Who audits the auditors?

Evidence-Bound Incentives for AI-Agent Security Evaluation

Howie Xu · Naysay · APA 7th edition · 15 September 2026

Read the paper   /   Download PDF (19 pages)
Download source & replication files (ZIP)   /   BibTeX references

The question

When an evaluator can profit from careless testing or concealed failures, what makes honest evaluation the better economic choice?

The argument

Bind each report to a test scope, an authenticated execution record, and an explicit dispute procedure. Then ask whether the probability of detecting misconduct, after accounting for false penalties, makes the collectible loss exceed the gain from cheating.

The paper derives that condition and examines its limits: shared auditor failures, audit costs, incomplete records, and collateral that loses value before enforcement.

One result

In a declared synthetic scenario, an 80% fall in native collateral reverses the incentive margin from +3.20 to −7.02. The token’s value at enforcement matters; its initial market value is insufficient.

What this establishes

A formal model, proofs, ten references, and reproducible synthetic stress tests. A native token may capture value from useful, paid evaluation work; customer demand, net surplus, and value growth still require evidence.

This is an AI-assisted working draft, not peer reviewed. The numerical results are simulations and model calculations. Real-agent performance and commercial demand have not been measured.

Inspect an illustrative evidence trail   /   Read the proposed network model