Serving max_tokens cut to 60: hermetically invisible, 434 tests green, and the
deployed gate dropped 30/30 to 19/30 on the scorecard.

The targeting is what makes it evidence rather than noise. Calibration stayed
10/10 and adversarial 9/9 — neither depends on long answers — while enrichment
fell to 12% and summarize to 62%. Both grading paths fired independently: the
schema assert on "Unterminated string", and the judge diagnosing truncation
from the answer alone, quoting the sentence it was cut off at.

Three attempts to get here. The row keeps all three, including the one that
broke nothing and the one that died in calibration, because the failures are
the part worth reading.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>