File size: 4,277 Bytes
573b28e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 | # Grading and stability statistics
## Task grader
Each Mode B task has one generated answer. The grader normalizes the submission before comparing:
- One surrounding code fence is stripped.
- The last balanced `{...}` JSON object is used when one is present; keys named `answer`, `result`, `final`, `value`, or `output` (or a single-key object) supply the answer.
- Text is normalized for case, Unicode form (NFKC), whitespace, quotes, and trailing punctuation.
- Numeric keys compare the first number token in the answer; `$`, `,`, `%`, and currency symbols are ignored, so `42`, `$42.00`, and `42 dollars` agree.
- Ordering tasks carry `metadata.options`. When the answer carries an order cue (`first`, `ahead`, `before`, `won`, `beats`, `leads`, `fastest`, `winner`), the check takes the option named nearest the cue on the winning side, so `Ari finished first, ahead of Bo` and `The winner is Ari` both score; without a cue the last option name mentioned wins, so `Bo finished behind Ari` also scores.
- Empty or non-string answers score 0.
Ways this grader can fail:
- A leaked public answer key can turn recall into an apparent reasoning success. The hidden split (`hid-` prefix, unpublished seed) exists for that reason; keep it outside version control. Hidden task metadata carries no seed, so a served task cannot reveal how to regenerate the split, and the generator redraws until no hidden question matches a public one. The hidden split stays hidden only while its seed is private and its prompts are not served publicly: the generator accepts seeds up to 2^53-1, but anyone who learns the seed can rebuild the file, and anyone served the prompts already has the tasks.
- Prose that mentions the expected value for the wrong reason still scores. The last-mention rule also credits a guess that happens to land on the key.
- Numbers embedded in a longer explanation are read as the answer; the first number token is treated as final, so a worked example placed before the answer can be misread.
- Generator mistakes become authoritative until a human reviews the task diff.
- The six synthetic families cover only a narrow slice of reasoning.
- MCP `score_answer.submission` is an open object schema in benchmark-core, so OpenAI strict tool mode may reject it; the shape is documented upstream, not here.
The E2E gate submits empty text, constant answers, a refusal, a repeated tool call, and a first-name heuristic to every task. Each policy must remain below 0.10 mean score. Two further policies answer with the first and the last number printed in each prompt; both must stay below 0.10 in every family, not only overall. A direct answer-key policy must score 1.00, and the ordering phrasings above must each score 1.00.
## Ranking analysis
`score-stability` groups input RunReport rows by model and aligns them by task ID. It computes model means, pairwise Kendall tau-b with tie correction, first-place probabilities, 95% rank intervals, and a `warnings` list. Task IDs ending in `-vN` are paraphrase variants of one item; each bootstrap resample draws items once and applies the same draw to every model, then picks one run per model, so comparisons stay paired and correlated variants are not double counted. A fixed seed makes the result reproducible.
When a task set has at least two variant indices the result adds `promptSensitivity`: a per-variant ranking table plus each model's best and worst rank across variants (charter Mode B).
Ways the statistics can mislead:
- Small task or run samples produce wide, unstable intervals. Fewer than 30 shared tasks, fewer than 30 shared items (the bootstrap resamples item clusters, not rows), one run per model, or under 500 resamples each add a `warnings` entry.
- Missing task IDs reduce the comparison to the intersection.
- Ties make rank order less informative even with tau-b correction. Identical models report tau 1 and split `probabilityFirst`.
- Reports from different benchmark versions may not be comparable.
- Bootstrap resampling describes the supplied sample, not all future prompts.
The command rejects malformed reports, duplicate task IDs within one report, fewer than two models, no shared tasks, and non-finite scores. Treat the SVG as a diagnostic, not a universal leaderboard.
|