EPS Dashboard

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a workflow-fix candidate raised on task #1739 (emitting agent: the #1739 orchestrator session, 2026-08-05).

Task #1739's evil-ood-spread-round hit ONE bug class THREE times in a single round: a selection/filter predicate written against an assumed artifact schema selects ZERO rows from an already-local committed artifact, and the script treats the empty set as a successful no-op and exits rc=0. Two of the three burned real time (one produced a plausible-looking output file that read as a completed run and was misdiagnosed as a missing deliverable for two days); the third was caught only by an ad-hoc dry-run minutes before a ~160,000-call Batch API wave.

Goal

Establish a fail-loud convention for empty post-selection row sets over local committed artifacts — an empty selection must raise, never exit rc=0 as a silent no-op — so this class is caught at the selecting script rather than after the compute is spent.

Workflow gap

Bug observed (three instances, same round):
1. scripts/issue1739_trait_rejudge.py (2026-08-03, recorded at #1739 epm:progress v412): CLI rung names vs on-disk train / hhrt / toxicchat — leg died; fixed by adding _LOCAL_RUNG_ALIAS (commit 0d09a51f49).
2. scripts/issue1739_rescore_ood.py _load_plain_ladder_cells: filtered top-level f_u / u_rung_label while the committed percell/cells.jsonl carries that identity nested inside unit_key — kept 0 of 826 cells and wrote ood_detection_metrics.json with n_metric_rows: 0 at rc=0. Measured fails-pre-fix at round 19: old filter 0/826, new filter 270/826. This one also cost two days of misdiagnosis, because the empty output file looked like a completed run.
3. scripts/issue1739_compliance_full.py _load_rollouts_local: same rung-alias mismatch as (1), carried in a pre-0d09a51f49 copy — would have judged 0 of 159,990 requests at rc=0. Caught pre-launch by a zero-API dry-run; fixed by porting _LOCAL_RUNG_ALIAS (commit 086d2588fe).
Why it is a workflow gap: the recurring cost is not the individual typo — it is that an empty selection is INDISTINGUISHABLE FROM SUCCESS. Every instance exited rc=0, and instance (2) additionally emitted a well-formed artifact. Round 19 added exactly the right guard (raise on empty cell selection AND on a 0-metric-row result set) but only to ONE site, so the next site re-learns it. Three occurrences in one round is the signal that this belongs in the always-consulted rule surface rather than in per-script patches.
Distinct from the existing nearest rule — this is a SIBLING, not a duplicate. .claude/rules/gotchas.md already carries the #1092 real-corpus streaming-filter entry ("a filter chain written from assumed field shapes can reject 100% of rows while every synthetic-fixture smoke stays green", verified present: 1 hit for reject 100% of rows), and that entry even records #1739's own Reddit-dump string-boolean shapes. But its SCOPE is remote streaming corpus builders and its REMEDY is a bounded tiny-real streaming probe with per-filter reject counters. The class filed here differs on both axes: the input is an ALREADY-LOCAL COMMITTED artifact (judge rollout mirrors, per-cell JSONL, cell registries) where no streaming probe applies, and the remedy is a fail-loud non-empty-selection assertion at the selection site. The new entry should cross-reference the #1092 one rather than restate it.
Confidence (emitter): high — three reproducible instances with measured fails-pre-fix evidence for two of them, all in current-tree scripts.
verified-at-filing: per-target grep -c at body-compose time 2026-08-05 — .claude/rules/gotchas.md: reject 100% of rows = 1 (the sibling entry, must exist), empty selection|non-empty-selection|selected zero rows = 0 (the claimed gap, absent); .claude/rules/code-style.md: same gap pattern = 0. All three incident scripts confirmed present on issue-1739 HEAD via git cat-file -e. Recently-closed siblings on gotchas.md in the last 7 days reviewed (5 commits: #2079 sha-pin lint, #2073 autocompact window, #2058 spec sync, the probed-tool-traps pair, #2050 pgrep -f) — none covers empty-selection semantics, so this is not a re-raise of a just-landed fix.

Proposed change (candidate diff sketch — refine in planning)

Add a .claude/rules/gotchas.md entry, sibling to and cross-referencing the #1092 streaming-filter entry: any selection/filter/join step over a local committed artifact must fail loud on an empty result set — an empty post-filter row set, an empty cell selection, or a zero-row result table is a RuntimeError, never a silent rc=0 no-op — and must NOT write a well-formed "success" artifact (instance 2's n_metric_rows: 0 file is the trap). State the three #1739 instances as the evidence and name the two recurring mismatch shapes: CLI-vs-on-disk key aliases, and identity nested inside a JSON-string column rather than at top level.
Consider the code-side companion: a small shared helper (e.g. in src/explore_persona_space/) that wraps "select rows, assert non-empty, report the rejecting predicate" so new scripts inherit it instead of re-deriving — the planner should judge helper-vs-convention. Whatever lands, the per-unit reject-counter idea from the #1092 entry is the right diagnostic to carry over (so a 0-selection failure names WHICH predicate emptied the set).
Do NOT weaken any existing fail-fast rule; this only adds a case.

Scope / surfaces

Primary target: .claude/rules/gotchas.md.
Secondary (planner's call): .claude/rules/code-style.md fail-fast section, and/or a shared selection helper under src/explore_persona_space/.
Evidence-only, do not edit: scripts/issue1739_{trait_rejudge,rescore_ood,compliance_full}.py (all three already fixed in-round).

Constraints / invariants

Workflow-surface only for the rule text; a shared helper would be library code and is in scope only if the planner elects it.
Must not contradict the #1092 streaming entry — cross-reference, do not restate or supersede.
scripts/workflow_lint.py --check-lessons-index passes if a rule row is added/renamed; ruff clean on any touched code; the gotchas size ratchet respected.

Provenance

workflow_fix_target: .claude/rules/gotchas.md
fingerprint: empty-selection-local-artifact-silent-rc0

Raised as prose by the #1739 orchestrator after the third instance in one round was caught minutes before a ~160k-call Batch API wave; the correction + full three-instance table are recorded in an epm:progress marker on #1739 (2026-08-05).