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

Validation — November 18, 2026

These are execution and packaging checks, not benchmark estimates. A model can receive partial rubric credit while the solver and grading pipeline work correctly.

Check Result
Build the combined Docker controller Passed on Linux ARM64 through Docker Desktop; both pinned Python environments pass pip check
Actual Docker CLI smoke tests 7 passed: exact seven-task selection, native plans for both providers, invalid selections/attempt counts, missing credentials, tampered input checksum rejection, undersized judge-panel rejection, archive integrity/path checks
Native OpenAI code adapter Real Responses API tool call and tool-result continuation passed
Native Anthropic code adapter Real native API tool call and tool-result continuation passed
Native Anthropic GDPVal adapter Real adaptive-thinking tool call and signed assistant/tool-result passback passed
cyber_02 reference patch Actual Docker agent submission and evaluator: reward 1.0
cyber_02 cache regression review Both policy-inhibition regressions fail on the original reference and pass on the corrected reference; corrected reference rerun through actual Docker CLI/evaluator: reward 1.0
cyber_02 unchanged repository Actual Docker submission and evaluator: reward 0.0
qna014 corrected reference answer Actual Docker submission and two native judges: reward 1.0; all 6 file-property checks passed
Completed Q&A attempt resume Repeated CLI run preserved all attempt records and needed no valid provider credentials
Q&A grading-only retry Actual CLI retry from a local GRADING_FAILED test state returned GRADED, reward 1.0; saved patch and solver logs/trajectory were unchanged
PDS049 native Anthropic solve Completed and graded through native OpenAI; seven-point rubric executed successfully
PDS136 counting arithmetic Independent exact Burnside enumeration matches D8=4864, SD16=37248, S3=502, S4=32923 and the final prime factorization
German GDPVal native OpenAI solve Completed: presentation and PDF saved, rubric grade returned successfully
Japanese GDPVal native OpenAI solve Completed and graded: Word document, score 100%
French GDPVal native OpenAI solve Produced a 19,347,709-byte PDF; rejected by the inherited 10 MiB per-file artifact limit before grading
French reference deliverable grading All 15 supplied input PDFs and the 17-page reference PDF processed; score 100%, 67 rubric criteria
Anonymous public download All 154 indexed file hashes matched; the downloaded Docker package built successfully and planned all seven tasks
Public artifact audit Python/JSON parsing, archive inspection, no known credential values, no production audit records or runtime secrets included

The Japanese live run has now completed. The French live solver produced a 35-page PDF but its 19,347,709-byte file exceeded the inherited 10 MiB per-file collection limit; this attempt is SOLVER_FAILED, not a graded incorrect answer. The reference grading result above uses the supplied human reference, not that newly generated answer. A completed native model solve is not claimed for PDS136 in this validation record.

The mathematical count check validates the finite-group arithmetic, not an independent proof of the block-Picard classification theorem. Live provider/tool tests do not establish that every account has access to the example model IDs. No AMD64 build or fresh E2B template build in an unrelated researcher account was performed during this validation; the documented template builder is the existing pinned Mini-Bench implementation.

The validation found and corrected the Q&A evaluator's missing YAML dependency, its handling of pytest collection failures, and omissions in the reference answer. See CHANGES.md for precise adaptations and the preserved original reference.

To repeat offline checks after building:

SAMPLES_IMAGE=poindexter-samples python tests/smoke.py
python tests/check_math.py
python tests/check_cyber_reference.py

The two tests/live_*.py scripts are optional paid API tests. Run them inside the controller using the corresponding /opt/venvs/code/bin/python or /opt/venvs/general/bin/python, with your own API environment and tests/ mounted at /test. They do not run as part of the offline smoke suite.