Audit loop max

audit-loop-max — accuracy-optimal variant

Load this skill together with audit-loop (~/.claude/skills/audit-loop/SKILL.md). This file documents only what differs. Everything else (why re-solve not critique, triage gate, characterize step, equivalence semantics, multi-turn handling, mandatory rules 1-4, all limits) 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)

Spawn 3 independent re-solves in parallel via the Agent tool. Diversity of Thought (https://arxiv.org/abs/2310.07088) shows +10-30 pp gains from prompt-method diversity over same-prompt sampling — vary the prompts deliberately:

• A — default model, problem verbatim, default approach.
• B — same problem; append: "Use a different solution approach than the most obvious one."
• C — cross-family if available (custom subagent in .claude/agents/auditor.md pointing at a different model family via MCP). If genuinely unavailable, replace C with a different problem decomposition (e.g., bottom-up vs top-down).

For the highest-stakes claims, add D and E with further method variation (up to 5 re-solves total).

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

Spawn 2-3 probes in parallel, each per a different falsification angle:

• P1 — primary falsification (trace on edge inputs / recompute via alternative decomposition).
• P2 — counterexample search ("find input violating property Y, or report 'none after honest search'").
• P3 (optional) — secondary falsification angle when applicable.

Step 4 — Aggregation

After Steps 2-3 complete (parallel), apply equivalence-semantics rules from audit-loop to count agreements across all outputs (draft + re-solves + probes).

Pattern	Action
Unanimous agreement	Strong evidence; proceed to report.
Clear majority for one answer	Adopt majority. If minority brings concrete evidence (counterexample, explicit contradiction), proceed to Step 5.
No clear majority	Step 5 mandatory.
Any probe finds explicit counterexample	Revise per probe regardless of re-solve agreement. Counterexample is constructive evidence; it dominates opinion-agreement.

Step 5 — Debate on disagreement (new)

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

• Pair the disagreeing verifiers (typically 3 selected from the pool).
• Show each the others' collated answers + brief rationales — but NOT the original draft.
• Each verifier updates or holds position with reasoning.
• Cap: 2 rounds × 3 agents (Du et al.'s reported plateau).

Spawn cost for debate: 3-6 additional spawns. 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 from "calibrated recommendation" to "required where available"

At the max-accuracy frontier, the small absolute gain from eliminating preference contamination (28-37% → ~±1.5%; https://arxiv.org/abs/2502.01534) is worth the setup. At least one re-solve and/or one probe must use a different model family. If cross-family is genuinely unavailable, note warnings=no-cross-family and accept the protocol operates at a degraded accuracy ceiling.

Spawn cap — removed

audit-loop's hard cap of 2 is a budget ceiling; this variant has no cap. Pool of 5-8 (3 re-solves + 2-3 probes); debate adds 3-6 if disagreement remains. Literature supports this range (BoN-MAV optima 6-14; debate plateau 5-10; self-consistency knee 10-20 — but those are for same-model sampling, less relevant once cross-method diversity is in play).

The mathematical floor on correlated-verifier accuracy still applies (https://arxiv.org/abs/2602.08003) — more spawns do not drive error to zero, only reduce it asymptotically toward a non-zero floor determined by inter-model correlation.

Design fallback — widened from 2 angles to 4-6

For design-type problems (no objective single answer), replace re-solves + probes with 4-6 failure-mode enumerations from different angles:

• Technical failure modes
• Adversarial-user / hostile-input
• Degraded conditions (failed dependencies, partial outages, scale extremes)
• Operational handoff
• Long-term maintenance / debt
• Cross-stakeholder (security, legal, performance, accessibility)

Main agent's proposal is checked against the union of all enumerated lists. Same independence rule as audit-loop — each subagent enumerates without seeing any proposed solution.

What stays identical to audit-loop (NOT repeated here — read audit-loop SKILL.md)

• Triage gate (3-way + bypass cases).
• Step 1 Characterize (CLAIM + FALSIFICATION SHAPE).
• Subagent prompt construction rules (verbatim transcription; no draft; no audit framing; structural omission for negative-prompting).
• Equivalence semantics (canonical forms; default to disagreement when unsure; no third-LLM judge).
• Multi-turn three-bucket handling (original / confirmed constraints / forbidden thoughts).
• Mandatory rules 1-4 (verbatim transcription; never show draft; no adversarial framing; report disagreement honestly).
• Explicit limits (mathematical floor, capability-driven correlation, bypass cases, cost considerations apply with adjusted thresholds).
• Handshake with red-team-process.