B1: gateway capability check now triggers on walking the sub2api
gateway, not on the gateway having filled a model; task-level
explicit models are audited (worker_capability_gateway_explicit)
but never blocked, preserving explicit priority; gateway-default
models still fail-fast under active.
I1: dry-run adds a gateway-model section, sourced only from
CLOUD_SUB2API_DEFAULT_MODEL (never queries the relay); an active
tier whose requires the gateway model fails or cannot verify
blocks, shadow only warns.
I2: shadow review gate now counts completed samples
(completed >= 20), same basis for sample_shortfall; the endpoint
attaches per-tier capability status (read-only policy load) and
flags capability-failing tiers in the recommendation.
I3: configured-but-broken routing policy logs a warning in
load_routing_policy (unconfigured stays silent); the worker sets
routing_policy_load_failed and audits
worker_routing_policy_load_failed per task.
I4: resume preamble wraps prior output/error as data with an
explicit do-not-execute constraint; position verified in run_one
(tree -> preamble -> project -> original) and the docs corrected
from the false 'closest to the original prompt' claim.
I5: decided model is written back with a CAS UPDATE
(model IS NULL OR model=''), so later attempts reuse it;
explicit models are never overwritten.
I6: agent_* columns added to worker._ensure_result_columns and
db.init_postgres_schema backfill; migrations backfills them for
tasks before insert (sender_task_id pattern); column-level tests
for agent_tool_approvals and task_attempts.
I7: coordinator requeue adds AND status=<snapshot> (rowcount 0
skips requeue/audit) and filters paused_by_user / user_cancel /
user_pause children, which are never revived or cleared.
I8: explicit-model tasks (including coordinator children that
inherited a parent model) audit a source=explicit worker_routing
row; evaluate_shadow counts it as per-tier explicit_n only,
never routed_n, never a dedup slot.
M1/M2/M3/M5/M6/M8/M10: dedup slot moved after tier validation,
shadow sort doc fixed, truncated recommendation notes truncation,
seed cleanup on gateway/quota/goal-cap early returns, tools-only
coordinator definitions rejected, preamble goal branch marked
effectively unreachable, DESIGN.md drift corrected (outside repo).
M4/M7/M9: documentation/new-field only (pass_rate_fully_measured,
whitespace-model shadow-drop note, bound-note race note); old
field semantics unchanged.
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).
host