Independent review for work made by AI
CamusTrust the work, not the model that made it.
One AI does the work. Another, from a different company, checks it against what you said must be true. Tests and sources settle what they can. You make the calls they should not.
The problem
A confident answer is not the same as a trustworthy one.
The model that made the work shares its own assumptions, omissions, and incentives. Asking it to check itself can produce a more polished version of the same blind spot.
The maker can miss its own mistake.
Stronger models make better work, but confidence is still not evidence.
Self-review shares the blind spot.
The same model family tends to defend the choices and style it already produced.
Review can detach from the final result.
A clean verdict is meaningless if the artifact changes after the reviewer saw it.
Camus keeps the artifact, the evidence, the independent verdict, and the human decision bound together.
A live example
The first model sounded right. The second model checked.
In a real research run, Sonnet drafted a plausible strategy from Hivemind material. GPT-5.4 found claims the source did not support, including a raw search score rewritten as “Relevance: 76%” after the contract explicitly forbade interpreting score semantics.
Sonnet produced the initial strategy using the frozen research context.
GPT-5.4 blocked unsupported interpretations instead of rewarding persuasive prose.
One decision reached the human: authorize one more repair round.
The repaired result passed verification and earned an independent clean audit.
The result did not merely say “done.”
standing verified
execution completed
verification passed
audit independent_clean
publication not_published
executor anthropic:sonnet
auditor openai:gpt-5.4
artifact 59ee19193b8b
receipt 0b8960d2e40dThe full receipt also preserves the acceptance contract, exact model identities, human decisions, checks, findings, and evidence bundle.
How Camus works
Four parts. One honest result.
Camus does not need to control every agent. It controls what the work must satisfy, what was actually checked, and who gets to decide.
Set the contract
State the goal, what must be true, the permitted knowledge, and the budget.
Let the executor work
Use a single model or its native agents. Camus treats the whole system as the maker.
Challenge it independently
A different model audits the artifact. Tests and captured sources arbitrate where possible.
Decide and seal
Ambiguity goes to you. Every verdict binds to the exact artifact it certified.
Human in the loop
You are not removed from the loop. You are removed from babysitting it.
Where Camus goes next
Better models make Camus more useful, not less.
Frontier models are becoming excellent orchestrators. Camus will let them own the workers while it compares what they produce under one contract and keeps any model from awarding itself trusted standing.
Available now
Independent trust
One executor, one cross-vendor auditor, deterministic checks, human checkpoints, and a sealed evidence pack for the exact result.
Planned: Compare & Learn
Evidence-backed improvement
Run the same goal and frozen knowledge across different models, efforts, or orchestration strategies. Keep failed arms. Compare blindly. Learn locally, and route only when enough evidence exists.
The quality floor comes first. Only then does Camus optimize tokens, time, and cost. Cheap failure never wins.
Two ways in
Built for work you stake your name on.
The trust protocol is the same. The interface meets you where the work lives.
For research and marketing
Loop Studio
Write a memo, investigate competitors, or turn Hivemind knowledge into a grounded deliverable. Use plain language, inspect every objection, and step in only for real judgment calls.
- browser interface, no JSON noise
- acceptance contract in your own words
- Hivemind grounding through your Claude MCP
- downloadable evidence pack
For developers
Camus CLI
Give an agent a real repository without giving up custody. Work stays isolated, reviews bind to exact commits, tests have the final word, and only proven changes reach your branch.
- isolated worktrees and controlled merge
- cross-vendor code review
- HEAD-bound verification
- crash-safe resume and named human halts
Run it locally
Your machine keeps custody.
Orchestration and receipts stay on your machine. Model and Hivemind requests go to the services you already authenticate. Camus does not bundle or proxy your subscriptions.
$ npm i -g camus-cli
$ camus install
$ camus check
> /camus-feat { feat: "Harden input boundaries", tasks: [...] }
✓ executor finished in an isolated worktree
✗ independent review found a missing guard
✓ repair passed 163 tests
✓ verified receipt bound to commit a1f9c2e
The principle
No intelligence should be the only judge of its own work.
Models will change. The need for independent judgment, inspectable evidence, and a meaningful human decision will not.