Skip to content

Pramāṇa

Pramana (Sanskrit for valid means of knowledge) defines a protocol-layer claim verification primitive for agent-to-agent networks. Every consequential agent output is wrapped in a typed ClaimAttestation whose verification operation is deterministic against the recorded source.

A2A and MCP standardize agent communication at the syntactic level (messages, tasks, agent cards, tool calls, tool results). Neither defines a typed vocabulary for the epistemic ground of an agent output. A tool result in MCP is a content blob; an A2A agent message is a text payload. Neither carries a typed attestation of source URI, measurement record, or inference chain.

Production verification today splits into two unstandardized halves. Probabilistic-verdict patterns (self-consistency voting, confidence-scored outputs, reviewer LLM ensembles) produce judgments about model outputs rather than the auditor-replayable artifact regulators ask for; aggregating verdicts does not change their category. Artifact-producing patterns (retrieval-augmented generation with citations, tool-augmented traces, generator-verifier loops as in FunSearch and AlphaEvolve, recent multi-agent research systems such as Google’s AI co-scientist) produce vendor-specific records but no shared wire format. An auditor inspecting a multi-vendor agent network cannot reconstruct verification across patterns without bespoke per-vendor integration.

Pramana defines the typed wire format that standardizes the verification artifact across both halves. Each agent output declares the epistemic ground that warrants the claim (measurement, inference, analogy, or citation) plus the metadata required to verify it independently against the recorded source.

PrimitiveEpistemic groundVerification operation
MeasurementClaimDirect observation / structured recordSource-record fetch and field match
InferenceClaimLogical inference from prior claimsInference-chain replay
AnalogyClaimSimilarity to a known reference caseSimilarity recompute against reference
CitationClaimAttribution to an authoritative sourceSource fetch plus faithful-citation check

For MeasurementClaim and CitationClaim, verify() is a deterministic function of (claim, source) and no probabilistic judge participates in the verification step. For InferenceClaim and AnalogyClaim with LLM-backed oracles, Pramana’s contribution is narrower: the verification step’s inputs and outputs are audit-replayable. Deployments that need full deterministic re-verification plug a deterministic oracle into the same dependency slot without changing the wire format.

Five named safety invariants exhaustively verified under TLC across three symmetry-reduced models (Lifecycle, Disclosure, Concurrency): 38,563 distinct reachable states, 0 violations.

InvariantProperty
P-1Single Emission: each claim has at most one emission audit entry
P-2Verification Determinism: each claim’s verification state is single-valued and terminal at most once
P-3Audit Completeness: every terminal verification, suppression, and disclosure is in the audit trail
P-4Disclosure Coupling: a claim shown to a principal has emit, verify, and disclose audit entries
SuppressionDisclosureDisjointNo claim is both suppressed and shown to a principal

The claim-attestation extension adds three deployment-grade invariants that make Pramana Core’s single-agent lifecycle guarantees checkable end-to-end across an agent network.

InvariantPurpose
CA-1 ReachabilityEvery emitted attestation includes verify_endpoint_hint so any receiver can round-trip verification
CA-2 SLA-BoundEvery accepted attestation reaches a terminal verification state within the declared sla_window_ms
CA-3 Offline Re-verifiabilityEvery VerificationOutcome is re-verifiable offline given (claim, source_digest, artifact_signature)

CA-1 and CA-3 are TLC-verified; CA-2 is enforced at runtime by the extension’s validators.