Skip to content

Latest commit

 

History

History

README.md

evals/results/ — the scorecard store

This directory is written by eidolons eval swe --matrix <arms.json> and is committed — it is the persistent, diffable measurement log the H-WIN campaign (and any future eval matrix) reads from.

What lands here

For every arm in an arms.json (schema: ../../schemas/eval-arms.schema.json), the matrix runner writes one scorecard:

evals/results/<UTC-date>-<suite>-<label>.scorecard.json
  • <UTC-date>date -u '+%Y-%m-%d' at run start.
  • <suite> — the suite file's basename without .yaml (e.g. swe-suite, kupo-keep-suite).
  • <label> — the arm's label from arms.json.

After every arm has run, one matrix summary is written alongside them:

evals/results/<UTC-date>-<suite>-matrix.json

It compares every non-control arm against the first control: true arm: resolved-rate delta, pass^k delta, and a per-task flip table (newly-resolved / regressed). Scorecards conform to ../../schemas/eval-scorecard.schema.json.

Re-running the same (date, suite, label) on the same UTC day overwrites the prior file for that day — this store is a daily log, not a run-by-run archive. Pass --no-store to eidolons eval swe --matrix to skip writing entirely (useful for local iteration).

Reading the store

eidolons eval baseline <suite> [--label <l>] [--against <file>] diffs the two most recent scorecards for a (suite, label) pair (or the latest vs an explicit --against file) and exits non-zero (5) on regression. See eidolons eval baseline --help.

⚠️ Smoke scorecards are NOT capability claims

A scorecard with harness.smoke: true was produced by --matrix --smoke: every arm's fix was the suite's own gold_fix reference patch, not a model. A 100% resolved_rate there proves the ORCHESTRATION works end-to-end (arm dispatch → suite run → scorecard shape), never that a model solved an unseen task. This is exactly the same honest-scope distinction evals/swe-suite.yaml and evals/kupo-keep-suite.yaml already document for the underlying eidolons eval swe harness — it just also applies per-arm here. Only harness.smoke: false scorecards (a real --fix-hook, e.g. one of evals/hooks/keep-bare.sh / evals/hooks/keep-system.sh) are evidence for a headline claim like H-WIN.

Compliance scorecards land here too

eidolons eval compliance --json output may be committed alongside the matrix scorecards as <UTC-date>-compliance-<qualifier>.scorecard.json (a different instrument, a different schema — see cli/src/eval_compliance.sh's emitted shape). Same store rationale: measurements are diffable history, not ephemera.