plainspeak Download

Get a second opinion with the audit

The audit asks two other AI models to read your rewrite and quote anything that still sounds generated.

Before you start

Turn on the audit

  1. Tick Audit with two other models below the Rewrite button.
  2. Select Rewrite.

When the rewrite finishes, the Audit panel on the right fills in. Each audit costs about two cents. PlainSpeak only picks models that cost at most $1 per million tokens of input.

How the verdict works

The two audit models are the judges. The model that wrote the rewrite does not vote: being told to remove every tell is not evidence that it did. Only a reply in the expected format counts.

The verdict sits at the top of the panel as a card: a badge with the verdict, and one plain sentence saying which model found what and which did not answer, such as "FAIL: 1 hit from z-ai/glm-5. mistralai/mistral-large didn't answer (rate limited)." A verdict brought back from History, which keeps no card for each model, says how many models replied instead, such as "2 of 2 responding auditors clean · 2 of 2 expected replied".

VerdictWhat it means
PASSBoth audit models replied, and neither found anything.
WARNBoth replied; one flagged something.
FAILEvery model that replied flagged something.
INCOMPLETEOne model replied clean, and the other errored, replied off-format or did not reply. One reply is not a full audit: select Retry.
INCONCLUSIVENo model replied in the expected format.
ERRORNeither model replied.

A verdict is two models' judgment of the wording. Models share blind spots, so their agreement is not proof of who wrote a piece or that its facts are right.

Each audit model gets its own card in the panel. It shows CLEAN, or the phrases it quoted, up to five, each as an AI-ism, jargon, cliché or filler. The audit models judge wording only.

How the audit models are chosen

Act on what it found

ButtonWhat it does
Fix N selectedRewrites only the quoted phrases you left ticked and leaves everything else alone.
RetryShown once: on the first card of a model that errored or replied off-format, or on the verdict card when no card carries it and the verdict is INCOMPLETE, INCONCLUSIVE or ERROR. Runs the whole audit again with the same two models. If it fails again, the card says why and that it tried again.
Audit againAsks the same two models to read the current copy again.
Audit with new modelsPicks two different models from other families.

Audit again and Audit with new models are a row of plain links at the bottom of the panel: they only measure again, while Fix N selected changes the copy.

Each finding has a tick box, ticked to start with. Untick a phrase you want to keep, and when you next select Fix N selected, Retry, Audit again or Audit with new models it is kept on purpose: the panel lists it under Kept on purpose, the audit models are told to leave it alone, and the next audit does not raise it. Changed your mind? Select Raise again beside it, and the next audit is free to raise it. A phrase both models quoted has one box, in the first card that quotes it. One fix takes up to 10 findings; any more are listed without a box and wait for the next round.

After a fix, the panel says how many of the earlier findings are gone. A phrase flagged three rounds running is marked Author's call and stays that way until you pick new models or select Raise again beside it, and after three fix rounds a note suggests editing the rest by hand: what is left is often a matter of taste. When both audit models reply and find nothing for two complete rounds in a row, the panel says the wording reads clean to them. A round where only one model replied does not count toward that, and neither do rounds saved in History before PlainSpeak recorded how many models replied: the panel marks those coverage unavailable, and a fresh audit is needed.

When you edit after an audit

Editing the rewrite by hand does not re-run the audit, which costs money each time. The verdict stays, and the panel opens with This audit read the copy before your edits. Undo the edits and the note goes; select Audit again to have the same two models read what is there now. Viewing an older version from its chip is not an edit: the panel names the version the audit read instead, such as This audit read v2.

If the audit does not run