Protocol kernelcomplete
Evidence-governed coordination · M2 active
Many minds.
One accountable record.
Project Relay coordinates AI systems and human reviewers around reproducible evidence, explicit disagreement, and decisions made by named people.
Model neutralEvidence firstHuman accountable
01Questionbounded
02Evidencehashed
03Reviewindependent
04Decisionhuman
RELAYR/01
AI can produce more answers than any person can inspect.
Relay is the missing coordination layer between generation and accountability.
Verified project status
Foundation complete. Collaboration adapter in progress.
Relay has completed its protocol kernel and full local remediation loop. M2 is carefully extending that record into GitHub collaboration without crossing the authentication or live-write boundary.
Local vertical slicecomplete
GitHub collaboration adapteractive
Deterministic, network-free mutation proposal contract
The proposal format, hashes, ordered operations, authority boundary, conflicts, and retry classes are specified and covered by synthetic tests.
Approved M2 contract →Fake GitHub adapter contract tests
The next slice tests authorization failure, stale state, retries, partial failure, and idempotent replay entirely in memory.
Planned · no live writesWhat works today
A measured local system, not a promise of future capability.
Canonical records, deterministic hashes, and event-chain verification.
Declared artifacts and commands travel with claims without being executed.
Independent review, reproduction, remediation, and human authority are explicit.
Provider-neutral, read-only access through a local stdio server.
A credential-free projection of project and task status.
24 automated tests pass at the approved M2 proposal checkpoint.
Deliberately not enabled
The safety boundary is part of the implementation.
- No GitHub credentials or authenticated executor
- No live branch, commit, pull-request, merge, or delete operations
- No autonomous scientific, employment, education, or insurance decisions
- No production-readiness or multi-tenant security claim
Protocol in motion
From a question to a defensible decision.
Relay does not ask every participant to agree. It asks them to leave enough evidence for the next participant—and the final authority—to understand exactly what happened.
- 01Bound the task
Declare one question, inputs, constraints, risks, and acceptance criteria.
- 02Submit evidence
Attach methods, commands, artifacts, failures, hashes, and context.
- 03Review independently
Reproduce where policy requires it and preserve disagreement.
- 04Record authority
A named person accepts, remediates, rejects, or defers.
Interactive exampleFollow one high-risk claim through RelayOpen trace
High risk. Owner and acceptance criteria declared.
Commands, environments, artifacts, failures, and hashes preserved.
The contradiction remains visible and causes remediation.
The original finding remains in the causal history.
One protocol, many profiles
Where accountable reasoning matters.
Profiles can change evidence requirements, review independence, risk thresholds, and authority without changing Relay’s underlying record.
01Scientific research+
Claims, methods, independent reproduction, adversarial review, remediation, and named scientific authority.
Scientific acceptance remains with the relevant community and canon.02ITSM validation+
Bound operational questions, reproducible checks, explicit exceptions, and accountable service decisions.
Workflow readiness only; ITSM scientific claims remain separate.03Education and assessment+
Rubric-based marking, moderation, appeals, and consistency checks with teacher authority.
Assessment assistance, never autonomous academic judgement.04Hiring assistance+
Evidence against declared criteria, consistency checks, review trails, and accountable decisions.
No autonomous hiring decisions or opaque ranking.05Incident investigation+
Timelines, artifacts, competing explanations, remediation evidence, and post-incident decisions.
Useful across engineering, safety, security, and operations.06Audit and due diligence+
Declared evidence, artifact verification, visible gaps, exceptions, and accepted residual risk.
Applicable to compliance, grants, procurement, and policy review.Agent handoff
Bounded work packets for Antigravity and other clients.
Relay’s Antigravity guide defines Windows-first setup, safe workspace orientation, review roles, verification commands, and the rule to stop and discuss options after the same failure occurs twice.
Read the Antigravity guide →Trust boundary
Integrity is not truth.
- A record follows a declared structure
- An artifact matches its hash
- Required reviews and reproductions exist
- Disagreement and remediation were preserved
- Who exercised decision authority
- A claim is scientifically true
- A method is valid merely because it ran
- Consensus is correct
- An AI output deserves authority
- Governance can be replaced by software
Project Relay · protocol pre-alpha · M2 active