Models generate. WINDI verifies.

AI systems can produce convincing responses without possessing independent evidence. The question is no longer just 'what did the AI answer?' but 'what evidence survives the model?'

✓ production

Verification

"Is this artifact registered?"

Verify receipts, hashes, and evidence without trusting the AI model.

Evidence: Ledger receipts with cryptographic proof

Proof limit: FOUND does not establish truth of content

👤 production

Human Approval

"Who actually authorized this decision?"

I9 invariant: AI proposes, human decides, system records.

Evidence: In governed sealing flows, human_approved=true is required

Proof limit: Does not cover all system surfaces

📋 production

Evidence

"What proof will survive after the model?"

Immutable receipts with cryptographic proof and audit trails.

Evidence: I11: Evidence integrity preserved by cryptographic proof

Proof limit: Append-only does not guarantee content accuracy

🪪 production

Identity

"Who acted and with what authority?"

Sovereign DIDs that separate session from agency.

Evidence: I17 invariant: Session proves presence, Identity proves agency

Proof limit: Identity does not verify external credentials

🔗 production

Provenance

"Where did this artifact come from?"

Track origin, transformations, and chain of custody.

Evidence: I19 invariant: Generation and receipt are atomic

Proof limit: Does not verify origin outside WINDI

🔌 candidate

Connectors

"How can my AI consult WINDI capabilities?"

MCP, APIs, and protocol bridges for AI-native interoperability.

Evidence: CANDIDATE · W-TUBE LIVE · 10 read-only governed tools demonstrated

Proof limit: SDK-native interop pending final gate

Core Semantics

GENERATED ≠ VERIFIED Production is not proof

FOUND ≠ truth Discovery is not certainty

NOT_FOUND ≠ false Absence is not negation

INTAKE ≠ authentication Ingestion is not validation

CANDIDATE ≠ SEALED Working is not ratified