mirror of
https://github.com/garrytan/gstack.git
synced 2026-11-16 05:17:24 +01:00
* Make /eval-plan check judges, zeros, failure counts and decisions before results count (GST-2754) Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com> * Check zeros in every results table, apply Decisions and Scoring with stack: none, count only valid judge labels (GST-2754) Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com> * Describe the /eval-plan result checks in the skills guide (GST-2754) Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com> --------- Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
55 lines
4.3 KiB
TypeScript
55 lines
4.3 KiB
TypeScript
/**
|
|
* Result codes for the /eval-plan preregistration (bin/gstack-eval-plan prereg
|
|
* --check, the priced pilot's arm list). Kept beside lib/result-codes.ts, which
|
|
* spreads this table into RESULT_CODES; the same rules apply: one stable
|
|
* docs/troubleshooting.md anchor, a summary and a fix clause each. The
|
|
* no-gstack arm rule is documented in docs/TESTING_INTERNALS.md#no-gstack-arm.
|
|
*/
|
|
import type { ResultCode } from './result-codes';
|
|
|
|
export const EVAL_RESULT_CODES = {
|
|
PREREG_INCOMPLETE: {
|
|
anchor: 'prereg-incomplete',
|
|
summary: 'the preregistration file lacks a required section, a section is empty, or the Comparison, Decisions, Scoring or Diagnostics declaration is malformed',
|
|
fix: 'fill every section `gstack-eval-plan prereg --check` names (bars, decisions, arms, comparison, scoring, decoys, stop rules, held-out exposure) before any priced call: at least two decisions with different actions, and a judge from another model family checked against at least 40 labels at kappa 0.6 or better',
|
|
},
|
|
// Field list 2 (2026-10-16), W8: preregistered diagnostics run before the freeze commit.
|
|
PREREG_DIAGNOSTIC_UNTESTED: {
|
|
anchor: 'prereg-diagnostic-untested',
|
|
summary: 'the declared summarizer, run on the declared synthetic fixture, did not report every preregistered diagnostic at its expected count',
|
|
fix: 'fix the summarizer or the declared definition (expected count, fixture) until every PREREG_DIAGNOSTIC line says ok, then commit the freeze; never freeze a diagnostic the summarizer has not counted',
|
|
},
|
|
// Garry, 2026-10-31: "all of our head to head tests have to test no gstack from here".
|
|
PREREG_NO_GSTACK_ARM: {
|
|
anchor: 'prereg-no-gstack-arm',
|
|
summary: 'the Comparison says stack: loaded and the No-gstack arm section is missing, lacks a key (masking:, install: and audit: included), says none for a masking key, or does not give every model its own <model>#no-gstack Arms line',
|
|
fix: 'add `## No-gstack arm` with arms:, inputs:, prompt:, panel:, report:, masking:, install: and audit:, and list each <model>#no-gstack arm under Arms; when no arm loads a skill stack, say stack: none with a why: line',
|
|
},
|
|
EVAL_REPORT_NO_GSTACK_ROW: {
|
|
anchor: 'eval-report-no-gstack-row',
|
|
summary: 'the results file has no single table with a filled row for every arm the preregistration lists, so a no-gstack row is missing or not beside its variants',
|
|
fix: 'put one row per arm identity (each <model>#no-gstack beside its variants) in one markdown table of the results file, then rerun `gstack-eval-plan results --check`; publish no number before it says ok',
|
|
},
|
|
// The result checks (docs/TESTING_INTERNALS.md#result-checks).
|
|
EVAL_JUDGE_UNCHECKED: {
|
|
anchor: 'eval-judge-unchecked',
|
|
summary: 'the preregistration names a model judge, and its hand labels are missing, fewer than the bar, all one class, or agree with the judge below the kappa bar',
|
|
fix: 'label at least the preregistered number of outputs by hand (never held-out rows), write them as {"id", "human", "judge"} JSONL, and pass them with --labels; a judge below its bar is replaced or its numbers are not published',
|
|
},
|
|
EVAL_REPORT_UNPLANTED_ZERO: {
|
|
anchor: 'eval-report-unplanted-zero',
|
|
summary: 'a column of a results table reports a zero, and the preregistration never showed the counter behind it finding a planted case',
|
|
fix: 'declare the counter under ## Diagnostics with a synthetic fixture that plants at least one case (expected 1 or more), run prereg --check --allow-repo-commands, and name the column after the diagnostic',
|
|
},
|
|
EVAL_REPORT_NO_ANSWER_ONLY: {
|
|
anchor: 'eval-report-no-answer-only',
|
|
summary: 'a results table counts failures without an answer-only column beside it, so a rise in failures cannot be compared with the grader\'s own reading',
|
|
fix: 'add a column headed answer-only to the same table, counting failures by the grader\'s rule on the final answer alone, and report both readings',
|
|
},
|
|
EVAL_ARM_DUPLICATE: {
|
|
anchor: 'eval-arm-duplicate',
|
|
summary: 'the pilot was given the same arm identity twice, so two arms would share one set of attempts',
|
|
fix: 'give each arm its own identity, <model> or <model>#<loadout> (for example claude-opus-5-5#no-gstack beside claude-opus-5-5); nothing was reserved',
|
|
},
|
|
} as const satisfies Record<string, ResultCode>;
|