EPS Dashboard

Overview / Motivation

Auto-filed by the /daily 2026-07-28 problem sweep (transcript miners over the day's 55 session transcripts). Sources: group-E P1/P2, group-G P2, group-J P1 (4 independent miners).

Goal

Extend the refusal-prevention surface with three measured 2026-07-28 gaps: Goal-frontmatter neutralization, a demonstrated-content-trigger fast path, and steering-surface read discipline.

Workflow gap

Bug observed: Refusal storms cost ~3h of wall across 4 sessions on 2026-07-28. #1769: 5 dead planner-phase spawns (3 usage-policy, 1 thrash, 1 degenerate) then an orchestrator refusal-wedge at 22:51Z — the task's goal: frontmatter itself carries intervention-control phrasing and rode verbatim into every brief. #1774: 3 implementer refusal kills + 1 sonnet thrash; round A ran clean at 100 calls while round B (the steering-module content) was killed at 14 calls — the content trigger was demonstrated, yet the next spawn went to the default model anyway. #1776: 2 kills at the final report turn. #1769's owner session was also killed while grepping steering-script hyperparameters (paging intervention-control vocabulary). (Kill counts are transcript-mined tallies — unverified hypothesis — verify at plan time against the sessions' events/markers; tasks #1769/#1774 both confirmed recovered by successor sessions.)
Why it is a workflow gap: the existing ladder (CLAUDE.md § Spurious usage-policy refusals) covers brief VOCABULARY (rung e) and post-kill recovery (b/b2), but not the Goal-frontmatter channel, not a skip-ahead once content (not model pathway) is the demonstrated trigger, and the trigger-dense-review recognition heuristic omits steering surfaces.
Confidence (emitter): medium (mechanisms inferred from kill patterns, not A/B-probed)
verified-at-filing: grep -c 'FIRST-PASS briefs carry BOTH disciplines' CLAUDE.md → 1; rung (b2) present (grep 2 hits); grep -n 'intervention-control phrasing' CLAUDE.md → clause (e) names brief phrasing only, no goal-channel clause; trigger-dense-review.md recognition heuristic has no steering-surface bullet (read 2026-07-29 UTC). NOTE: the first-pass real-corpus brief discipline half of this incident class ALREADY LANDED as #1748 (§ Real-corpus datagen briefs) — deliberately excluded from this filing.

Proposed change (candidate diff sketch — refine in planning)

See the three amendments in the title fields; each is a short additive clause to an existing section — no ladder restructuring.

Scope / surfaces

Primary targets: CLAUDE.md (§ Spurious usage-policy refusals, clauses (e)/(b2)), .claude/rules/trigger-dense-review.md (recognition heuristic + § Orchestrator ordinary turns)

Constraints / invariants

Workflow-surface only — never experiment code, configs/, or tasks/.
scripts/workflow_lint.py --check-asks passes; ruff on touched files passes.
The spawned session runs under a workflow_fix_target: Provenance line — recursion guard applies (it parks, never auto-routes, its own subagents' workflow-fix candidates).

Provenance

fingerprint: cb0051488fb6
workflow_fix_target: CLAUDE.md