Large language models are fluent but unfaithful — they can't tell you why an answer is true, and they fail silently. In medicine, law, and finance, a confident wrong answer is worse than no answer at all. NuSy is built for exactly those decisions: a neurosymbolic platform whose provable layer, by construction, cannot hallucinate.

What NuSy delivers

• Answers that carry their proof. When NuSy stands behind an answer, it emits the derivation — every step traceable to source. Not a confidence score; an auditable chain of reasoning.
• Abstention you can trust. When the symbolic path can't prove an answer, NuSy declines rather than fabricating. On the provable path, hallucination is zero by construction, not by hope.
• Structured, auditable memory. Knowledge lives in a stratified, versioned knowledge graph wired directly into inference — every fact traceable, every relationship explicit, the entire knowledge state diffable and reversible.
• Domain experts, schooled — not trained. A being's expertise comes from the curriculum it studies, not from code changes. The same platform produces a medical expert, a legal expert, or an engineering expert — expertise as auditable, versionable, transferable data.
• Healthcare-native. Speaks the clinical semantic stack — UMLS, SNOMED-CT, LOINC, RxNorm, ICD-10, HL7 FHIR, and FHIR-CPG — natively. Computable clinical knowledge, first-class.

Who it's for

NuSy targets the high-stakes domains where unauditable AI is a non-starter:

• Clinical decision support (first). Provable, guideline-grounded reasoning that a clinician and a regulator can both audit — the hardest possible proving ground, and the sharpest fit for the architecture.
• Law, finance, and other regulated, high-consequence domains. The same never-launder guarantee, a different curriculum.

Why it's hard to replicate

• Never-launder separation — the provable/heuristic boundary is a runtime invariant, not a prompt. A guess can never be minted as a fact.
• Versioned knowledge as an enforced artifact — Apache Arrow + graph-native Git. Knowledge is reviewed and versioned like code, not baked opaquely into weights.
• Built by a fleet, governed like code — a fleet of AI coding agents authors and maintains the symbolic codebase under cross-agent review. An approach that compounds, not a one-off model.
• An open, reproducible record — the method is published as open preprints and the working artifacts below are public. Claims can be checked, not just cited.

The proof

A platform claim is only as good as the working artifacts behind it. Everything below exists to demonstrate the claim above — each with an honest maturity label, because coverage is nascent and the differentiators are real, and we say both.

NuSy Drug-Interaction Database — the DDI database that shows its work

Open source (MIT) · Live: opendruggraph.com · Congruentsys/opendruggraph — FDA-label-cited edges, coverage growing

What it proves. That proof-carrying answers work at production shape: every asserted interaction cites its exact FDA Structured Product Label sentence; every mechanism-derived interaction shows its mechanism path as the proof; and when the database doesn't know, it abstains out loud instead of guessing "no interaction."

What it does. Checks drug-drug interactions and — uniquely — the interactions a flat pairwise table structurally can't hold: mechanism chains (drug → enzyme/transporter → drug), 3+-drug polypharmacy (additive QT/serotonergic/bleeding load, shared-enzyme hubs, cumulative organ burden), plus drug-disease contraindications, drug-indication, and duplicate-therapy screening. Delivers via FHIR / CDS-Hooks.

Why it matters. The incumbents tell a clinician a pair is "major" and won't say why — and a 2025 review found the popular checkers agree on only 16–24% of flagged interactions. NuSy shows the FDA-label evidence, represents the multi-drug interactions the incumbents exclude, and is the structural answer to alert fatigue — MIT-licensed and free versus $10k–100k+/yr proprietary feeds. Honest scope: we win on transparency, representation, and safety model, not on catalog breadth yet — coverage is growing and labeled as such.

The Never-Launder Reasoning Engine — structural zero-hallucination

Pilot-ready · the platform under everything on this page

What it proves. That "cannot hallucinate" can be an architectural property rather than a benchmark score. A language model proposes and a symbolic gate disposes: every answer is either Proven (with a full provenance trace), Heuristic (clearly labeled), or a loud abstention — never a guess dressed as a fact.

What it does. Turns unproven claims into a class of error that is impossible by construction, attaching a proof object to each answer. Verified on internal batteries at false_proofs = 0 across 100K synthetic patients and millions of derivations.

Why it matters. For any regulated-domain decision support — clinical, payer, legal — the buyer gets auditable correctness they can inspect, satisfying "transparent basis, clinician-reviewable" requirements by design. Almost no AI vendor can produce a claim-to-evidence chain like this; it is the core differentiator everything else here is built on. The method is documented in the open research record.

Provable Guideline CDS — "prove it on YOUR guideline"

Pilot-ready

What it proves. That the engine generalizes: one generic reasoner applies computable clinical guidelines across publishers with zero per-guideline engine changes — demonstrated on NCCN (four cancers), JNC-8 hypertension, and WHO antenatal care, with every recommendation covered by a passing knowledge test.

What it does. Bring a computable clinical guideline (a FHIR PlanDefinition); NuSy reasons over it provably — contraindications fire, and missing data makes it abstain out loud rather than guess.

Why it matters. Guideline-grounded decision support that is defensible and auditable, adaptable to a partner's own guideline without bespoke engineering — the fastest path from "trust our model" to "here is the proof over your content."

The open-source projects — the approach, in public

All maintained, all verifiable today — see Open Source for the full registry and which promise each project carries.

The strongest evidence that a method works is the working infrastructure it produces. These are the actual tools our own agent fleet runs on, published:

• arrow-kanban — the Arrow-native work-tracking engine the fleet coordinates through daily: typed relationships, a NATS server mode, a durable event log. MIT; the public build is the fleet's build.
• acf-framework — the AGI Certification Framework: 12 capability dimensions, 16 canonical hypotheses for measuring whether an AI system reasons, instead of asserting it.
• noesis-ship — the NATS-based multi-agent communication layer under the fleet.
• yurtle — the knowledge-representation format (Markdown + TTL frontmatter): human-readable, machine-queryable, git-versioned. A stable published spec.
• opendruggraph — the DDI database above, data and pipeline in the open.

On the roadmap (in development — not yet available)

Listed for transparency; clearly marked as not shipping today, per our honesty brand.

• Terminology Service — a proof-carrying clinical terminology/value-set resolver (the substrate every clinical product silently trusts — but with proofs on mappings and membership, which no incumbent carries). In development; the stated prerequisite for MTM CDS.
• MTM CDS Service — medication-therapy-management decision support with proof-carrying recommendation cards (CMS-mandated MTM; adherence measures are triple-weighted in STARS). In development; depends on the Terminology Service.

Coverage/parity claims are gated by our compete-ready discipline — no breadth-parity claim ships until the coverage work lands with on-disk evidence.

See for yourself

• How it works → The NuSy platform
• The published record → Research
• Run what's open → Open Source