Catalog/E family/E06
Proposed specificationSource-located exemplarH3

Joint harness-policy evolution

Adapt runtime structure and policy together.

JOIN POINTcoadaptive_training_loopRECORD STATUSCandidate · v0.1.0
UML class view for Joint harness-policy evolution, showing participant roles, received calls, and analytical dependencies
FIG. E06 / UML CLASS VIEW
Object interaction diagram for Joint harness-policy evolution, showing role collaboration and a numbered message sequence
FIG. E06 / OBJECT INTERACTIONS
Context
Runtime harness and model policy can both change.
Trigger
An evolution round evaluates task behavior and proposes changes to either or both.
Structure
Task interface, harness representation, policy representation, and joint evaluator.
Effect
The selected pair is evaluated as a pair with attribution for harness and policy changes.
Requires
Policy and harness variants are versioned jointly.
Ensures
Results identify the evaluated pair and separate objective contributions where possible.
Invariant
A policy gain is not reported as a harness gain without an ablation.
Violation
Failed candidates or incompatible pairs are excluded with reason.

Negative control

Retraining only the model with a fixed harness is not joint evolution.

Permitted variation

Curated pairs, coevolution, alternating optimization, and ablation controls.

Test obligation

Test pair compatibility, harness-only, policy-only, joint, and transfer.

Evolution question

Did research move from separate optimization to paired variants?

PROMPT SKETCH
Version policy and harness variants as a pair. Every result must identify the evaluated pair, and separate each objective's contribution where the design permits.
CODE DESIGN / PYTHON
def evaluate_pair(harness, policy, suite):
    result = suite.run(harness=harness, policy=policy)
    return PairResult(
        harness_version=harness.version,
        policy_version=policy.version,
        metrics=result.metrics,
        ablations=suite.ablate(harness, policy),
    )

Illustrative design sketch. A score without the pair identity cannot support a joint-evolution claim. This example is a teaching aid, not implementation evidence or a conformance claim.

Join point
coadaptive_training_loop
Source locator
Chen et al.
Lexical cue
(?i)\b(?:joint[_-]?evolution|evolve\w*\s+(?:harness|policy)|policy[_-]?evolution)\bFrozen expression, v0.1.0. A match is a retrieval cue, not a confirmed implementation.
Aliases
coevolution, jointoptim, harnesspolicy, policyevolution
Canonical exemplar
mingju-c/HarnessForgeHarnessForge_4B/evolved_pairs/pairs.yaml:1Support: README.md:15
Negative control
Retraining only the model with a fixed harness is not joint evolution.

Scope-anchored review. pohe-agent analyze <repo> --pattern E06 anchors the cue to its enclosing declaration and checks a separate operational signature for all four pointcut dimensions. Missing roles and negative-control risks remain visible; a structural match is not E2/E3. Calibration report ↗

Source-located exemplar

A concrete source anchor is available for this proposed record. It is one exemplar, not evidence that all repositories—or every contract field—conform.

Review boundary: this candidate and its specification come from an AI-assisted study. Independent human validation remains open.

ANY-EVIDENCE FAMILY COVERAGE1/36

2.8% of repository families in the frozen panel have at least one current cue, historical cue, or canonical source layer. This is not confirmed prevalence.

Current lexical-cue repositories: 1 / 41
Canonical source anchormingju-c/HarnessForge 8bef9488c35aHarnessForge_4B/evolved_pairs/pairs.yaml:1
Supporting surfaceREADME.md:15

Curated harness-policy pairs.

Exploratory retrieval cue Regex / v0.1.0
(?i)\b(?:joint[_-]?evolution|evolve\w*\s+(?:harness|policy)|policy[_-]?evolution)\b

This expression is a discovery instrument only. A match is a retrieval cue—not a confirmed implementation, pointcut match, or contract proof.