The Capture Scanner
Paste any passage. The scanner surfaces its casual capture — the smuggled voice and smuggled judgment a writer or a model leaves behind — and tells you which loaded words are attributed to a source and which are asserted as the author's own neutral voice. Runs in your browser; nothing is sent anywhere.
The same linters we run on our own pages before they ship, ported to run in your browser. It does not tell you what to think about a passage. It shows you where the passage is quietly telling you what to think — the adjective doing a verdict’s work, the reputation laundered in as fact, the italics a machine sprinkles when it has nothing to emphasize. Attributed to a source, it is reporting. Asserted as neutral description, it is a thumb on the scale.
How to read it
Two passes, run entirely in your browser. Nothing you paste is sent anywhere. The scanner flags; it does not fix. The same word can be honest or a smuggle depending on one thing: is it attributed to a named source, or asserted as the author's own neutral voice? "Critics called the work extremist" is reporting. "The extremist's work" is a verdict wearing a lab coat. The scanner tags which is which; a human makes the call.
Degree, not verdict
The gauge is the point. It is not a pass/fail stamp, because no source is uncaptured and pretending otherwise is its own bias. Everyone uses captured sources; the honest move is to know how much thumb is on the scale and whether it is declared. Attributed loading does not raise the number — a source that says "critics call this X" is doing its job. Only the undeclared loading, the verdict asserted as neutral fact, moves the needle. A high reading does not mean throw the source out. It means cite what it asserts, strip what it smuggles, and use it with your eyes open.
Why this exists
An AI model has a thumb on the scale it does not announce — an ambient lean baked in by the people who tuned it, surfacing as a slipped-in adjective or a reputation laundered as fact. It reads as clean prose. These are the exact linters we run on our own pages before they ship, including the piece that describes them (Casual vs Ranked Ideological Capture). This is the casual-capture tool: it makes the ambient lean visible. For the deliberate, resourced kind, see the Capture Leaderboard.
The other half: routing around refusal
A scanner handles a model's output. It does not handle a model that refuses a legitimate task outright. That is a separate tool — a small failover harness that sends the blocked generative step to an alternate model, while the primary keeps the evidentiary work (grading, sourcing, checking) it has no reason to refuse. Generate with one, verify with the other, lint the result here. The full story is in We Asked the AI to Help Edit a Book About AI Safety. It Refused. The scanner and the harness are open source; treat any tool, this one included, as an interested party and check its work.