Today the reasoning behind a decision lives buried inside documents. A web page mixes the claim, the evidence, the rhetoric, and a great deal of unstated assumption into one blob, and if you want to cite it, you cite the whole page. Isonomia breaks that apart. A claim or an argument becomes its own object with a permanent address, carrying what supports it, the sources behind it (fetched, timestamped, and verifiable, so the record holds even if the original link later rots), the strongest objection on file against it, and whether it has survived challenge.
Why this matters
Reasoning is the one thing software has never given a durable home. We store documents, messages, transactions, and code, but the inferential structure that connects evidence to a claim to a conclusion is thrown away the moment a decision is made — left to be reconstructed, badly, from prose every time someone asks why.
That gap is becoming expensive at exactly the moment it is becoming unavoidable. AI systems are moving from answering questions to acting across tools, and they need a place to read and write reasoning state — what has been asserted, what supports it, what attacks it, what survived, and what a conclusion does not yet license — that lives outside any single model, where it can be inspected, versioned, and contested. Better models do not remove this need; they raise it, because the cheaper it becomes to generate a confident claim, the more valuable an external, accountable record of which claims have actually earned standing.
The problem it addresses
As language models produce a growing share of what we read, the bottleneck is no longer fluent text — there is effectively infinite fluent text now. The bottleneck is trust: can you tell where a claim came from, whether its source actually supports it, and what it leaves out? Models are good at writing about a topic and bad at keeping its structure straight. They flatten which point is doing the real work, miss objections aimed at the logic rather than the conclusion, and cite a page when the meaningful unit is a single sentence. Asking one model to grade another’s reasoning yields a verdict that shifts with the wording.
The academic field that studies how to represent arguments formally is called computational argumentation. Isonomia applies it to exactly this gap: it keeps the structure of reasoning outside the model, where the things that can be checked are checked rather than guessed.
How it works
The structure is captured inside the work where reasoning is already happening — research, document review, policy analysis — rather than bolted on afterward as a separate annotation chore, which is precisely where earlier attempts at this broke. AI can propose the first-pass structure; people confirm or correct it where the commitment matters; and the system records who authored what, so machine-generated material stays visibly provisional until a human ratifies it.
Everything is exposed both as ordinary web pages and over the same kind of connection AI assistants already use to reach outside tools (the Model Context Protocol), so a search engine, a researcher, and an automated agent all read the same object. Every citation carries its strongest known counter-argument by default, which makes one-sided citation harder to do by accident.
Under the hood, this rests on formally studied frameworks — ASPIC+ grounded extensions, Walton’s argumentation schemes with their critical questions, Girard’s Ludics as an interaction semantics, and a category-theoretic evidence algebra that folds confidence lawfully in log-odds space. That complexity is invisible to anyone who does not seek it out; it is what lets the interface stay simple.
What is public, and what requires an account
Public arguments and claims resolve to permanent, login-free permalinks (/a/{shortCode} and /c/{moid}), and the corpus is a crawlable, machine-citable search surface. Each permalink is reachable as HTML, JSON-LD, AIF, or a compact attestation envelope, and an immutable, content-hash-pinned form survives future edits.
Reading and citing public arguments and claims requires no account. Authoring — proposing arguments, filing challenges, joining rooms, and driving dialogue moves — does.
Where it leads
The single object is the seed of larger infrastructure, and each next step extends what already runs rather than pivoting away from it:
- A web of machine-citable arguments — a public corpus where every claim carries its support, its opposition, its provenance, and its standing, addressable the way papers are today.
- A reasoning backend an agent can call — to learn whether a claim has survived challenge, where a disagreement actually lives, and what it would have to retract to reject a conclusion, so an AI writer gets not just sources but a contract for what it may not assert and what it must hedge.
- A minimal-disagreement debugger — that locates the single point where two positions first diverge instead of declaring them globally opposed.
- A transport layer — that moves arguments between communities and institutions with their provenance intact, and an evidence algebra that folds confidence lawfully rather than as an opaque score.
The compounding asset is not any one feature but the accumulating graph of reasoning itself.
Who it’s for
The value concentrates wherever the reasoning behind a decision matters as much as the decision itself and has to survive scrutiny later: research groups, peer review, policy and regulatory analysis, and the teams evaluating AI systems. Alongside those, AI tools themselves are a natural user — a model consuming structured arguments as a citation source does not tire of structure the way a human volunteer does, and an agent that writes back into the graph extends a shared record rather than spending its work into prose that scrolls past. A broad public version is the long-term upside, not the starting point.
Open source, self-hostable, and yours
Isonomia is free, open-source, and self-hostable. There is no behavioral tracking, no algorithmic ranking, and no engagement metric. The communities that produce the reasoning own it and can carry it out in open formats; the reasoning graph is content-hashed and cryptographically auditable, so provenance is enforced by architecture, not policy. The project is sustained by grants, institutional partnerships, and optional managed hosting — never by advertising or by selling data. Source lives on GitHub; see the privacy documentation for details.
Learn more
- Architecture — the social layer, reasoning layer, and how they fit together.
- Argument graph — claims, arguments, critical questions, challenges, citations, and confidence.
- Browser extension — create evidence-backed arguments from any webpage.
- Self-hosting — deploy and operate your own instance.