same signals + same policy version → same Authorization Decision
2. No environmental dependency

Decisions must not depend on:

timestamps
system load
runtime state
external randomness
3. Replay protection

Each decision is bound to a unique executionId.

Duplicate execution attempts are rejected.

4. Independent verification

Any party can verify a decision using:

policy version
verified signals
public key

No system access is required.

5. Fail-closed behavior

If evaluation cannot be completed reliably:

→ execution is rejected

There is no fallback mode.

Confidence is not governance

AI confidence scores are probabilistic.

Governance is not.

Even high-confidence AI outputs:

cannot be used directly for execution
must be evaluated through deterministic rules
must be independently verifiable
Role of AI

AI generates signals.

Governance evaluates signals.

AI → Signals → Governance → Authorization Decision

AI is not part of decision logic.

How Parmana enforces determinism
Rule	Enforcement
Canonical input format	deterministic serialization
No runtime randomness	disabled in evaluation layer
Versioned policies	immutable policy bundles
Replay protection	executionId tracking
Signed decisions	Ed25519 attestation
Stateless evaluation	no external dependency
System impact

Determinism enables:

auditability
compliance verification
reproducible decisions
cross-system trust
Summary

Governance must be deterministic because execution is irreversible.

Parmana ensures:

AI generates signals
Governance evaluates deterministically
Authorization decisions are reproducible
Execution is strictly controlled
Every decision is verifiable