EPS
← All tasks·#1974Completed

daily-fix: pooling-convention row in mapping disclosure

kind: infra#wf-fix#wf-fix-fp:c9f88143e1b6#daily-auto-filed
track:

Highlight text on any card (body, plan, or an event) to anchor a comment, or leave a whole-task comment here. Mention @claude to summon a reply.

No comments yet.

Overview / Motivation

Auto-filed by the /daily 2026-07-31 problem sweep (CONSOLIDATED M10; miner-5:P2). Source: session 75f66748 (#1768; binding convention change also posted to #1947) — #1768 shipped span-mean context pooling as an unexamined inherited convention, mismatching the #779 last-token line it compares against. The user caught it ("why are we fitting span mean …???"); the assistant conceded "it wasn't a deliberate measurement decision — it's an inherited convention… The plan noticed the mismatch… but only in one place: the base-fit sanity gate." Cost: a ~15–18 GPU-h re-pool round. Plan critics did not flag the pooling mismatch against the cited comparison line.

Goal

Add a pooling-convention row to the both-arms representation-mapping disclosure: every mapping plan names the pooling of every context/answer vector (span-mean vs last-token vs response-avg) AND its match to the cited baseline line's convention; a mismatch is a plan-time REVISE, not a sanity-gate footnote.

Workflow gap

  • Bug observed: A representation-mapping plan inherited span-mean pooling while its headline comparison target (#779) used last-token pooling; no plan-time lens required the pooling convention to be named or matched, so the mismatch survived the full critic ensemble and cost a re-pool round after user catch.
  • Why it is a workflow gap: The mapping disclosure surfaces (planner §4 Design both-arms rule; Statistics lens item 15 "Mapping-baselines pair" in .claude/rules/critic-lens-reference.md line ~984) require the prefix/context arms and the identity+kNN baseline reads, but say NOTHING about the pooling convention of the vectors being mapped or its parity with the cited baseline line — pooling is a load-bearing measurement choice that silently rides in from reused capture code.
  • Confidence (emitter): high (absence probed; incident is the session's own concession)
  • verified-at-filing: grep -n "pooling\|span-mean\|span mean" .claude/agents/planner.md .claude/agents/statistics-critic.md .claude/rules/critic-lens-reference.md .claude/rules/planner-section-reference.md → 0 hits in all four files (absence claim — no pooling-convention row exists anywhere in the mapping-disclosure surface; the 0-hit result is the evidence). Statistics lens item 15 confirmed at .claude/rules/critic-lens-reference.md:984 (identity+bias / kNN only; context read — no pooling clause). git log --oneline --since='7 days ago' -- .claude/agents/planner.md → 5 commits (smoke-slice probes, c39 prose sync, off_pod_phases, §10 durability, §3 relocation) — none pooling-related; no landed fix (2026-07-31).

Proposed change (candidate diff sketch — refine in planning)

.claude/agents/planner.md §4 Design (the both-arms mapping clause):
+ Pooling-convention row: for every representation mapping, name the pooling
+ of EVERY vector entering the map (context/prefix/answer: span-mean |
+ last-token | response-avg | other) AND state whether it MATCHES the pooling
+ of the cited comparison/baseline line (e.g. the #779 last-token line);
+ a deliberate mismatch carries a one-line justification.

.claude/rules/critic-lens-reference.md, Statistics & Measurement lens
(extend item 15 or add a sibling item):
+ REVISE a mapping plan that does not name its pooling convention per vector,
+ or whose pooling mismatches the cited baseline line without a stated
+ justification — a mismatch is a REVISE, never a sanity-gate footnote
+ (#1768: span-mean inherited vs #779 last-token; ~15-18 GPU-h re-pool).

.claude/agents/statistics-critic.md (v2 twin): mirror the same item.

Scope / surfaces

  • Primary target: .claude/agents/planner.md (§4 Design)
  • Secondary: .claude/rules/critic-lens-reference.md (Statistics lens), .claude/agents/statistics-critic.md (v2 owner), .claude/rules/experiment-guidelines.md guideline 11 (one sentence), .claude/agents/codex-statistics-critic.md if it inlines the lens text
  • Grep before editing: grep -rn 'item 15\|Mapping-baselines' .claude/agents/ .claude/rules/ and update every surface that mirrors the lens; list them in the plan.

Constraints / invariants

  • Workflow-surface only — never experiment code, configs/, or tasks/.
  • Keep the v1 lens numbering stable (extend item 15 or append; do not renumber).
  • scripts/workflow_lint.py --check-lens-coverage (if the lens ledger is touched) + --check-asks pass.
  • This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — it MUST NOT auto-route any of its own subagents' workflow-fix candidates.

Provenance

  • fingerprint: c9f88143e1b6

  • workflow_fix_target: .claude/agents/planner.md

  • fingerprint: (driver-computed; tag authoritative)

Origin: CONSOLIDATED M10 (miner-5:P2), /daily 2026-07-31 — "#1768 shipped span-mean context pooling as an unexamined inherited convention, mismatching the #779 last-token line it compares against" (session 75f66748).

Activity