Release
Now you can argue with the critique
Push back on any finding and Thinkr rules on it — upheld or withdrawn. It can only withdraw a finding by quoting the part of your PRD that proves it wrong.
Until today a critique finding gave you two moves: fix it, or delete it. Neither of them is "you're wrong."
That's a real gap, because sometimes the finding is wrong — it missed a section, or it's flagging something your team decided on purpose. And deleting it left no trace, so a rejected finding looked exactly like a fixed one. The review stopped being auditable at the exact moment it mattered most.
What shipped
Push back on a finding. Open the Findings tab on any PRD, and every finding now has a Push back action next to Resolve and Delete. Say why you disagree. Thinkr re-reads your PRD as it stands right now and rules: finding upheld or withdrawn by Thinkr. The whole exchange stays on the finding.
It can only back down on evidence. This is the part worth knowing. Thinkr's default is to hold its position, and it's allowed to concede for exactly one reason: your PRD already answers the finding. To do that it has to quote the passage — character for character. If it decides to concede but can't produce a real quote from your document, the concession is rejected in code and the finding stands.
That rule exists because of a specific failure mode. A model asked "the user disagrees, were you wrong?" will fold, every time, because being agreeable is what these things are trained to do. A critique that folds under pressure is worthless — it's the same flattery problem that lets a smooth-reading draft ship a blocker. So the verdict isn't left to the model's judgement. Thinkr can't be persuaded, only shown.
A good argument can still win something. When Thinkr upholds a finding it can raise or lower its severity based on what you said. You'll see a Severity → badge on the ruling. Not every argument has to be total to be worth making.
Structural findings answer instantly. Generic user-story roles, missing MoSCoW tags, untestable acceptance criteria, event names that aren't noun_verb — those come from deterministic checks, not the model. Push back on one and Thinkr just re-runs the check against your current draft. Still fires, it stands. Doesn't, it's withdrawn. Instant, free, and identical every time.
Withdrawn is not resolved. A conceded finding is labelled Withdrawn by Thinkr, not Resolved. A withdrawal isn't a fix, and the record shouldn't blur the two.
You get two pushes per finding. If two rounds haven't settled it, a third won't.
What's next
The disputes you file are the most honest signal we have about where the critique is over-flagging. We're watching what gets withdrawn and why, and feeding it back into the passes that produce those findings — so the eleven passes get sharper at the specific things you keep having to argue with.