Architecture & reproduction

Compile once, cache, verify deterministically — and prove it offline.

Cold path

One model call per unique task. The Contract Agent sees the task, the tools and the domain schema — never a trajectory, a state snapshot, a final response, a case id or a gold file. The compiled contract is validated, semantically linted, hashed and written to a bundle.

3 calls · 29,889 tokens · 53.6 s

Warm path

The bundle is loaded, every hash re-derived, and each case's task fingerprint recomputed from the task in front of it. A mismatch fails closed rather than recompiling — silently filling a gap would turn a measured warm run into a partly cold one.

0 calls · 0 tokens · 386 ms

Gold isolation

Gold contracts and gold verdicts live behind a package boundary (@stateproof/benchmark/gold). The prediction phase imports only the agent-facing surface, so it cannot reach them, and predictions are written to disk before the scorer opens its first gold file. A test observes every case-file read and asserts that ordering directly.

Reproduce this

No API credential is required. The replay loads the committed contract bundle, re-verifies the eight development cases and compares canonical predictions to the pinned warm run.

pnpm install
pnpm reproduce

Expect, at a high level:

Core-12 and Hard-12 fixtures validate
the pinned registry and every artifact hash verify
zero model calls, zero tokens, no raw response files written
predictions byte-identical to RUN-stateproof-hard-development-warm-20260829T022344Z
SVR / FVR / CDR / BVA identical to the pinned report
RESULT: PASSED

Other commands: pnpm reproduce:check (artifacts and provenance only), pnpm dashboard:build, pnpm dev, pnpm check:provenance <runId>.

Pinned registry: submission/reproduction-manifest.json
Contract bundle: RUN-stateproof-hard-development-cold-20260829T022133Z-contracts
Assertion schema: 2.1.0
Source commit: 42135267c23841b7c8bb960c01749f58bb53481a
Judge summary: artifacts/submission/development-summary.md

Limitations

Validation is demonstrated in a synthetic refund-operations domain. Nothing here establishes behaviour on real production systems.
The semantic lint's task-fact extraction is template-oriented and regex-based. A broader domain needs typed task adapters rather than pattern matching.
Eight development cases, four held out. The locked cases were evaluated exactly once after the source freeze and were never used for tuning.
USD figures are an estimate against a dated, sourced price list, not an invoice. Token counts are measured; the money is derived from them.
Two historical provenance defects are preserved and documented rather than repaired: the Gate 3A run predates its own commit, and the Gate 3C cold manifest carries a stale stage label. Both are cosmetic-to-the-result and both are visible in the artifacts.
This is not a claim of production readiness.

Human review for consequential actions

StateProof does not autonomously approve or execute consequential actions. In a real-world deployment, failed or uncertain verification results should be escalated to a qualified human reviewer before any consequential decision is made.