feat(compliance): prove metric container subsumption end to end - #6887
feat(compliance): prove metric container subsumption end to end#6887bokelley wants to merge 3 commits into
Conversation
available-metric.json's container-subsumption rule was prose-only — nothing pinned independent implementations to the same evaluation, the divergence hazard tracked in #6785. Two halves close it: The reference seller now implements the rule through one shared helper (metric-subsumption.ts): required_metrics filtering on get_products existed for vendor metrics only — the standard-metric filter was silently ignored — and now evaluates container-subsumes-leaf semantics; get_media_buy_delivery implements requested_metrics narrowing (leaf resolves to its carrier, impressions and spend always included) across totals, package rows, and creative rows, with viewability mirrored to package grain. The metric_container_subsumption storyboard grades all four exercises from #6785 with validations anchored to the normative schema text: container satisfies leaf requirement, no empty mixed intersection, leaf request returns the full carrier object, and a leaf declaration does not satisfy a sibling leaf (filter-not-fail). Closes #6785. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The #6815 coverage rebaseline set current-source floors above what main actually measures (creative 48/50 clean with 200 steps vs 49/209 floor; creative-builder 48 vs 50, 172 vs 184; sales 553 steps vs 556) — every gap is authored known-failing skips citing packaged-runner blockers, with zero step failures. The workflow has been red on main since the merge and the mirrored local pre-push gate blocks every compliance- source push. Set floors to the values measured on main's own CI run so the ratchet grades regressions from reality again. Closes #6876. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
![aao-secretariat[bot]](https://avatars.githubusercontent.com/u/221690821?s=60&v=4){kind=small}
| // Sellers may report per-package viewability; the training agent's | ||
| // simulate_delivery viewability param is buy-scoped (mirrors totals), | ||
| // so it applies to every package uniformly. | ||
| ...(simDelivery?.viewability ? { viewability: simDelivery.viewability } : {}), |
There was a problem hiding this comment.
(non-blocking) Duplicate spread. packageDeliveryMetrics already spreads ...(simDelivery?.viewability ? { viewability: simDelivery.viewability } : {}) three lines up (L12958), before ...byCreative. This second spread of the identical key/value is a no-op — nothing between them mutates viewability. The comment describes real intent (buy-scoped viewability applied uniformly per package), but the first spread already achieves it. Drop the added line and, if the rationale is worth keeping, move the comment to L12958.
aao-secretariat
Bot
added
the
ladon/needs-human-review
There was a problem hiding this comment.
Ladon verdict: Escalate to human review
Escalate — gated path requires human/CODEOWNERS review.
This PR modifies .github/workflows/training-agent-storyboards.yml, which matches the deterministic gated-path glob .github/workflows/**. gated_paths is true and review_decision is REVIEW_REQUIRED (not APPROVED), so decision-table row 2 fires: the hard approval gate is not satisfied and the outcome must be escalate.
There are no blocking (critical/high) findings — the wire-shape logic is sound, the metric-narrowing is schema-safe, and changesets are present and correctly scoped. The only reviewer finding is low (a no-op duplicate viewability spread in task-handlers.ts:12964), which does not block.
Escalation reasons
.github/workflows/training-agent-storyboards.yml(modified) is under a hard approval gate (.github/workflows/**); requires human/CODEOWNERS review (review_decisionis currentlyREVIEW_REQUIRED, notAPPROVED).
Once a human/CODEOWNERS approval lands (review_decision: APPROVED), this PR falls through to a clean approve on its findings.
Why human review
- .github/workflows/training-agent-storyboards.yml (modified) is under the hard approval gate
.github/workflows/**; review_decision is REVIEW_REQUIRED, not APPROVED — human/CODEOWNERS review required. - This PR touches a path under a hard, non-overridable approval gate (.github/workflows/training-agent-storyboards.yml (modified) matches
.github/workflows/**) and the current GitHub review decision is 'REVIEW_REQUIRED', not APPROVED. This is a hard gate enforced in code — Ladon cannot auto-approve until a human/CODEOWNERS approval is recorded, regardless of how clean the diff is.
1 participant
Summary
Implements #6785 (deferred at triage as blocked on #6618; unblocked since it merged). The container-subsumption rule in
available-metric.jsonwas prose-only — nothing pinned independent implementations to the same evaluation. Two halves close it:Reference seller (
server/src/training-agent/): one shared helper (metric-subsumption.ts, 14 unit tests) now backs both surfaces. Notably,required_metricsfiltering onget_productsdid not exist at all for standard metrics (onlyrequired_vendor_metricswas implemented — the documented filter was silently ignored); it now evaluates container-subsumes-leaf semantics.get_media_buy_deliveryimplementsrequested_metricsnarrowing (leaf resolves to its carrier, impressions/spend always included) across totals, package rows, and creative rows, with viewability mirrored to package grain.Conformance storyboard (
metric_container_subsumption, 10 graded validations): the four exercises from #6785, each anchored to normative schema text — container satisfies leaf requirement; mixed container/leaf declarations don't produce empty intersections; a leaf request returns the full carrier object (viewability.viewable_rateAND.measurable_impressionspresent,clicksabsent, impressions/spend always present); and a leaf declaration does not satisfy a sibling leaf (filter-not-fail exclusion).Verification
Includes the floor re-baseline commit from #6885 (merge-order safe: whichever lands first, the other dedupes).
Closes #6785.
🤖 Generated with Claude Code