skinny-cloud/crucible-measurement-integrity

The CRUCIBLE Protocol — Auditing Measurement Integrity in AI-Assisted Software Development

Author: Asif Waliuddin · Affiliation: NXTG.AI, Chicago, USA · License: CC BY 4.0

AI-assisted development can make the same agent responsible for implementation, tests, and the metrics used to judge both. CRUCIBLE is a protocol for auditing whether that measurement system provides evidence not controlled by the same producer.

This repository holds the paper (preview PDF) and pointers to the public code and evidence for an industrial-experience report — not a causal-effectiveness or prevalence claim.

What's here

• crucible-arxiv-preview.pdf — the full paper (27 pages, 4 figures).

The empirical core (three forensic cases, one 15-project AI-governed portfolio)

1. 3,277 passing tests coexisted with silent graph-metadata loss — a green suite over a broken measurement.
2. A hardcoded 77% coverage badge sat over an omit list excluding 1,145 lines of core-engine code (audit estimate ≈ 15% real coverage).
3. 4,726 passing tests did not distinguish isolated-unit evidence from real-database reachability — prompting a separately reported 138-test integration suite.

The public reference implementation passed structural smoke tests on three unrelated Python repositories while exposing precision limits in the silent-exception heuristic. Because the cases, protocol, and evaluation come from the same organization, these results establish operational feasibility and failure-mechanism evidence, not prevalence or causal effectiveness.

Why it matters (the frontier)

The same failure mode — a producer grading its own work — is what the emerging agent-reliability / measurement-integrity literature (e.g. Holistic Agent Leaderboard, "Towards a Science of AI Agent Reliability") attacks from the leaderboard side. CRUCIBLE contributes the audit-forensics side from a real production deployment.

Links

• Code: https://github.com/nxtg-ai/crucible-protocol
• Portfolio: https://nxtg.ai

Citation

@misc{waliuddin2026crucible,
  title  = {The CRUCIBLE Protocol: Auditing Measurement Integrity in AI-Assisted Software Development},
  author = {Waliuddin, Asif},
  year   = {2026},
  note   = {NXTG.AI industrial-experience report},
  howpublished = {Hugging Face}
}