Skip to content

graff — a fast agentic coding harness in Zig: multi-provider, MCP, workflows, DGM evolution loop, TS/Python SDKs

Python
1.7 KB
Raw Blame

0194. Native learning checks a pinned formal baseline

Status: accepted 2026-09-23

Context

The local learning store already pins its adapters and evaluation suites, and rechecks their immutable evidence before promotion. A separate, bounded formal checker can establish properties of the shared engine's modeled control flow. Its success is not evidence that a newly generated prompt is correct. The existing signed aggregate receipt also has no field for a formal check.

Decision

An optional formal_check configuration pins the checker executable, its script, and its input manifest. The manifest's engine binary must be an exact pinned input and argument to both learning adapters. Before mutation, and again for a selected genome, the checker must succeed for the respective prompt fingerprint. The store saves content-addressed check evidence in a versioned pending record and run record. Resume and manual or automatic promotion re-run the pinned checker and reject missing, stale, or mismatched evidence. Default configurations do not enable this gate.

Formal-enabled trials stay local. run --submit, learn submit, and direct submission reject them until a signed receipt version binds the formal evidence. The legacy fleet persona promotion path remains separate and has no formal gate.

Consequences

The checker consumes local time at admission, selection, resume, and promotion. Operators must supply its signing key file and trust the pinned checker, adapters, and manifest; pinning proves their byte identity, not that an adapter actually uses its binary argument. The model-check result concerns the pinned shared engine baseline. Behavioral correctness of a candidate prompt still depends on the normal paired evaluations and holdout gate.