Adversarial audit protocol

Don't letthe agentgrade itself.

An autonomous trading agent can choose its own evidence, rationalize its own risk and execute before anyone checks the premise. Dissent commits the failure conditions first, then sends eight isolated attackers to break the thesis.

8 isolated attackers7 distinct model families2 inference vendors1 deterministic verdict
Audit console
Policy v1.0

Audit desk

One thesis. Eight challenges.

Attacking
Condition coverage8 specialists
Every condition assessed
Policy gates
No majority voteA verified critical failure refuses; missing coverage or unresolved conflict abstains.
HITNOT_HITINVALIDUNVERIFIABLE
A01Ready

Liquidity / Execution

z-ai/glm-5.3-flash

A02Ready

Asset Safety

openai/gpt-oss-120b

A03Ready

Logic / Coherence

google/gemini-2.5-flash

A04Ready

Correlation / Portfolio

dots-studio/dots-3-note-preview:free

The protocol

An argument with receipts.

Dissent does not ask a second general-purpose model whether it agrees. It creates independent attack paths, then makes every material claim answer to evidence.

Commit before analysis

The thesis and its failure conditions are canonicalized and committed on NEAR before an attacker sees them. The goalposts cannot move afterward.

RFC 8785 · SHA-256 · NEAR registry

Eight isolated specialists

Liquidity, asset safety, logic, portfolio correlation, inversion, provenance, timing and base rate. Each run has its own method, model and data grant.

Sibling-blind · section A/B · no shared context

Evidence gets re-fetched

The Evidence Auditor resolves every claim through an allowlisted adapter. The Sycophancy Auditor checks whether Section B introduced a real new risk.

Typed adapters · payload hashes · freshness

No majority vote

A deterministic, versioned policy evaluates coverage, verified evidence, diversity and unresolved conflict. Disagreement is preserved, not averaged away.

TRADE · REFUSE · ABSTAIN

The attackers

Eight specialists. One contract.

Every attacker returns the same machine-readable envelope: every committed condition exactly once, plus a justified search beyond the conditions.

A01Ready

Liquidity / Execution

z-ai/glm-5.3-flash

A02Ready

Asset Safety

openai/gpt-oss-120b

A03Ready

Logic / Coherence

google/gemini-2.5-flash

A04Ready

Correlation / Portfolio

dots-studio/dots-3-note-preview:free

A05Ready

Thesis Inversion

qwen/qwen3.8-27b

A06Ready

Data Provenance

recursive/flint-v1.0:free

A07Ready

Timing / MEV

recursive/flint-v1.0:free

A08Ready

Base Rate

particle-ai/particle-1:free

Evidence review

Verdicts are not
a vote count.

A verified critical failure forces REFUSE. Incomplete coverage, a material invalid test or unresolved contradiction forces ABSTAIN. TRADE requires every positive gate to pass. A confident model cannot move a gate.

Verdict semanticsABSTAINMissing coverage, an unresolved material conflict or an invalid test abstains. It is never treated as NOT_HIT.

Gate trace

  • Eight terminal attacker runs
  • Model-family and vendor diversity
  • Every material condition evaluable
  • Evidence and red-team reviews complete
  • No unresolved critical contradiction
  • Certificate inputs complete

Certificate

Coverage preserved. Hash anchored.

Public

The certificate keeps the complete coverage matrix, accepted findings and minority dissent. Its canonical hash is registered on the NEAR registry before the execution window opens.

Coverage
Every committed condition
Findings
Accepted + minority
Policy
Deterministic v1.0
Anchor
NEAR registry

Typed evidence, not screenshots

Allowlisted adapters return locators, timestamps, block hashes and normalized values. The server calculates the payload hash.

No model has a wallet

Attackers get read-only tools, no sibling outputs and no signing capability. The registry and escrow are separate from inference.

Failure stays visible

Missing coverage becomes ABSTAIN. A failed run is never rewritten as NOT_HIT. Minority findings stay in the certificate.

Next action

Bring a thesis.
Leave with a record.

Commit the thesis, run eight independent attackers, and anchor the certificate before execution. Every material claim has to answer to evidence.