How it works
Plain-English guide to the eleven lenses the engine applies, the philosophical frameworks behind them, and the limits of what the tool can and won't do.
What the engine actually does
You paste a piece of argumentative text — an op-ed, a thread, a draft, a transcript. The engine reads it as an argument: not as prose, not as opinion, not as a verdict on who's right. It returns a structural view of how the argument works.
Eleven analytical lenses run in parallel. Each catches a different failure mode in reasoning. The output is a structured report — claims, evidence, hidden assumptions, fallacy patterns, opposing positions — with verbatim quotes from your text, severity ratings, and confidence scores on every finding.
The eleven analytical lenses
Each lens has a name in the code (used by API consumers) and a plain-English label (used in the UI). The philosophical pedigree is given because the lenses come from somewhere — they aren't arbitrary.
1. Argument structure — Toulmin decomposition
The claim, grounds, warrant, qualifier, and rebuttal. Plus the weakest link — which of these components is least supported. From Stephen Toulmin's The Uses of Argument (1958).
2. Reasoning patterns — 26 named fallacies
From Ad Hominem to Motte-and-Bailey, Selection Bias to Genetic Fallacy. Each finding includes the verbatim quote that exemplifies it, the severity (high / medium / low), and the confidence (50–100%). Catalogue grounded in Aristotle's Sophistical Refutations, extended by Walton and Hansen.
3. Loaded language — 8 rhetorical techniques
Emotionally charged terms, weasel words, glittering generalities, scare quotes, presupposition smuggling, euphemism, false-precision numbers, name-calling. The argumentative work done by diction rather than logic.
4. Unstated warrants — hidden assumptions
The premises the argument needs but doesn't state. Most arguments are enthymemes — Aristotle's term, in the Rhetoric, for syllogisms with a suppressed premise. The hidden premise is often where the disagreement actually lives.
5. Key-term scrutiny — Wittgensteinian
Tracks whether load-bearing words are used consistently across the text. Detects stipulative smuggling, cross-language-game equivocation, and family-resemblance overreach. Grounded in Wittgenstein's Philosophical Investigations.
6. Vague reference check — Russellian
Appeals to groups whose referent is empty ("real Americans"), vague ("the establishment"), or presuppose facts not established ("the proven link between X and Y"). From Russell's On Denoting (1905) and his theory of definite descriptions.
7. Testability check — Davidsonian falsifiability
For each major claim: what evidence would refute it? Claims with no truth conditions, circular truth conditions, or that are unfalsifiable while dressed as substantive are flagged. Grounded in Davidson's truth-conditional semantics.
8. Certainty claims — modal scope check
Where the argument substitutes necessity ("will", "must") for possibility ("might", "could") without the evidence to license the upgrade. Or where premises hedge but the conclusion drops the hedge. Grounded in Kripke's modal logic semantics.
9. Counterarguments — Studio tier
The 2–3 strongest opposing positions a thoughtful opponent would deploy — each steelmanned, each with its own Toulmin structure. Standalone free version available at /strongest-opposing-case.
10. Philosophical commitments — Studio tier
The ethical, epistemic, political, and methodological framework the argument operates within — consequentialist vs deontological, empiricist vs rationalist, individualist vs structuralist. Made visible so the writer can decide whether to defend, soften, or replace the underlying frame.
11. Evidence-weighted likelihood — Studio tier
For empirical claims only, we search 200M+ papers via Semantic Scholar and synthesise the scientific consensus — strong support, contested, opposed, insufficient data. Value claims are explicitly flagged as not empirically assessable. Grounded in Hume's is–ought distinction (1739).
Three ways to use the engine
Reader · free
Paste text or a URL at /reader. Returns the full lens suite minus the Studio-tier features. Five audits per day, no account required.
Creator Studio · $22/mo
Full lens suite at /creator/studio. Document persistence, revision tracking, inline highlights, counterarguments, citation audit, evidence-weighted likelihood, transcript audit. For writers who care.
Devil's Advocate · Chrome extension
Audit any argument inline while reading on the web. Highlight text or analyse the whole article. Inline highlights overlay the original page. Coming to the Chrome Web Store.
What it won't do
- It won't adjudicate claims. The tool analyses argument structure, not factual accuracy of specific assertions. Whether a cited study is real, or a statistic holds up, is your job — though the citation audit and evidence-weighted likelihood lenses help.
- It's not a political moderator. The same engine runs on a climate-policy argument and a climate-sceptic counter-argument. The frame is neutral; the conclusions you draw remain yours.
- It's not a writing improver. The engine audits but does not rewrite. Preserving the writer's voice is a deliberate design choice — the work of fixing flagged issues stays with the writer.
- It is not always right. Findings carry confidence scores (50–100%) precisely because the tool can be wrong. Treat findings as prompts to look more carefully, not as verdicts. Calibration improves with feedback — the thumbs on each finding feed the prompt-iteration process.
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.
View on GitHub →