Conversation & work item
You tell Claude: “Use CW.” The request becomes a durable Conversation and a claimed work item with an operator-approved scope. What is allowed, and what is excluded, is recorded before anything runs.
Claude works. ChatGPT and Codex challenge. You stay in command.
ClearWright is the Council of AI, and you are Chairman of the Board.
ClearWright is a local control plane for AI-assisted work. It coordinates independent review, records disagreement, enforces operator-defined authority, and refuses completion until required verification passes.
The operator commands.
Independent reviewers challenge.
Verification decides DONE.
What Just Happened?
In a recent governed session, a single operator request initiated a two-model review council. ChatGPT and Codex independently reviewed the plan, Claude reconciled their findings with recorded evidence, and the work proceeded only inside the scope the operator had approved.
The finished result then went back to ChatGPT and Codex for final verification, and ClearWright recorded DONE. The session’s audit chain is logged harness invocations end to end, with no manual-relay step anywhere in it.
And When It Fails
In another recorded session, verification did not pass. ClearWright refused to record DONE and said so. That refusal is the product working, not the product failing.
How ClearWright Works
In the open reference implementation, one request runs a full governed pipeline: a durable conversation, a claimed work item with approved scope, a Claude plan, independent ChatGPT and Codex review, evidence-backed reconciliation, a deterministic agreement threshold, execution inside scope, final verification, and a recorded outcome, DONE or operator required.
Independent ChatGPT + Codex review
Claude drafts a plan. ChatGPT and Codex each review it independently as real recorded calls and return structured findings: blockers, required changes, dissent, and questions. A reviewer that cannot be reached is recorded as unavailable, never simulated.
Evidence-backed reconciliation
Claude accepts valid criticism or rejects it with recorded evidence. Every blocking finding must be resolved by explicit reference. The agreement threshold is deterministic: while blockers remain, work does not proceed.
Execution inside approved scope
Only after agreement does work begin, and only inside the approved scope. Progress is posted to the record. Glitches go to an incident consultation instead of silent retries.
Final ChatGPT + Codex verification
The actual result, with the real diff and outputs, goes back to the reviewers in a verification council. Verification examines what was done, not what was promised.
DONE, or the honest state
ClearWright records DONE only when required verification has passed. Otherwise it stops and says so: the work item is marked verification incomplete and the operator decides. Terminal outcomes get a canonical, harness-generated summary.
1. You tell Claude: "Use CW."
2. Claude records the task and approved scope.
3. ChatGPT and Codex independently review the plan.
4. Claude accepts valid criticism or rejects it with evidence.
5. Work proceeds only inside the approved scope.
6. ChatGPT and Codex verify the actual result.
7. ClearWright records DONE, or stops honestly.The Review Council
Not a vote. Structured review: independent reviewers, recorded dissent, evidence-backed reconciliation, and a deterministic agreement threshold under operator authority. ChatGPT and Codex hold the first council seats; more agents will be incorporated in the future.
Real reviews, never simulated
In the recorded councils behind this page, every ChatGPT and Codex review was a real, recorded adapter call: ChatGPT over the OpenAI API, Codex through the Codex CLI, with per-review telemetry in the audit record. A reviewer that cannot be reached is recorded as unavailable, and the implementation is open source, so you can inspect the dispatch path yourself.
Structured findings
Reviewers return structured verdicts: blocking findings, required changes, non-blocking findings, disagreements, assumptions, and questions, not a thumbs-up emoji.
Recorded dissent, evidence-backed rejection
Claude may reject a reviewer finding only with recorded evidence: the record schema requires a reason and a non-empty evidence list for every rejection. Disagreement is kept in the record, not smoothed over.
Deterministic agreement
Two to five substantive rounds, bounded by the implementation. Agreement is a deterministic threshold, not a vibe: no agreement while any blocking finding is unresolved.
Proof, Not Promises
Claims here come from ClearWright’s own durable records: canonical summaries, council files, reviewer telemetry, and invocation logs from governed alpha sessions. Nothing below is hypothetical.
It refused to say DONE
In a recorded governed session, verification did not pass, and ClearWright refused to record DONE. The work item was marked verification incomplete and handed back to the operator, with the refusal itself in the audit record.
Fail-closed by implementation
Completion is fail-closed in the published implementation: the complete command refuses DONE whenever a bound verification council has not reached agreement, or required verification was never run. Absence never bypasses it.
Governed work has completed cleanly
Recorded work items have reached DONE with verification satisfied: plan and verification councils both meeting the agreement threshold, with zero unresolved findings.
Corrections ran in both directions
Reviewers have blocked unsupported claims by the implementing model, forcing them to be withdrawn or evidence-bound before agreement. And reviewer findings have been rejected, on the record, with the required evidence attached.
No manual relay
In a recorded end-to-end session, the audit chain runs unbroken from review dispatch through reconciliation, verification, and completion, and every step is a logged harness invocation. The recorded chain contains no manual-relay step.
This page is an example
This page was rewritten under ClearWright governance: planned, independently reviewed by ChatGPT and Codex across multiple review rounds, and reconciled with recorded evidence, and under ClearWright's fail-closed rule it does not ship as done until verification passes.
Where It Fits
ClearWright is not a replacement for tool-access protocols, agent-to-agent communication protocols, or workflow-orchestration frameworks. Those systems are complementary.
They give agents a way to reach tools, exchange messages, and run execution graphs. ClearWright answers a different question: who is allowed to act, is the lane clear, what clearance was granted, who can override, and what stays in the audit trail.
In One Line
ClearWright is not the tool bus, the message bus, or the workflow graph. It is the review, clearance, and audit layer that helps those systems operate under authority and discipline.
Open Source, Early Alpha
ClearWright is a local, single-operator, early-alpha proof of concept. It is not a multi-user platform, not a certified command system, and not production-ready for regulated or safety-critical deployment. Planning for its first self-improvement capability is complete and has passed a two-reviewer plan gate; no implementation authority has been granted and no such code exists yet.
Beyond that governed two-model council, review and advising for this work have also been exercised through ChatGPT (API and WebUI), Codex, Claude, and Grok Pro WebUI. Grok currently participates via the WebUI as an out-of-band production advisor; full API integration into the governed review lane is pending.
The implementation is public: the protocol, the council engine, the Use CW workflow with its Claude Desktop and Claude Code skill, and the operator tooling are on GitHub. Inspect how reviews are dispatched, how reconciliation is validated, and how completion is refused, in the code, not in a whitepaper.
Inspect It
github.com/LimitedEnergyX/ClearWright — source, docs, and the ClearWright Protocol specification.
Reviewed Before It Acts. Verified Before It’s Done.
Human-commanded AI work for operators who understand chain of command, delegated authority, separation of duties, and audit.