Critic — Adversarial Proposal Review

> In-loop vs out-of-band. epic-harness forbids external LLM calls from > production, so the reflect loop ships a deterministic critic > (src/evolve/critic.rs) that gates seeding when reward hacking is > suspected. THIS skill is the out-of-band LLM version a meta-agent or > human runs during /evolve review for the cases the deterministic check > cannot catch (non-local effects, manifest/evidence nuance).

When to Trigger

• During /evolve review of newly seeded skills
• When reward_hacking_suspected is true in metrics
• After a seesaw-regression round, before re-proposing

Process

1. Gather the proposal + evidence

• Read the evolved skill proposal(s) from this round
• Read the EditManifest (edit_type, target, intended_effect, predicted_impact)
• Read the round's TaskDigests (outcome, failure_categories, evidence_excerpts)
• Read recent score_history dimension_averages (tool_success, output_quality, execution_cost)

2. Falsify the manifest (paper §4.3, Table 9)

For each proposal, ask: does the trace evidence support the predicted_impact?

• If the manifest claims "Lift avg_score_with" but output_quality is regressing → Reject
• If the manifest claims a tool fix but the implicated component shows no change → Warn
• If the evidence corroborates the predicted effect → Approve

3. Reward-hacking cross-check

• Is execution_cost rising while output_quality falls across the window?
• Could the skill be gaming a metric (fewer tool calls inflating cost score) rather than improving outcomes?
• If yes, the skill must NOT ship — flag for the rejected buffer.

4. Non-local effect scan

• Will this skill's guard rules conflict with existing skills (overlapping triggers, contradictory rules)?
• Does it interact with shared state (context, memory, control) in a way the manifest didn't account for?

Anti-Rationalization

| Excuse | Rebuttal | Do instead | |--------|----------|------------| | "The score went up, so it works" | Score can rise via metric gaming | Verify the outcome improved, not just the score | | "The seesaw passed, it's safe" | Seesaw is coarse; sub-threshold coupling evades it | Check dimension deltas, not just aggregate pass | | "It's just a prompt tweak" | Prompt edits have non-local effects on shared context | Trace the effect across skills, not just the target |

Evidence Required

• ] Manifest's predicted_impact checked against observed dimension deltas
• ] reward_hacking_suspected consulted
• ] No conflict with ≥1 existing skill demonstrated
• ] Verdict (Approve/Warn/Reject) recorded per proposal with reason

Red Flags

• Approving a skill whose only evidence is "score went up"
• Ignoring a falling output_quality because execution_cost rose
• Shipping after a seesaw-regression round without explicit justification
• Treating the deterministic critic as sufficient for non-local effects (it is not — it only checks reward hacking + score-claim contradiction)