Someone makes a decision on a page that another page contradicts, and nothing in your stack was looking. Curio gives you one entry per thing your documents describe, with its disagreements listed, both sentences quoted and both sources named. That is possible because the layer holds rules about what can be true at the same time, instead of ranking which passage looks most like the question.
Most documentation starts clean. Then it spreads across pages, branches, translations and copied runbooks. Search gets noisy, answers disagree, and the assistant you put in front of it inherits every disagreement without knowing it.
The same process gets rewritten in several places, and each copy drifts. Teams end up comparing pages instead of getting an answer.
Search returns old material because it is popular, not because it is correct. Decisions that were reversed still read as current.
Once trust drops, people ask in chat instead. The answer lands in a DM, and the documentation gets worse.
Retrieval over a corpus that disagrees with itself still produces a confident answer — built from whichever chunk scored highest. The model can't tell that the passage it found describes a different version, a different distribution or a different population, and it has no way to tell you the sources disagree.
Three approaches to the same pile of documents, on the questions a buyer actually has. These are approaches, not products.
| Your question | Search over documentsRAG | Search over a graph of factsGraphRAG | A checked knowledge layerCurio |
|---|---|---|---|
| What happens when two sources disagree? | Returns the passage that scores highest and answers from it. The disagreement never surfaces, because nothing compared them. | Stores both. Two facts about the same thing are two entries; nothing decides they cannot both hold. | A thing is removed once, and one deadline is one deadline. Two different answers in the same slot are a contradiction the layer derives, and it shows both sentences with their sources. |
| Do I have to already suspect the problem to find it? | Yes. It answers questions. A contradiction nobody asked about stays invisible. | Largely yes. Traversal follows the question you ask. | No. Everything comparable is compared, so findings arrive without anyone suspecting them first. |
| What decides that something is a finding? | A similarity score, which moves with the query, the chunking and the phrasing. | A judgement made at answer time about two stored facts. | A rule written down before the run. That is what lets a finding sit in a release gate or a review record rather than in a chat window. |
| When is it finished? | Never. A ranked list has no bottom. | Never, for the same reason. | A finite, triaged queue. You can work through it and say so. |
| What about an answer that is right for a different version, region or population? | Passes every check. The text matches; the scope is invisible. | Only if scope was modelled as data, and only if the query asked about it. | Scope is a field on every claim. “Right answer, wrong version” is its own verdict. |
| What does a finding come with? | A passage and a confidence score. | A path through the graph. | Both sentences, both sources, the byte range in each, and the rule that made them incompatible. |
Contradictions are the hardest case for similarity, not the easiest. Changing a number, a date or a “not” is the smallest possible edit to a sentence, so the contradicting version looks more like the original than a harmless rewording does. That is a property of how similarity works, not a gap that a better model closes.
Curio does not replace your search or your assistant. It sits underneath them, so that what they retrieve has already been checked.
One layer produces all of them.
One entry for every thing your documents describe, with every source it came from listed.
Pre-sorted into genuine disagreement, extraction artifact, and needs-a-human.
Constraints that hold across your sources and are written down in none of them.
Supported, contradicted, unsupported, or out of scope — each with the sentence behind it.
Filtered to your version, your variant, your jurisdiction.
The description of what can be true in your field, which stays yours and keeps working as your documents change.
Indexing finds passages that look like your question. Curio does something different: it works out the things your documents describe, records what each source says about each one, and checks those statements against each other.
All your sources, including the ones in other languages and the old branches nobody ever deleted. Nothing is excluded for being out of date — being out of date is a fact worth knowing.
Not keywords — the real things: a setting, a process, a component, a threshold. Every way each one gets written is pulled together into a single entry, so four spellings of the same version number stop being four different facts.
Every statement is attached to the thing it describes, tagged with the version, product variant, region or population it applies to, and linked back to the exact sentence it came from.
The rules are written down before the run: a thing is removed exactly once, one deadline is one deadline. Whether two statements can both hold is decided by those rules, not by a score that moves with the query, the chunking or the phrasing — so every finding arrives with the rule that produced it.
“PodSecurityPolicy was removed in v1.25.”upstream · k8s/website
“PodSecurityPolicy was removed in v1.24.”vendor · distribution docs
The layer is an OWL ontology. Extraction produces entities, relationships and claims; contradictions are entailed by rules stated in advance, not inferred from similarity. Properties that can only hold one value — the release something was removed in, for example — turn two different canonical answers into a logical contradiction. Obligations that do not contradict but bind at the same time are surfaced as a co-applicability cluster, with the tightest value identified. Every claim carries its source document, byte span and content hash, so any statement is citable down to the sentence. Scope is explicit and typed: release, distribution, language, population, region, jurisdiction, notification stage.
