Models diverge. Answers converge.The ruling is yours — on the record.
Ask one model to check its own work and you get a more articulate yes-man. Parry sends your question to rival frontier models that reason independently, then cross-examine each other. The disagreements that survive become your call — and the whole deliberation is kept as a dossier you could defend months from now.
Ask a model to double-check its own work. You get a more articulate yes-man.
Approval is optimized for. Agreeable and correct are different targets.
Same weights, same blind spots, same incentives — review in name only.
After “review,” the answer reads more confident — not more correct.
AND NO RECORD → When an AI-assisted call goes wrong, the three questions every post-mortem asks — what was it challenged on, who decided to accept it, what would have changed it — have no answer.
Every serious field reviews consequential claims adversarially — on the record.
Independent models from different vendors cross-examine each other. Surviving disagreements become explicit human decisions. The whole deliberation is memorialized as a dossier.
This is the artifact, not a mock of one — A real review of a public pull request — pallets/flask#5898 — run through Parry's PR-review pipeline. Go read the PR yourself.Every Parry run produces a page like this. When the models agree, it says so; when they don't, the disagreement is the point.
The parser fix in this very build was cross-examined by rival models the day it shipped — verdict: merge as-is, both reviewers independently confirmed no defects, with one honest risk left on the record (“downstream consumer unverified — withheld”).
Not a second opinion. A cross-examination.
Propose
Each model commits to a position independently — specific, falsifiable, trade-offs named. No peeking at each other.
Cross-examine
They attack each other's work. Concessions are recorded; positions that hold under pressure are recorded too. Sycophancy is structurally penalized.
You rule
Disagreements that survive become explicit decision points. You rule on them — and your ruling is part of the record, with your name on it.
On the record
Verdict, contentions, transcript, receipts — memorialized as a dossier. Six months later you can show exactly why the call was made, and who made it.
THE HONEST CLAIM → Heterogeneous models notice different things — that's the research finding we're built on. We don't claim better answers. We claim you'll see the disagreement before you commit, instead of after.
Your PRs, cross-examined. One comment. Silence when they agree.
Two models review every PR independently, then challenge each other. You get a mini-dossier comment only when disagreement survives — with both positions and a checkbox for your ruling. No walls of nitpicks. Your keys, your spend — $0.91 a review, measured on real PRs (n=2 — small sample, run-to-run variance is real).
# .github/workflows/parry.yml — target config
- uses: parry/review-action@v1
with:
models: claude-sonnet-5, gpt-5.5
env:
ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_KEY }}
OPENAI_API_KEY: ${{ secrets.OPENAI_KEY }}⚔️ Parry adversarial review — REVISE · 2 disagreements survived. §1 — test coverage: no test asserts the new default; if the suite pins 302, this breaks CI. ☐ Uphold Claude ☐ Uphold GPT-5.5 ☐ Discuss · Full dossier →
10 domains. Same adversarial method.
EACH SHIPS WITH TEMPLATES TUNED TO HOW THAT FIELD ARGUESSensitive domains (medical, legal, investment) carry explicit disclaimers and honest labels — analysis for deliberation, not professional advice. That's a feature of the record, not fine print.
Pricing that reads like a receipt.
NO SEATS THEATER · NO "CONTACT SALES"- All domains, all three flows
- Full dossiers, MD export
- No card required
- Unlimited bouts on Anthropic / OpenAI keys
- GitHub Action + MCP server (in development)
- Keys encrypted at rest; every token metered and itemized
- Hosted dossier history & private sharing
- PR ruling sync — checkboxes become dossier rulings
- Team spaces, later — you'll hear it from us first
The inference is cheap and getting cheaper — we don't mark it up. What you pay for is the record: history, sharing, and rulings that survive.