SATYA is the workflow and product layer from Invariant Research for verifying AI-generated work before people rely on it. Instead of asking another model whether an answer sounds correct, SATYA checks the output against source evidence, evaluates explicit constraints, and produces a signed verification receipt showing what was checked. SATYA is designed for high-stakes workflows where a chat transcript, confidence score, or human-review policy is not enough audit evidence.
The Problem SATYA Solves
AI outputs reach legal filings, compliance records, scientific publications, and agent actions every day. The verification layer between generation and consequence is missing.
Most AI verification tools use a second language model to estimate whether the first model's output is acceptable. That approach is probabilistic: a different prompt, a different temperature, a different day can produce a different verdict. SATYA replaces that pattern with deterministic structural verification. The same input always produces the same result, with zero learned parameters in the verification loop.
The output of every SATYA verification is a signed receipt, not a score. The receipt records the input hash, the source hashes, the constraints evaluated, the verification result, the verifier version, and a timestamp, all bound to an Ed25519 cryptographic signature. Any third party with the public key can verify the signature and replay the verification independently.
How SATYA Works
SATYA sits underneath your existing AI systems. Your LLMs generate. Your RAG retrieves. SATYA verifies.
SATYA uses SIGMA, the streaming structural verification engine built on cellular sheaf cohomology, to check whether AI-generated outputs are structurally consistent with their source evidence. It models claims, citations, policy assertions, protein constraints, and other structured relationships as a constraint graph, then checks whether the constraints can all hold simultaneously.
When constraints contradict, SATYA localizes the failure: which claim, which citation, which residue, which policy control failed, and why. When everything passes, SATYA produces a signed receipt recording the verification work performed.
InputAI-generated output + source evidenceA legal citation and its cited case. A compliance assertion and its policy source. A protein candidate and its native reference. A claim and the documents it relies on.
VerificationStructural constraint checkingSATYA evaluates whether the output is consistent with the source evidence across every constraint dimension. No ML model in the verification loop. No probabilistic scoring.
OutputSigned verification receipt (SVR)Ed25519-signed artifact recording what was checked, what the sources said, and whether the output passed. Independently verifiable by any third party.
What SATYA Verifies
SATYA applies to any workflow where AI-generated output becomes action, evidence, or liability.
LegalCitation and claim verificationChecks whether cited cases, statutes, and authorities exist and support the claims made. Catches fabricated citations before they reach a filing or docket.
CompliancePolicy contradiction and coverage mappingChecks whether policies, controls, and audit findings are consistent with each other. Identifies coverage gaps and contradictions across documents.
ProteinStructural coherence verificationChecks whether a predicted protein structure is coherent with its native reference across backbone geometry, native contacts, disulfide bonds, and steric constraints.
AgentsPre-action verification for tool callsChecks whether an agent's planned action is consistent with its policy constraints and prior state before the action executes.
What SATYA Does Not Claim
SATYA does not claim that a verified output is universally true.
A signed verification receipt records the verification work performed against specific source evidence and explicit constraints. It does not assert that the underlying sources are correct, that the constraints are complete, or that the output is true in every possible context. SATYA verifies structural consistency, not factual omniscience.
This distinction matters in regulated environments. A verification receipt is evidence of what was checked. It is not a guarantee that nothing else could go wrong. That honesty is a feature: auditors, regulators, and courts need to know exactly what was verified and what was not.
Relationship to SIGMA, SVR, and Locus
SATYA is the product layer. SIGMA is the engine. SVR is the receipt format. Locus is the transparency layer.
SIGMA is the streaming structural verification engine built on cellular sheaf cohomology. It maintains coherence checks as state changes, at O(1) amortized cost per edit, verified at 5 million vertices with 35 microseconds median latency.
SATYA applies SIGMA's verification to specific workflows: legal, compliance, protein, agents, scientific computing. It handles the domain translation, source retrieval, constraint construction, and receipt generation.
SVR (Signed Verification Receipt) is the output format. Every SATYA verification produces an SVR receipt: a cryptographically signed artifact binding the verification details to an Ed25519 signature.
Locus is the transparency layer that anchors receipt commitments into an independently checkable log. It records hash commitments so a verifier can confirm whether a receipt was registered without trusting a private database.
Frequently Asked Questions
Common questions about SATYA.
What is SATYA?
SATYA is the workflow and product layer from Invariant Research for verifying AI-generated work against source evidence before people rely on it. Instead of asking another model whether an answer sounds correct, SATYA checks the output against source evidence, evaluates explicit constraints, and produces a signed verification receipt showing what was checked.
How is SATYA different from using another LLM as a judge?
An LLM judge is probabilistic: a second model estimating whether the first model's output seems reasonable. SATYA is deterministic: it checks structural constraints against source evidence with zero learned parameters in the verification loop. The same input always produces the same result.
Does SATYA prove an AI answer is true?
No. SATYA does not claim universal truth. It records what was checked, which sources were used, what constraints were evaluated, and whether the output passed structural verification. The signed receipt is auditable evidence of the verification work performed.
What does SATYA verify?
SATYA verifies structural consistency of AI-generated outputs against source evidence. This includes citation verification, claim-source alignment, policy contradiction detection, protein structure coherence, and compliance coverage mapping.
Can verification receipts be audited by a third party?
Yes. Every SATYA verification produces an SVR receipt signed with Ed25519. The receipt binds the input hash, source hashes, verification result, and timestamp to a cryptographic signature. Any third party with the public key can verify the signature and replay the verification independently.
Does SATYA require sending documents to the cloud?
No. SATYA runs locally on Windows. Source documents stay on your machine. The verification engine operates on structured constraint checks, not cloud API calls.
Try SATYA.
Verify a citation, audit a compliance document, or check a protein structure. No signup. No cost.