Production code had no golden tasks, scorer, or rubric
(harness2.py states this explicitly); routing shadow gives
cost/status only and admits counterfactual not_measured,
so 'is the suggested model actually better' had no
mechanism. This track adds a pure module + CLI that scores
already-completed task records against human-curated golden
tasks and compares two runs, without running tasks, spending
model cost, touching the DB, or activating anything.
Golden tasks: JSON {id, prompt, tier, checks}; only four
deterministic check types (contains, not_contains,
status_completed, max_tokens, max_cost_usd, the last in
cost.py microdollar units); invalid definitions fail loudly.
Scoring: per-task score = passed / measured checks;
cost checks with missing/unpriced data are unknown, excluded
from the denominator and reported separately (never counted
as pass or fail); empty result_text is a string-check fail,
not unknown. all_pass means 'all measured checks pass' and
is reported alongside fully_measured_n so a pass rate is
never misread as 'all checks passed'. Comparison: per-run
stats, paired delta over golden tasks present in BOTH runs
(unpaired listed separately, missing listed, unmatched
records excluded with a warning), duplicate task_id rejected.
Paired n < 20 (same threshold as shadow) => ready_for_
decision false, no significance or p-values computed.
Deliberately NOT included: a task runner (golden tasks
still need a separate run-and-record step or manual task
creation), an LLM judge or quality score, and any routing
activation suggestion. A contains check can be gamed by
echoing, so golden sets must be human-curated with checks
that require actually doing the work; the example file in
eval/ is format-only, not a decision-ready set. Tests:
test_eval_loop.py (8 tests).
