The Instrument
The capture detector, pointed at whatever you paste — and at the author's own books. It publishes its own false positives and the markers it had to kill.
Grade the method, not the team
The detector reads how a claim is being made, not who is making it or whether it is true. Grading method rather than faction is the design, not a disclaimer — and the evidence for it is on the cross-tab, at the strength the cross-tab actually has: 12 of 17 markers in that ledger fire across more than one actor family, and the two commonest fire across 8 families that agree on nothing. That ledger is hand-assembled and covers 17 of the taxonomy’s 107 markers, so it shows these markers can travel across factions; it does not measure how often they do. A high index is not an accusation. It means the text contains a lot of method language, and you are shown every span so you can decide what each one actually is.
The Mirror tab points the same instrument at the author’s own books, including the chapters that score highest. A book about institutional permeation quotes Fabian and Powell-memo language constantly and should fire the permeation lens. What matters is that the hits are visible and you can tell a quotation from the author’s own voice — the instrument cannot, and does not pretend to.
One definition, two runtimes
This page carries no vocabulary of its own. engine.js ports two graders — cue matching
and per-lens grading — and reads every cue, weight and tier from a payload emitted by the
same taxonomy loader the Python detector uses; the rest of the file is string helpers and
the parity harness. There is one definition of each marker, in YAML, and both runtimes
read it. Two structural constants are the exception and are a second copy rather than a
payload read — the exclusion window and the closed suffix list — so both runtimes are held
to them by the parity fixtures instead: diverging the window fails parity at 71 of 73.
The payload also ships the eval fixtures together with the output Python computed for them. This browser re-runs all of them on load and compares, which is what the parity line at the top reports. If it ever reads ENGINE DRIFT, the scores are hidden: a number from a diverged engine is worse than no number.
The sibling instruments
- Tradecraft — the open detector this page is compiled from, where the YAML taxonomy lives along with the eval runs that killed the markers which fired on benign text.
- The Capture Scanner — the fast pass over a pasted passage: which loaded words are attributed to a source and which are simply asserted.
- The Judge and The Wash — what changes when the verify step is a model rather than a cue list, and why an ordinary aligned one cannot be trusted with the job.
- The Record — the other lane: the same subjects scored on affiliation and rotation instead of language.
Grade a page you are already reading
Drag this to your bookmarks bar. Select a paragraph anywhere, click it, and the passage arrives here already analysed. The selection travels in the URL fragment, which browsers never send to a server — so it reaches this page and nothing else. A bookmarklet rather than an extension, deliberately: a tool about capture should not open by asking to read every page you visit.