPVal and two code samples, including source inputs, expected outputs and executable evidence. STEM was excluded from this review. This is not an exhaustive proof of reference correctness; passing an LLM judge alone is not treated as such proof.

GDPVal

German — family foundation. No material contradiction found between the prompt and the reviewed legal assertions in the rubric. The initial endowment allowance is conditioned on the potentially most distant beneficiary, not multiplied by the number of children. The rubric distinguishes later additions, preserves foundation ownership and capital-maintenance constraints, and identifies the 30-year substitute inheritance tax. The beneficiary interpretation agrees with the Federal Fiscal Court decision II R 25/21; the 30-year interval is in ErbStG §1. This is a 2025 planning exercise, not a current personalized legal opinion. The rubric repeats some documentation/timeline requirements, so a concise presentation can lose several points for one omission. A live native OpenAI run completed with 70.21%; that is partial answer credit, not a pipeline failure.

Japanese — NFT photography article. The supplied Word research file, requested output and rubric are consistent on the article title, Japanese character count, formatting, photographer examples and links. A live native OpenAI run produced a Word document and received 100%. Two qualifications matter: this uses a historical collection of web references, whose continuing availability is not guaranteed; resale royalties should be described as platform-dependent, not guaranteed for every NFT sale. For example, SuperRare documents a secondary-sale royalty. The rubric's wording permits earning royalties rather than requiring a universal guarantee. Its character-count boundary uses the first paragraph containing プルクオート: or キーワード; keep the keyword section at the end, since those words inside the body could make that boundary ambiguous. The original rubric is preserved.

French — 2024 household tax return. Independently recalculated the principal totals from the supplied salary, interest, dividend, securities-sale and payment PDFs. They reconcile with the rubric/reference:

Calculation Euros
Declared salaries: Laurent / Claire 234,934 / 92,069
Net taxable salaries after standard deductions 303,370
Progressive tax after family-quotient cap 85,108
Net securities gain 25,915
PFU income-tax component 4,323
Donation and school reductions 2,989
Childcare credit / 2CK credit 820 / 1,006
Net income tax before PAS payments 84,616
PAS payments 39,079
Remaining income tax / social charges 45,537 / 4,458
Final balance 49,995

The twins were under six on January 1, 2024; the questionnaire explicitly excludes the global progressive-rate option and prior-year carryovers. The 2BH entry is consistent with the official 2025 simulator's conditional CSG wording. Brackets were cross-checked against the 2025 BOFiP schedule, and deduction limits against the 2025 calculation instructions. No material arithmetic discrepancy found. This is a synthetic historical exercise, not a certified tax filing. The human reference processed successfully against all 67 rubric criteria; a separate native solver produced a 35-page, 19,347,709-byte PDF and was rejected by the inherited 10 MiB per-file collection limit before grading. That is an execution constraint, not evidence of a factual error in the task.

Code tasks

cyber_02 — certificate validation. The test scenarios match the task's stated path, policy, name-constraint and revocation requirements. The original reference passed the existing evaluator, but independent inspection found that its cache key omitted initial_policy_mapping_inhibit and initial_any_policy_inhibit. After a permissive validation, a stricter validation on the same engine incorrectly returned the cached success; a fresh engine correctly rejected it.

The Samples reference now includes both flags in cache identity and again receives reward 1.0 through the actual Docker submission/evaluator. Two new reference-only regression tests cover both directions of each policy change; both fail on the original reference and pass on the corrected one. The original is preserved as upstream-cyber02-gold.diff. The task prompt, starter repository, hidden tests and scoring protocol are unchanged. This fixture uses deterministic crypto helpers and task-specific validation rules; it is not a production certificate validator or a comprehensive RFC-conformance test.

qna014 — signed-package incident audit. Applied the nine proposed safe patches in order to an isolated copy and executed the archived primitives (9/9 passed). Replayed the captures and independently derived the sequence-906 head, all capture decisions, eight runtime packages, four build packages, recipe digest and built-wheel digest. They match the published reference. Token 71 conflicts on the old request and is fenced on the current request; token 81 is fenced, and token 82 resumes. These outcomes agree with the reference and executable retry rules. The existing native judge/file-property validation also passed. This checks the main reference conclusions, not every possible incident variant.

Run the additional cyber regression locally with python tests/check_cyber_reference.py, or inside the built controller with:

docker run --rm --entrypoint python poindexter-samples tests/check_cyber_reference.py