Curio's own reference corpora: 228 documents and 7,349 claims across 27 entity classes in software; 145 documents and 6,916 claims in medicine; 4 instruments and 445 claims in EU law. These are public documents we selected, not customer results.
Narrowing is what makes a finding list finite. In software, 491 restatements were collapsed, 13 duplicate claims removed, 3 candidates adjudicated and cleared and 1 deduplicated, leaving 13 findings: 6 version drift, 4 numeric divergence, 3 logical inconsistency — 10 of them spanning more than one source and 4 more than one language. In medicine, 28 false positives were quarantined, 10 candidates adjudicated and all 10 cleared, 7 polarity candidates cleared and 9 deduplicated, leaving 24 findings — 15 numeric divergence, 6 temporal drift, 3 polarity — alongside 13 observations, of which 9 are projections and 4 are stated trends; 20 of the findings span more than one source and 13 more than one language.
The same layer serves all three. It is not three products.
Cleanup stops being an endless manual audit and becomes a reviewable queue.
TopologyManagerPolicyBetaOptions — default disabledv1.27 docs
TopologyManagerPolicyBetaOptions — default enabledv1.30 docs
deleted after one releaseone page
remain available for useanother page
An answer is only trustworthy if something can verify it. The layer is that something.
Retrieval hands a model whatever text scored highest. The layer hands it the answer, its scope and its caveats.
Verification decomposes an answer into atomic claims and resolves each one against the ontology. Support and contradiction are entailment results, not scores; scope mismatch is a separate verdict because a claim can be entailed and still be about the wrong release, distribution, population or jurisdiction. Retrieval is entity-first: resolve the mention to an entry, then select claims whose scope satisfies the query context, then return them with their provenance spans.
It's one layer. Curating your knowledge is the same work that makes your agents reliable — you do it once.
Curio runs today on medical literature with a metabolic-health focus, on Kubernetes ecosystem documentation — kubernetes/website, k3s, minikube and OpenShift, across several release branches and several languages — and on EU regulatory instruments: GDPR, NIS2, DORA and the DORA delegated RTS, in their in-force consolidated versions from the official EU source. Not three configurations of a tool: the same code, unchanged. Every piece of domain knowledge lives in the ontology pack for that domain, never in the engine, and an automated test swaps one domain for another to show the engine has learned none of them.
That is the reason to expect it to work on your corpus, which resembles none of the three.
An incident at a regulated financial entity that also processes personal data. GDPR, NIS2 and DORA each apply, under a different supervisory regime, and five separate initial-notification duties bind at the same time.
The EU corpus produced no conflicts at all. Its finding is this cluster: obligations that do not contradict each other and all hold at once. Across 4 instruments and 445 claims there are 3 cross-instrument co-applicability clusters.
“72 hours after becoming aware”GDPR · Art. 33 · consolidated version
“72 hours of becoming aware”NIS2 · Art. 23 · consolidated version
“24 hours”NIS2 · Art. 23 · consolidated version
“no later than 24 hours from becoming aware”DORA delegated RTS · Art. 5 · consolidated version
“4 hours from classification as major”DORA delegated RTS · Art. 5 · consolidated version
Every member of the cluster carries all six of these:
These are public documents we selected as a reference corpus. They are not a customer result.
The EU material is reused under Commission Decision 2011/833/EU, with the source acknowledged.
Only the Official Journal edition is authentic. Consolidated texts are documentation tools and have no legal value.
6.1 to 6.9 mmol/L (110 to 125 mg/dL)WHO
5.6 to 6.9 mmol/L (100 to 125 mg/dL)ADA
No domain literals in engine code, no branching on domain name, and a domain-swap parity test that runs one domain's configuration through another domain's pipeline and back. All of it is enforced in CI. Adding a domain means writing an ontology pack, not changing the engine.
Private beta. Every corpus on this page is public material we selected ourselves, not a customer result.
Curio is not a one-time cleanup project. The layer is rebuilt as your documentation changes, and everything new is checked on the way in.
Wiki, docs, runbooks, repos, and every version of them. Old branches and other languages included.
One entry per thing your documents describe, with every claim attached to its scope and its source sentence.
Reviewed side by side with citations. Extraction artifacts have already been filtered out.
New, restatement, or contradiction. You find out before it's published, not after someone acts on it.
Scoped context in, answers checked on the way out.
If your documentation has stopped being something you'd point a new hire at — duplicated pages, stale branches, answers that disagree — the layer is what turns that into a queue you can actually work through.
And if you're putting an AI assistant in front of documentation nobody has verified, the layer is what its answers get checked against before your users believe them.
Curio is in private beta. Leave your email and we'll get in touch when access opens.
