EPS Dashboard

Overview / Motivation

Auto-filed by the /daily 2026-07-28 problem sweep (transcript miners over the day's 55 session transcripts). Source: group-A P3 (miner-probed).

Goal

Fail loud at item-construction time on Batch-API custom_id grammar violations instead of dying mid-wave on an anthropic 400.

Workflow gap

Bug observed: #1739's judge wave for the evil lane died on an anthropic HTTP 400: a custom_id carried a ~k segment, violating the API's ^[a-zA-Z0-9_-]{1,64}$ grammar. The crash happened inside a detached driver and was caught only by log monitoring (~10 min fix/relaunch; the fixed id shape f"{context_id}_k{k:02d}" landed on the issue worktree).
Why it is a workflow gap: judge_dispatch.py constructs and submits custom_ids with no grammar check — a malformed id is a deterministic 400 (an invalid_request_error, correctly NEITHER retried NOR dropped per llm-judging rule 24(iii)), so the right fix is pre-submit validation in the request builder.
Confidence (emitter): high (probed)
verified-at-filing: grep -n 'custom_id' src/explore_persona_space/eval/judge_dispatch.py → construction/threading sites at 29/36/219/429/499; grep -c 'a-zA-Z0-9_-' .../judge_dispatch.py → 0 (no grammar pattern anywhere in the module) (2026-07-29 UTC).

Proposed change (candidate diff sketch — refine in planning)

A small _validate_custom_id() at JudgeItem construction (and/or at dispatch_judge_items entry) raising ValueError naming the offending id; unit test with a tilde id.

Scope / surfaces

Primary target: src/explore_persona_space/eval/judge_dispatch.py

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: 584cb811c222
workflow_fix_target: src/explore_persona_space/eval/judge_dispatch.py