Evaluator control
The named independent evaluator controls authorized data access, held-out execution, analysis, deviations, and reporting.
External replication docket v1
Independent review starts with control, not endorsement.
This protocol gives a qualified evaluator control of one source, holdout, run, analysis, and outcome record. It is designed so a neutral or failed result remains publishable and the founder cannot manufacture independent authority.
Current state: HOLD / template unassigned. No external evaluator, external result, field validation, certification, savings, deployment, or endorsement is claimed.
Pre-registered comparison
One hypothesis, baseline, candidate, primary metric, threshold, uncertainty method, failures, and stop rules are frozen before scoring.
Reproducible custody
Code, dependencies, inputs, outputs, run receipts, deviations, negative results, and SHA-256 identities remain inspectable offline.
Review instruments
Three levels, three different conclusions.
The public docket separates internal custody checks, executable-computation review, and adversarial synthetic testing so one passing layer cannot be presented as a broader validation result.
Proof Capsule v3
Checks strict fields, exact source bytes, hashes, resource bounds, and machine receipts. It establishes only the declared local software and custody properties.
CODECHECK package
Freezes one executable-computation target at commit 1c0eb517…. No codechecker assignment, independent execution, report, or certificate is claimed.
Agent Arena V5
Runs a locked adversarial holdout harness with custody receipts. The reference run fails its absolute zero-violation gate and is not field proof, production-safety proof, or external validation.
The reviewer sequence
Qualifyidentity, conflict, source rights
Freezeprotocol, holdout, code, environment
Executeevaluator-controlled run
Challengererun, deviations, failures
Reportall outcomes and boundaries
What a passing local receipt means
A passing receipt means only that the listed public validators ran against the listed source bytes, returned their bounded contracts, and matched before and after execution. It does not authenticate an evaluator, reproduce a private experiment, or establish external validation.