martin-loop & codex-residual-fidelity

[tylorsaling-source]

@Keesan12 — following up on our exchange in openai/codex#35528.

I ran the deterministic "codex-residual-fidelity" checks through MartinLoop "0.4.5" using the verifier-only lane. MartinLoop handled the boundary correctly: it verified the checks and receipt integrity while preserving "completeEvidence: false", rather than claiming that evidence already discarded by Codex had somehow been recovered.

That suggests a useful division between the projects:

• "codex-residual-fidelity" defines and tests the producer-side residual across capture, model-visible, and durable state;
• MartinLoop binds that evidence to the run and refuses to overclaim when the evidence is incomplete.

Before I build anything around that relationship, I wanted to ask directly for your blessing. Would you be comfortable with—and interested in—an optional MartinLoop verification integration in "codex-residual-fidelity"?

My thought is a reproducible verifier-only lane that emits a redacted integrity receipt, preserves MartinLoop’s "EVIDENCE_BOUNDARY" behavior, and clearly distinguishes verified execution from recovery of omitted evidence.

I also don’t want to describe it as a supported MartinLoop integration or imply your endorsement without your explicit approval. Would you prefer it to remain entirely within "codex-residual-fidelity", or would you like us to coordinate the receipt fields and integration boundary so it can also be useful to MartinLoop?