One review, end to end — written for an engineering leader deciding whether independent AI review is worth process weight. This recounts a real pre-release review of Impasse's own code; the resulting commit and CHANGELOG entry are public, and the limitations of this kind of evidence are stated at the end.
When an AI system tells you "this work was independently reviewed," the value of that sentence depends entirely on whether the claimed independence is real. Impasse's job is to attach an honest independence label to every review. The failure that matters is quiet: a system that overstates independence hands you correlated opinions dressed as a second set of eyes — and you make the decision with more confidence than the evidence supports.
In July 2026 the maintainer reworked how Impasse computes that label, so it would hold up no matter which AI agent was driving the process. The pre-release implementation carried a deliberate compatibility choice: when the driving agent couldn't be identified, keep the old, more favorable labels. The author had reasons — don't change behavior for existing setups, and software can't identify a driver that refuses to identify itself. Tests were written that locked the choice in.
Before the change shipped, it went through Impasse's own protocol: a reviewer from a rival provider (OpenAI's Codex, reviewing work produced with Anthropic's Claude — the authorship and fresh-context details are the maintainer's account; the published records identify only the reviewer backend) read the change cold. Its top finding, paraphrased:
The fallback grants a positive independence claim in exactly the case this change exists to fix. An agent that fails to identify itself — by omission, misspelling, or misconfiguration — gets "independent cross-provider review" as a label, with no warning. The new tests don't protect that boundary; they enshrine the hole. This is fail-open at the most important gate.
The author had framed the fallback as backward compatibility. The reviewer reframed it as a security default — and under that frame the design was indefensible. Same code, same facts; a different reading.
Impasse's record format has one unusual rule: a saved reconciliation cannot record a rejected
finding without at least one piece of contradicting evidence — "I meant to do that" doesn't
validate. That rule constrains the record, not the person: the workflow is directed by the
host following the protocol, and a host that ignores it can. What the rule did here was make the
honest paths explicit: produce evidence against the finding, or concede it, or escalate it as a
dispute. The author checked the finding against the code (accurate), against the project's own
precedent (an unidentified environment already failed safe — the inconsistency was real), and
conceded. The fallback was reversed the same day: an unidentified driver now gets
"undetermined," never a positive claim, and the tests assert the fail-safe instead of the hole.
The CHANGELOG entry and the commit are public, and — as a deliberate exception to Impasse's
records-stay-local policy, safe here because the reviewed artifact was this public repository's
own code — the reviewer's full response and the reconciliation for this case are published
verbatim in evidence/host-independence-review/. Two
further author decisions fell to the same process that week (summarized in the CHANGELOG).
- The class of error is the interesting part. This wasn't a bug the author lacked skill to see — it was a decision the author had reasons for. Review by the same assistant that helped make a decision tends to inherit the reasoning behind it; a fresh reviewer, plus a rule that dismissals need evidence, gave the objection somewhere to land.
- One case is an anecdote, not a controlled result. Cross-provider diversity is Impasse's design hypothesis — different training plausibly means partially different blind spots — but this record can't isolate that from fresh context or plain variation between runs. It is offered as a compatible example, not causal proof. Agreement between two models, likewise, is evidence and never proof.
- The human stays the decider. Genuine judgment calls escalate to the operator rather than being settled between models; in this period one operator ruling reversed the operator's own earlier spec — on the reviewer's evidence.
- The costs are real. The full cross-provider tier needs access to two AI providers (a same-provider fallback exists, with a disclosed weaker guarantee); reviews take minutes and cost tokens; and this account is the author reporting on his own tool.
Impasse is the open (MIT) implementation of the workflow described in AI's Second Opinion: When Rival Models Disagree. Moving Average Inc. advises engineering organizations on exactly this class of AI-process design.