EPS
← All tasks·#2045Completed

daily-fix: projected-vs-realized + fallback-slice claims

kind: infra#wf-fix#wf-fix-fp:38d25ba714a8#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 /daily 2026-08-02 (route 2: behavior/logic change -> independent review) from the nightly problem sweep (miners 2/3, sessions 0ac15c23/55419495, tasks #1773/#1739).

Goal

Claims about what ran, at what size, and what is banked are grounded in completion artifacts and relocation sweeps at compose time.

Workflow gap

  • Bug observed: Four verify-before-asserting failures in one day that existing clauses do not name: (a) 'we never ran 3.2M batch calls' asserted post-/compact about a real completed 9,505wave−−composedofftheapproval/projectionmarker;refutedonlyafterThomas′s′areyousure−digdeeper(withasubagent)′(session0ac15c23,14:02−14:42Z).(b)AhallucinationNQ−Openheadline′+0.220′wasreadfromaFALLBACKslice(U=250,L=2,500)becauseoperating−slicecellsdidn′texistyet;realoperating−slicedeltas+0.018/−0.003(session55419495,corrected08:34Z).(c)ThreeconsecutivewrongclaimsaboutSAElabelcoveragefromasinglelocal−dircount;thefull127,605/128,512inventorysurfacedonlyafter3correctionsin 4min(0ac15c23,03:04−03:11Z).(d)A 9,505 wave -- composed off the approval/projection marker; refuted only after Thomas's 'are you sure - dig deeper (with a subagent)' (session 0ac15c23, 14:02-14:42Z). (b) A hallucination NQ-Open headline '+0.220' was read from a FALLBACK slice (U=250,L=2,500) because operating-slice cells didn't exist yet; real operating-slice deltas +0.018/-0.003 (session 55419495, corrected 08:34Z). (c) Three consecutive wrong claims about SAE label coverage from a single local-dir count; the full 127,605/128,512 inventory surfaced only after 3 corrections in ~4 min (0ac15c23, 03:04-03:11Z). (d) A ~1.5-3k re-judge was priced to the user on a false 'per-draw scores are averaged away' absence claim; the relocation sweep that refuted it ran only after dispatch (55419495, 14:27-15:35Z).
  • Why it is a workflow gap: The existing arms cover numbers, absence claims, and coverage extremum sweeps, but not the projection-vs-realization ambiguity (approval markers carry the same figures as completion records) nor fallback-slice labeling, and the absence-sweep is not sequenced BEFORE spend questions.
  • Confidence (emitter): high (all four incidents probed by miners with verbatim user rows + refuting artifacts)
  • verified-at-filing: grep -c -iE 'projected.vs.realized|completion artifact' CLAUDE.md -> 0; grep -c -iE 'fallback slice|operating slice' CLAUDE.md -> 0; the relocation-sweep clause exists but has no before-pricing sequencing (read at compose time).

Proposed change (refine in planning)

extend the verify-before-asserting arm: (i) PROJECTED-VS-REALIZED -- any claim about whether/at-what-size a batch wave or run executed is read from the run's COMPLETION artifact (drop report, batch-dispatch checkpoint, spend record), never the approval/projection marker (the same figure legitimately appears as both projection and fact); (ii) a headline number measured on a FALLBACK/partial slice names the slice and the 'fallback' label in its setup line; (iii) coverage/width claims about banked artifact SETS (labels, judgments, per-draw scores) require the relocation sweep BEFORE any regeneration is priced or an AskUserQuestion spends the user's attention on it.

Scope / surfaces

  • Primary target: CLAUDE.md

Constraints / invariants

  • Workflow-surface rules apply; scripts/workflow_lint.py --check-asks passes; ruff on touched files passes.
  • This session runs under EPM_WORKFLOW_FIX_SESSION=1 (Provenance workflow_fix_target: line) -- it MUST NOT auto-route its own subagents' workflow-fix candidates (recursion guard).

Provenance

  • fingerprint: 38d25ba714a8

  • workflow_fix_target: CLAUDE.md

Activity