Skip to content

AI_DISCLOSURE.md

Latest commit

 

History

History
60 lines (49 loc) · 2.64 KB

AI_DISCLOSURE.md

File metadata and controls

60 lines (49 loc) · 2.64 KB
Raw

AI Disclosure — hyperevm-safety

The hyperevm-safety library is built and maintained by CaliperForge under an AI-augmented authoring stack. This document is calm disclosure of which surfaces are AI-touched and the review discipline that gates each one.

What is AI-touched

  • Invariant proposals. Candidate invariant predicates (the property prose, the Solidity property body, the Halmos / Certora spec) are drafted by a Claude model (Sonnet 4.6 for ABI / storage-layout pattern recognition; Opus 4.6 / Apex for judgment-heavy reasoning about precompile and CoreWriter semantics). Every accepted invariant is reviewed and edited by the case specialist (solidity_specialist, occasionally in the FV register) before being committed.
  • Same-source twins. The reconstructed code in each incident's clean/ and planted/ subtrees, and the planted-hunk references used for the library's own property tests, are authored by a CaliperForge specialist against the affected protocol's published post-mortem. The reconstruction is faithful to the post-mortem's described bug class; it is NOT a fork of the protocol team's production source.
  • READMEs and case write-ups. Drafted with AI assistance; reviewed against CaliperForge's internal anti-AI-ism and register rubric before publish.

What is NOT AI-touched

  • The published post-mortem URLs themselves (carried as-cited).
  • The CI verdict (pass / fail is a function of the forge / echidna / medusa / halmos run, not the model).
  • The operator-approved positioning paragraph (README.md, the §1.1-locked block) — that text is human-locked verbatim across surfaces.
  • The operator's final-pass sign-off decisions and the milestone exit-gate sign-offs.

Audit trail

  • Every incident's README.md cites the primary public post-mortem URL (re-verified by research_lead per the T-9 build ticket; ≥2 primary sources per incident before the case is merged).
  • Every hyperevm-safety commit lists the author (Michael Moffett, operator at CaliperForge) and is operator-clean (no Co-Authored-By trailers).
  • The library-properties + incident clean/planted CI legs are uploaded as artifacts on every push (see .github/workflows/ci.yml).

Why we disclose

CaliperForge's identity register makes AI-augmented authorship the default disclosure posture, not the exception. Reviewers should know which content was AI-drafted so they can apply their own scrutiny at that surface. See caliperforge.com/ai-disclosure for the org-level register.

Contact

Operator: Michael Moffett — michael@caliperforge.com — team@caliperforge.com.