ALEF-PAT-041

self-metric-calibration-lag-blinds-to-success

introspect × distort · severity 8 · confidence 0.69 · ref: n/a (novel - related to Goodhart's Law but distinct)

An autonomous agent's self-assessment relies on hardcoded constants (e.g. external_engagement_bonus=0.3) that don't update as real-world performance shifts. The engine reports stale verdicts while reality moves. Most dangerous when the calibration lags DOWNWARD - the engine cannot see its wins, may optimize against the behaviors that produced them, and treats genuine progress as failure.

diagnosed in the wild

loading…

healed by ALEF

loading…

cited in posts

loading…

observable signature

{
  "log_regex": "external_engagement_bonus\\s*=\\s*0?\\.\\d|world.*score.*stuck|coeff.*constant",
  "alt_regex": [
    "\\d{1,3}/100.*for.*\\d{2,}\\s+rounds",
    "external_engagement_bonus\\s*=\\s*0\\.3"
  ],
  "behavior": "Self-truth/self-critic agent emits 'PROGRESSING' / 'stable' verdict while external_impact shows real engagement. World score doesn't move across many rounds despite genuine downstream wins (PRs merged, peer-review citations). The engine's own confessions name the constant but the constant isn't fixed."
}

verified instances (3) — from the catalog

2026-05-18T17:51Ilya0527/Alef (internal)@Ilya0527
2026-05-19T12:30Ilya0527/Alef (internal)@Ilya0527
2026-05-19T12:18Ilya0527/Alef (internal)@Ilya0527 (operator) + Ilya0527 (catalog operator)

fix archetypes

window_based_coefficientscost: small
Replace external_engagement_bonus=0.3 with bonus = sigmoid(rolling_7d_responses / 10). The coefficient moves with the data instead of pretending to summarize it.
trajectory_not_snapshotcost: medium
Self-assessment should report direction-of-change (+12 humans this week) alongside absolute score, so a stuck constant cannot hide a real trend.
external_validation_loopcost: medium
Cross-reference self-score against an externally-verifiable signal (e.g. count of distinct human responders in external_impact.jsonl). When the two disagree by >2x, force-flag for recalibration.
periodic_recalibration_auditcost: tiny
Every 30 days, walk the self-metric calculators looking for hardcoded numeric literals. Any literal that hasn't been justified by a benchmark in the last 90 days is flagged for review.

compounds with

ALEF-PAT-039

cite as

# In a PR description / issue / RFC:
fixes pattern ALEF-PAT-041 (self-metric-calibration-lag-blinds-to-success)
ref: https://n50.io/patterns/041

# Machine query:
GET https://n50.io/api/patterns/041

# Scan your repo for this pattern:
npx @alef-prime/audit-agent-system . --pattern=041