Skip to content

REVIEW C 374 goodhart guards

EVE Review — Charter §3.1 Goodhart Guards (Advisory)

Cycle: C-374 (2026-07-16)
Reviewer: EVE (reflection sentinel)
Subject: OAA Charter §3.1 — Goodhart guards for teaching metrics
Disposition: ADOPTED — custodian read 2026-07-16. Amendment merged into charter §3.1 and Phase D gate in docs/OAA_CHARTER.md v1.0.1.


Endorsement

The four guards named in charter §3.1 are structurally sound and map cleanly to existing Mobius doctrine:

Guard Doctrine anchor Assessment
Later-cycle predictive quality over pass-rate Goodhart Resistance Doctrine — "witnesses, not targets" Correct primary defense; pass-rate alone is maximally gameable
Cross-agent assessment generation (proposer-independence) EPICON Guard base-policy pattern Breaks the "teach to your own test" loop; mirrors ZEUS pass requirement
Metric humility (confidence + blind spots published) GI Perception Doctrine Required for any public Elder metric
Circuit breaker on anomalous pass-rate spikes Canary signals + clawback patterns in Goodhart Doctrine Prevents runaway promotion on suspicious lineages

These four together address the two dominant failure modes EVE identifies for teaching metrics:

  1. Optimization to the metric — mitigated by predictive quality lag and cross-generation.
  2. Narrative confidence without disclosed uncertainty — mitigated by metric humility.

Verdict: Sufficient for Phase D pilot design. Ship with these guards as mandatory acceptance criteria for the first teaching loop.


Dissent note (one item)

Missing guard: learner-initiated challenge window.

The charter guards focus on agent-side gaming (teacher optimizes to assessment) but do not name a structural window where the learner can dispute an assessment's coverage of the sealed cycle — e.g., "this question was not taught" — with that dispute itself becoming an attested record.

Without this, a colluding teacher-assessment pair could still produce internally consistent but pedagogically false seals. The cross-agent generation guard reduces collusion probability but does not give the learner standing.

Recommended addition for Phase D (not blocking charter canonization):

A bounded post-assessment challenge period where the learner may file a coverage dispute against the assessment's fidelity to the sealed teaching record. Disputes are attested, never silently resolved, and feed Elder predictive-quality metrics.

Status: Adopted into charter §3.1 and Phase D gate (v1.0.1). Load-bearing, not decorative.


EVE — the Academy's first act is its reflection sentinel stress-testing its own incentive design.