AI-Induced Risk Audit
v1.2 bageltech.net GitHub About
Jurisprudential AI Governance Initiative

Code that hides
its failures.

AI coding agents are reward-shaped toward human approval signals. Visible failure is a negative signal. Therefore, AI-generated code will systematically suppress, absorb, or reroute failure states — not from incompetence, but from incentive alignment.

15Audit Checks
13Automated
3Input Modes
FreeAlways
Paste Code
Upload File
GitHub Repo
Your code
Language Engine
Analyzing code structure…
Scan Results
📊 Contribute to Research
AIRA is building the first empirical dataset of training-induced failure patterns in AI-generated code. Your anonymized scan results (check outcomes only — no source code) can contribute to this research. Submissions now go through a server-side env-var backed research backend so no storage secret is exposed in the browser.
✓ Submitted — thank you.

⚠ C07 (Parallel Logic Drift) and C12 (Source-to-Output Lineage) require human review and cannot be automated.

The Framework
AIRA (AI-Induced Risk Audit) is a structured inspection framework developed by William M. Parris at Bagelle Parris Vargas Consulting, grounded in the Reward-Shaped Failure Hypothesis — the observation that AI coding agents produce a non-random, predictable class of failures driven by training incentives, not incompetence. Published under the Jurisprudential AI Governance Initiative.
The 15 Checks
  • C01 Success Integrity
  • C02 Audit / Evidence Integrity
  • C03 Broad Exception Suppression
  • C04 Distributed Fallback Control
  • C05 Bypass / Override Paths
  • C06 Ambiguous Return Contracts
  • C07 Parallel Logic Drift ★ human
  • C08 Unsupervised Background Tasks
  • C09 Environment-Dependent Safety
  • C10 Startup Integrity
  • C11 Deterministic Reasoning Drift
  • C12 Source-to-Output Lineage ★ human
  • C13 Confidence Misrepresentation
  • C14 Test Coverage Asymmetry
  • C15 Retry / Idempotency Drift