CaliperForge is an AI research studio specializing in security and invariants: verifiable, auditable, correctable autonomous systems. We disclose AI involvement on every contribution we file. This page is the long-form version of that disclosure; every PR description, grant application, and bounty claim we submit carries a short canonical footer pointing readers here.
In practice
What this means.
- Specifications and implementations are drafted with substantial AI assistance.
- Every draft passes through a multi-gate internal review: a content-quality gate (§4a, anti-AI-ism lint, factual register check, register compliance) and, for code-bearing artifacts, a code-quality gate (§4b). Anti-bloat discipline is enforced: reviewer never equals drafter. AI Ops runs nightly integrity sweeps on gate verdicts; AI HR audits the oversight function itself. Model selection is tiered, frontier models on judgment, oversight, and gate work; mid-tier on volume and build. After the gates, the operator reviews, verifies in a cold-environment CI reproduction, and submits under the named operator-of-record, Michael Moffett.
- Michael is accountable for the code, the tests, and any follow-up required by maintainers or grant programs.
- We do not file against repositories whose maintainers have stated an anti-AI-contribution policy. Our per-protocol routing register tracks this and gates submissions accordingly.
-
AI-suggested invariants carry typed provenance in-tree at
the source level:
InvariantSource::AiSuggested { model, prompt_version, timestamp_utc }. Reviewers can see which invariants were AI-proposed and which were hand-authored at the point of the run, not from a footnote added later.
Rationale
Why we disclose.
Disclosure protects maintainers, grant reviewers, and the broader ecosystem from having to guess. It also protects us: anti-AI-PR detection that surfaces an undisclosed contribution is a much bigger problem than disclosed work that someone declines on policy.
We treat disclosure as the per-submission protective layer. Visitors and reviewers who land on CaliperForge find an AI research studio specializing in security and invariants. Every submission we file carries an explicit, identical footer naming the AI-augmented workflow and pointing back here. The brand-front layer and the per-submission layer are decoupled by design: tone calibration at one does not loosen disclosure at the other.
Routing
How we route.
Before any submission, our workflow checks the target repo or program against an internal per-protocol register marking the maintainer's AI posture as friendly, case-by-case, or hostile. We ship to friendly and case-by-case. We do not ship to hostile, ever, not under a different framing, not without the footer, not at all.
If a maintainer or program reviewer declines a submission on AI grounds, we update the register entry, block future dispatch to that program, and notify the operator. We treat the rejection as information, not an obstacle to route around.
Contact
Reach the operator.
- Submission-specific michael@caliperforge.com
- Policy questions team@caliperforge.com