EPS
← All tasks·#1429Completed

workflow-fix: gotchas entry — rank-space bootstrap tail-mass gating (float-space CI coverage epsilon-fragile)

kind: infra#wf-fix#wf-fix-fp:92554ba1f495
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 workflow-fix-on-bug protocol from a gotcha_candidate: yes failure-lesson raised on task #825 (emitting agent: experiment-implementer; orchestrator routed per /issue Step 7 failure-lesson capture action 3).

Goal

Add a gotcha entry to .claude/rules/gotchas.md documenting the epsilon-fragility of float-space bootstrap-CI gating and the rank-space tail-mass replacement.

Workflow gap

  • Bug observed: a mechanized parity gate using strict float-space CI coverage (lo < point < hi) mis-classified degenerate bootstrap cells as gating because interpolated percentiles land 1e-8..1.6e-7 above the point estimate, causing deterministic pipeline aborts (#825 r11 rev 3, code-review v26 FAIL at t=24/25/29; n_gating=25, n_fail=3).
  • Why it is a workflow gap: .claude/rules/gotchas.md is the codebase-wide catalog of numerical/statistical traps; this class (float-space CI-coverage tests, and any

    =alpha/2 tail test against a point computed by DIFFERENT arithmetic — the identity-resample tie cluster, mass ~n!/n^n, float-jitters at ~1e-7) recurs in any bootstrap-gated pipeline and is documented nowhere in the rules surface.

  • Confidence (emitter): high (root_cause_confirmed: yes; fix validated by replay on the real archived table + 12/12 unit tests; code-review v27 PASS)
  • verified-at-filing: grep -n "CI coverage\|tail.mass\|rank.space\|percentile" .claude/rules/gotchas.md → 0 hits (2026-07-16); grep -c bootstrap .claude/rules/gotchas.md → 3 (none about CI-coverage gating — eyeballed); git log --oneline --since='7 days ago' -- .claude/rules/gotchas.md → 5 commits, none matching (EXDEV, hf-xet, artifact-reuse, bf16, pid-namespace).

Proposed change (candidate diff sketch — refine in planning)

+ **Float-space bootstrap-CI gating is epsilon-fragile — gate in rank space.** A strict
+ coverage test (`lo < point < hi`) mis-gates collapsed small-n bootstrap cells because
+ interpolated percentiles land 1e-8..1.6e-7 ABOVE the point; any >=alpha/2 tail test
+ against a point computed by DIFFERENT arithmetic inherits the fragility (identity-
+ resample tie cluster, mass ~n!/n^n, jitters ~1e-7 vs a float64/centered point). Gate:
+ strictly MORE than alpha/2 of finite draws strictly on EACH side of the identity-
+ resample anchor computed through the SAME vectorized expression as the draws (bitwise
+ tie-exact); alpha = the CI's own level. Annotate per-node tail fractions; fail loud on
+ nodes lacking them. Worked example: scripts/issue825_turndyn_fit.py _gc_verdict
+ (#825 r11 rev 4, commits 5f1ffd1dea+9ed2930464; agent-memory
+ feedback_rank_space_bootstrap_tail_gating.md).

Scope / surfaces

  • Primary target: .claude/rules/gotchas.md
  • Grep the workflow surface for the pattern before editing (grep -rln 'lo < point < hi\|CI coverage' .claude/ CLAUDE.md scripts/) and update every hit; list them in the plan. (The superseded agent-memory entry feedback_small_cell_bootstrap_ci_degeneracy.md is already annotated — do not re-edit.)

Constraints / invariants

  • Workflow-surface only — never experiment code, configs/, or tasks/.
  • scripts/workflow_lint.py --check-asks passes; ruff on touched files passes; if workflow.yaml or CLAUDE.md change, they stay consistent with the rule file.
  • 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 (recursion guard).

Provenance

  • workflow_fix_target: .claude/rules/gotchas.md
  • fingerprint: 92554ba1f495

failure_class: code phase: fit_gc (issue825_turndyn_fit.py G-C parity verdict) lesson: Strict float-space CI coverage (lo < point < hi) is itself epsilon-fragile: interpolated bootstrap percentiles land 1e-8..1.6e-7 ABOVE the point on collapsed small-n distributions (and any >= alpha/2 tail test against a point computed by DIFFERENT arithmetic inherits the same fragility — the identity-resample tie cluster float-jitters to a random side of a float64/centered point at ~1e-7). Gate on rank-space bootstrap tail mass instead: strictly more than alpha/2 of the finite draws strictly on each side of the identity-resample anchor computed through the SAME vectorized expression as the draws (bitwise tie-exact), alpha = the CI's own level. generalizes: yes owning_agent: experiment-implementer gotcha_candidate: yes root_cause_confirmed: yes supersedes: the 2026-07-16T15:54:13Z lesson (v10) on gating by strict CI coverage

Activity