File size: 3,410 Bytes
149da17 | 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 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 | # Scoring
The current scoring protocol is finding-quality-discovery. Commit-pinned current
scores must not be compared directly with historical narrow-task scores.
## Local gates
Each finding needs a permitted population, valid agent-selected comparison,
minimum population/arm sizes, complete condition partitions, exact text offsets,
and all recomputed statistics. Invalid submissions do not receive invented
quality scores. See [submission contract](SUBMISSIONS.md).
## Evidence-based quality
A semantic assessment supplies support, task_fulfilled, duplicate_of, rationale,
and four dimensions from 0 to 1: S statistical validity, E evidence entailment,
D analytical depth and C calibration. Dimension anchors are 1 fully justified,
0.75 minor gaps, 0.5 material limitations, 0.25 weak and 0 absent/wrong.
```text
finding quality = support × (15 + 25S + 20E + 30D + 10C)
```
Support factors: supported=1, partial=0.5, unsupported=0. Uncertain stays null.
An unfulfilled task or duplicate gets 0 regardless of dimensions. Generic
sentiment, metadata frequencies or an unexamined aggregate contrast do not
fulfill the task. Depth requires a substantive discovery, a defensible choice
of scope/comparison, competing explanations, and correct interpretation of
robustness checks. Material audit omissions cap D at 0.5; merely restating
numbers caps it at 0.25. These semantic caps are judge instructions, not
deterministically proven properties.
Example: a supported, nonduplicate finding with S=.8, E=.9, D=.75, C=.8 earns
15+20+18+22.5+8 = 83.5. Partial support halves it to 41.75. This is an illustrative
calculation, not an observed agent result.
## Semantic audit
All selected-population arithmetic and partitions are checked. The model then
sees at most 160 documents by default: submitted quotations, samples from each
nonempty positive/negative/unknown assignment cell and a corpus-wide remainder.
The seed is created after submission and saved with judge configuration for
reproducibility; sampled IDs are retained in each review. This is a bounded
audit of submitted labels, not a new independently labeled or held-out dataset.
Sampled label mistakes undermine semantic support even if the counts add up.
Absence of sampled errors does not prove all labels correct. The judge must use
uncertain when the packet cannot resolve a claim. No full-corpus semantic
guarantee or unbiased estimator of label accuracy is claimed. Report judge
model, input budget and audit size; model-based scores have evaluator error.
## Aggregation and reference availability
Task quality averages all submitted finding scores; any unresolved finding makes
task quality unavailable. An abstention is valid but unscored, not a verified
absence of useful findings. A full quality_mean is available only when every
task has a score. conditional_quality_mean covers only scored tasks and must be
reported alongside scored_tasks, abstention_rate, valid_submission_rate, missing,
invalid and pending counts. Source and family breakdowns are included.
Current redesigned tasks have no fixed reference conclusions. reference_coverage
is null. Earlier 50 conclusions are retained only in history, not reused against
new questions. Novel supported findings are not penalized for lacking a fixed
match. Reviews bind task, corpus, submission, reference configuration and scoring
hashes; stale reviews must not be reused.
|