01 · The problem
Most AI code review is confident right up until you ask what it actually read.
You can't attach it to a proposal. You can't hand it to a client. You can't tell which findings are grounded in your code and which are the model being agreeable. Ask it what it looked at and the answer dissolves.
POVV produces a verdict with its evidence attached: the files that entered evidence, the commit they came from, the coverage that implies, and a cryptographic seal anyone can check without trusting us.
- Evidence base disclosed
- Verify without trusting us
- Free · no signup
One verdict, free, no account and no card. The run is reported here as it happens — that is the trailer; the verdict itself arrives by letter, roughly three minutes from now. One free public run per address and per browser.
02 · The difference
An opinion tells you what a model thinks. A verdict tells you what it saw.
A model opinion
- Scope is whatever fit in the context window. It won't say which files.
- Findings read plausibly. Nothing distinguishes a quote from a recollection.
- Agrees with you, because it was trained to.
- Re-run it tomorrow and get a different answer with equal confidence.
- Nothing to hand a client. Nobody else can check it.
A POVV verdict
- Names its evidence base: files read against files auditable, at a specific commit.
- Every finding quotes your file verbatim, re-checked against the source before it can seal. Quotes that don't match are struck and counted.
- Built to disagree — the benches are scored against each other, not against your feelings.
- Signed over its contents. Change one character and the signature breaks.
- A permanent URL and a README badge that a client, an acquirer or a reviewer can verify without asking you or us.
03 · How it works
Map before judgment. Quote before claim. Seal what survives.
Map
Before anything is graded, POVV builds a graph of the repository's trust boundaries — routes, policies, secrets, datastores and what gates what — so findings are anchored to structure rather than to whichever file happened to be read first.
Scope
What entered evidence is disclosed before the score: how many files were read against how many are auditable, the commit, the byte count. On a large repository POVV reads a bounded slice and says so.
Contest
Two opposed benches examine the same evidence and are forced to reconcile. Every finding must attach a verbatim quote from your file, which is machine-checked against the source before it can enter the record.
Seal
The verdict and its evidence base are hashed and signed with Ed25519. The public key is published, so the seal can be checked offline, by anyone, years later, with POVV nowhere in the room.
05 · What POVV does not do
The limits, stated by us, before you find them.
- It does not read your entire repository. On anything large, POVV reads a bounded slice and publishes the coverage ratio with the verdict. A verdict over 19% of a repository says 19% on its face.
- It does not report a control as absent because it didn't look. Something POVV did not read is reported as unverified. Absence of evidence is never written up as evidence of absence.
- It is not a security certification — not a compliance attestation, not a penetration test, and not legal or engineering advice. It is an automated adversarial opinion over a disclosed evidence base.
- A high score is not a warranty and a low score is not proof of a defect. It is a risk index. Don't make an acquisition or deployment decision on this document alone.
- It does not publish a verdict about a repository unless its owner ran it. POVV does not audit strangers to get their attention, and does not send findings to anyone who didn't ask.
- It does not keep your code. Source is held in memory for the duration of the audit and flushed after the seal. It is never used for training.
Point it at a public repository and read what comes back. One verdict, free, no account and no card — and you'll know inside three minutes whether this is an instrument or another chatbot.