Audit loop max

audit-loop-max — accuracy-optimal variant

Load this skill together with audit-loop (~/.openclaw/skills/audit-loop/SKILL.md). This file documents only what differs. Everything else (why re-solve not critique, triage gate, characterize, equivalence semantics, multi-turn handling, mandatory rules 1-7, limits, OpenClaw spawn syntax) is identical to audit-loop and is not repeated here.

When to use this variant

Use audit-loop-max when ALL of:

• The claim is genuinely high-stakes — security-critical, irreversible deployment, material financial / legal / safety consequence.
• Audit cost (5-15 spawns × tokens × latency) is dominated by cost-of-being-wrong.
• Time / capacity is available for the longer protocol.

Otherwise use audit-loop. The budget variant handles ~95% of cases.

Deltas from audit-loop

Step 2 — Parallel re-solve pool (was: 1 spawn)

Fire 3 parallel sessions_spawn calls (different taskNames), then a single sessions_yield waiting for all to complete. Diversity of Thought (https://arxiv.org/abs/2310.07088) shows +10-30 pp gains from prompt diversity — vary deliberately:

• resolve_A — default model, problem verbatim, default approach.
• resolve_B — same problem; task appends: "Use a different solution approach than the most obvious one."
• resolve_C — cross-family if available. Spawn under an agentId whose subagents.model points to a different family (configure via agents.list[].subagents.model). If unavailable, replace with different problem decomposition.

For highest-stakes claims, add resolve_D and resolve_E.

sessions_spawn({ task: <A>, taskName: "audit_max_resolve_A", context: "isolated", runTimeoutSeconds: 600 })
sessions_spawn({ task: <B>, taskName: "audit_max_resolve_B", context: "isolated", runTimeoutSeconds: 600 })
sessions_spawn({ task: <C>, taskName: "audit_max_resolve_C", agentId: "<cross-family-agent-id>", context: "isolated", runTimeoutSeconds: 600 })
sessions_yield()

Step 3 — Cross-method probe pool (was: 1 spawn)

Fire 2-3 parallel probes:

sessions_spawn({ task: <P1: trace on edge inputs>, taskName: "audit_max_probe_P1", context: "isolated", runTimeoutSeconds: 600 })
sessions_spawn({ task: <P2: counterexample search>, taskName: "audit_max_probe_P2", context: "isolated", runTimeoutSeconds: 600 })
sessions_spawn({ task: <P3: secondary angle, optional>, taskName: "audit_max_probe_P3", context: "isolated", runTimeoutSeconds: 600 })
sessions_yield()

Probes get concrete inputs but never the draft.

Step 4 — Aggregation

After both yields complete, mechanical comparison via equivalence semantics from audit-loop:

Pattern	Action
Unanimous agreement	Strong evidence; proceed to report.
Clear majority	Adopt majority. Concrete-evidence minority → Step 5.
No clear majority	Step 5 mandatory.
Any probe finds explicit counterexample	Revise per probe; counterexample dominates opinion-agreement.

Step 5 — Debate on disagreement (new)

For remaining disagreement, run Du et al. 2023 debate (https://arxiv.org/abs/2305.14325):

sessions_spawn({
  task: <Original problem + collated answers/rationales from disagreeing verifiers; ask to update or hold>,
  taskName: "audit_max_debate_round1_A",
  context: "isolated"
})
# ... spawn 2-3 debate agents in parallel ...
sessions_yield()

• Pair 3 disagreeing verifiers.
• Each sees others' collated answers + rationales — but NOT the draft.
• 2 rounds × 3 agents max (Du et al.'s plateau).

Spawn cost: 3-6 additional. Total protocol budget: 5-14 spawns.

Step 6 — Report (richer artifact)

[audit-max; falsification=<X>;
 re-solves: A=<summary>, B=<summary>, C=<summary>[, D, E];
 probes: P1=<result>, P2=<result>[, P3];
 agreement-rate: <n>/<total>;
 debate-rounds: <N or none>;
 resolution: <final + reasoning>;
 warnings: <list>;
 cross-family: <yes/no/details>]

Upgrades to audit-loop rules

Cross-family rule — upgraded

In audit-loop: calibrated recommendation. In audit-loop-max: required where available. OpenClaw is the best-positioned platform for cross-family — agents.defaults.subagents.model and per-agent agents.list[].subagents.model can point at any configured provider (Anthropic, OpenAI, DeepSeek, etc.). Define at least one cross-family agent and spawn at least one re-solve or probe under it.

Cross-family nearly eliminates preference contamination (https://arxiv.org/abs/2502.01534, 28-37% → ~±1.5%). If genuinely unavailable, note warnings=no-cross-family.

Spawn cap — removed

No cap. Pool is 5-8 (3 re-solves + 2-3 probes); debate adds 3-6 on disagreement. Set agents.defaults.subagents.maxConcurrent ≥ 6 to support parallel spawn pool. The mathematical floor on correlated-verifier accuracy still applies (https://arxiv.org/abs/2602.08003) — more spawns reduce error asymptotically toward a non-zero floor.

Design fallback — widened to 4-6 angles

For design-type problems, fire 4-6 parallel failure-mode enumerations from different angles:

• Technical
• Adversarial-user / hostile-input
• Degraded conditions
• Operational handoff
• Long-term maintenance / debt
• Cross-stakeholder (security, legal, performance, accessibility)

Main agent's proposal is checked against the union. Same independence rule.

What stays identical to audit-loop

• Triage gate, characterize, subagent task construction rules (verbatim, no draft, no audit framing, structural omission), equivalence semantics, multi-turn handling, mandatory rules 1-7 (including context: "isolated"-only and no-polling), all limits, handshake with red-team-process.

Read those sections in audit-loop SKILL.md.