[isadominguez314]

A run whose final assistant message was destroyed before the harness read it still sent the stand-in text to the OutcomeValidity judge, which confidently scored it 0.0. Observed live on run_20260911_172304_ocvertex-oc (the #32 validation run): openclaw's own sanitizer stored [Malformed diagnostic JSON redacted] over the final message inside its sqlite, the placeholder landed in the record's output, and the judge graded it — OutcomeValidity 0.0 on a run with a 149-step trajectory and a passing verification report. On a task with no verification_spec, that false zero becomes the published correctness.

What changed

• core/run_status.py — new shared is_placeholder_output(): full-string match on the bracketed redaction family ([Malformed diagnostic JSON redacted], [Oversized diagnostic JSON redacted], and shape-siblings). Anchored to the whole string, so a real answer that merely quotes a placeholder never matches. Lives beside is_unscoreable_run — same "one definition, read by every layer" rationale.
• metrics/pipeline.py — _build_context computes final_output_missing: a redaction placeholder always counts; an empty output counts only beside a non-empty trajectory (there the blank is a capture failure — an empty output beside an empty trajectory is the agent genuinely turning in nothing, and that zero is real and still judged).
• metrics/outcome_validity.py — when the flag is set the metric abstains instead of calling the judge: {"score": null, "reason": "withheld: …"} plus a warning log. A null score is skipped by the correctness preference chains in both metrics/pipeline.py and results/normalize.py (verified — no changes needed there), so correctness falls through to the deterministic sources, and when nothing measured anything the composite is withheld: a null row, not a confident zero nobody measured.

The abstention is deliberately scoped to OutcomeValidity: it is the only metric that grades the final text alone. The checklist / safety / grounding judges read the combined case that embeds the full trajectory, so they keep real signal even when the final text is gone.

Validation

• uv run pytest tests/unit → 1861 passed (9 new); ruff check + format clean.
• New tests cover the predicate edges (both observed placeholders, whitespace, quoted-inside-an-answer, empty/non-string), the four _build_context cases, and an end-to-end batch test reproducing the live run's exact record shape — asserting the judge is never invoked and no composite is emitted.
• Behavior documented in docs/components/metrics.md beside the existing OutcomeValidity notes.

Relation to other follow-ups

This is the scoring-side backstop. The parsing-side fix (mapping the placeholder to empty in openclaw/parsing.py so the assistant.message fallback can recover the real text) is a separate, complementary change — once it lands, this layer still protects records where the fallback also comes up empty, and historical results.json artifacts on re-score.