Method
How the audit works
Every position in a briefing, and every text run through the Audit, goes through the same engine. This page is the method: what it looks for, what the labels mean, and what it deliberately won't do.
What the engine analyses
Eleven lenses, each grounded in a real tradition of logic and argument analysis: from Toulmin's argument model to key-term, referential, and testability scrutiny. Six of the headline outputs:
Argument structure
Toulmin decomposition: claim, grounds, warrant, and the weakest link. The skeleton of the argument, stripped of rhetoric.
26 reasoning patterns
Post Hoc to Motte-and-Bailey, Selection Bias to Genetic Fallacy, each with the verbatim quote that exemplifies it and a severity.
Counterarguments
The two or three strongest objections a thoughtful opponent would deploy, each steelmanned, each with its own Toulmin structure.
Citation audit
Every cited URL is fetched and checked: does the source actually support the claim it is cited for?
Frameworks made visible
The ethical, epistemic, and political framework the argument operates within, surfaced so you can defend or replace it.
Evidence-weighted likelihood
For empirical claims, we search 200M+ academic papers and synthesise where the scientific consensus actually sits.
Two rules the findings live by
Every finding quotes its evidence, verbatim.
A finding must point at exact words in the text. Quotes that don't appear verbatim in the source are dropped before anything is shown: a finding you can't check against the text isn't a finding.
Every finding declares its kind of ground.
Logic: verifiable in the text's own structure. Judgment call: depends on a reading a careful reader could contest. Factual: checkable against external sources. No fake percentage scores; the label tells you how to weigh the finding.
How a briefing is built
A briefing maps the argument space of one question: it does not sample the media coverage. See a live example: Does raising the minimum wage cost jobs? →
One contested question
Each briefing takes a single question that serious people genuinely disagree about, stated plainly and kept evergreen.
A per-question spectrum
Sources are mapped on an axis named for what actually divides them: "competitive market vs monopsony power", not left vs right. The question chooses the axis.
Positions, each audited
The strongest form of each position, quoted verbatim from a real source, followed by a structural audit card: the named reasoning pattern, where it sits, and why it matters.
The shared assumption
The premise both sides quietly stand on. No other outlet tells you what the disagreement agrees about.
The editor's view, walled off
A position, taken openly, labelled as opinion, and followed by the devil's advocate: the strongest reason to doubt it.
Other takes, audited
Curated external takes each carry their own structural annotation. Where none are surveyed, the briefing says so.
How positions are placed on the map
Only sources that argue a stance are plotted: every plotted point carries an audit card. Sources cited as factual input (a measurement, a dataset, a technical result) are listed as evidence beneath the map, never plotted: a paper that reports a finding does not thereby take a side.
The stance bands
- −2 argues strongly for the left-hand pole of the question's axis
- −1 leans toward the left-hand pole
- 0 genuinely mixed or split
- +1 leans toward the right-hand pole
- +2 argues strongly for the right-hand pole
Five bands, not a 0–100 number: a two-digit coordinate would claim a precision the judgment doesn't have. The axis itself is named per question: what actually divides the positions, not left vs right.
The confidence tag (drawn as marker width)
- high we've read three or more pieces of this source's argument on this question (tight marker)
- med one or two pieces (medium marker)
- low a single article; provisional (wide bracket, honestly "somewhere around here")
Scope, stated once: placements describe a source's argument on this question, not the outlet in general; a briefing maps the argument space, it does not sample the media coverage.
What it won't do
It won't adjudicate claims.
The engine analyses argument structure, not the factual accuracy of specific assertions, though the citation and evidence lenses help.
It's not a political moderator.
The same engine runs on a climate-policy argument and its counter. The frame is neutral; the conclusions you draw remain yours.
It's not a writing improver.
It audits but does not rewrite. Preserving your voice is deliberate. Fixing flagged issues stays with you.
It is not always right.
Findings flag structure, not verdicts. Treat them as prompts to look more carefully; calibration improves with feedback.
Open and inspectable
Every prompt, schema, and analytical framework is in the public repository. A tool that shapes how people read arguments should itself be transparent: if a finding seems off, you can read the prompt that produced it.