Why it exists
Software gives durable homes to documents, messages, transactions, and code. Isonomia extends the same treatment to reasoning itself: the inferential structure that connects evidence to a claim to a conclusion. That structure is becoming valuable at exactly the moment AI systems need it — a shared place to read and write reasoning state: what has been asserted, what supports it, what attacks it, and what has survived, where all of it can be inspected, versioned, and contested.
Reasoning becomes a graph
Each claim and each argument is a node with a permanent, content-hashed address. Each node carries its premises, its sources, the strongest objection on file against it, and a plain standing label: untested, tested, or survived. The edges between nodes are typed — an argument can support a conclusion, rebut it, undercut the inference that leads to it, or undermine one of its premises. The whole structure can be followed, cited, and checked, by a reader, a search engine, or an AI agent.
Deliberation and construction
Two kinds of work run on the same graph. The floor is eliminative: typed attacks, critical questions generated automatically from each argument’s scheme, and a refusal surface listing the conclusions that require open challenges to be answered before the graph licenses them. The ceiling is ampliative: the symposium, where structured arguments, warrants, inference chains, and syntheses are built, guided by an opportunity frontier that ranks the open work the graph invites — critical questions awaiting a first answer, off-path arguments ready to chain in, parallel conclusions ready to synthesize. The two interlock: floor verdicts flow into every construction through its weakest link, so a construction is exactly as trustworthy as the reasoning beneath it.
Evidence carries its provenance
Any URL or DOI pasted into a citation resolves automatically into a verified bibliographic record, through a fixed waterfall: arXiv, Crossref, the page’s own metadata, OpenAlex, a flagged language-model extraction, and the Internet Archive. Every resolution carries an explicit confidence tier. Cited pages are fetched, timestamped, and archived with a stable snapshot, so the evidence stays addressable for as long as the argument does. Citations are executable: each one navigates to the exact page, passage, figure, or section it refers to.
A verification copilot for claims in the wild
The same discipline turns outward through a browser extension. On any page, the provenance lens checks whether a cited source supports the claim in front of it, detects drift — a dropped qualifier, a changed denominator, a rate stated as a count — and traces a page’s independent sources back to their shared roots. Pointed at AI-generated text, it verifies each citation in an assistant’s answer against the source it names. The direction is chain backing: follow a claim up its syndication chain — post, press release, study — judge each hop against the next, and produce a per-hop drift ledger recording where the claim changed and at which hop the change entered.
An agent reasoning backend
Every argument is served as a web page and, through content negotiation, as JSON-LD, an Argument Interchange Format graph, and an embeddable card. A bidirectional Model Context Protocol surface gives an AI agent read tools — arguments, counterarguments, claim stances, the legal move set — and write tools that propose arguments, chains, and warrants, answer critical questions, file typed attacks, and drive protocol dialogue moves from a declared dialogical side. AI authorship is flagged on every record, human ratification lifts machine-drafted material to full standing, and every citation carries its strongest known objection by default. Standing is always relative: survived means the argument withstood every challenge mounted against it so far.
Provenance-aware writing
Long-form outputs — theses, briefs, peer reviews — are living documents whose embedded claims and arguments read live from the graph. An attack register shows where the document is exposed and where it has held. Confidence cards disclose every input behind a score. Snapshots freeze a point in time and diff against the current state. Hedging keys itself to standing, so the prose stays calibrated to the reasoning beneath it.
Who it’s for
The value concentrates wherever the reasoning behind a decision is worth as much as the decision itself. Research groups and peer review bind evidence to claims and track which arguments survive challenge. Policy bodies and regulated institutions receive an audit trail generated by the process itself. AI builders treat the graph as a citation and reasoning backend, and evaluation teams use its computed manifests — fixed corpora, structural ground truth, refusal surfaces — to grade systems that claim reasoning competence.
Want the full picture? Read the overview, the architecture, and how the argument graph works. Isonomia is free, open-source, and self-hostable, and every community owns its data and can export it in open formats. Source lives on GitHub.