Assurance Doctrine

Disciplined Nondeterminism

We do not pretend AI is deterministic. When a process is stochastic, adaptive, learned, agentic, or heuristic, we make it measurable, bounded, replayable where possible, auditable, and tied to explicit evaluation criteria.

Recorded runs

Replays measured

—

Avg determinism score

—

Eval gate pass rate

1 · Measurable

Every run records a self-reported confidence and, on replay, a normalized determinism score in [0,1]. Reproducibility is a measured number, never an assertion.

2 · Bounded

Models are pinned, sampling temperature is capped (≤ 0.5), and latency is held to a declared budget. Out-of-bounds runs fail the gate.

3 · Replayable

Exact inputs, seed, model, and params are captured so a run can be re-executed and compared. Where true determinism is achievable, the score is 1.0; where it is not, the residual variance is reported.

4 · Auditable

Each run is content-hashed (SHA-256) and bound to the git commit SHA of the code that produced it. Records are append-only — nothing is rewritten in place.

5 · Explicit criteria

Outputs are scored against named, versioned criteria with thresholds (determinism ≥ 0.85, confidence ≥ 0.7). Required criteria form a hard pass/fail gate.

Mathematics anchor

Stochastic outputs are verified against the Sovereign 144 Master Equation and 49 deterministic proofs — a non-probabilistic floor beneath the learned layers.

RTOS Hard Deadline Gate

NIST SI-6 · SA-11

Inspired by real-time operating systems used in industrial control environments (IEC 62443 / AtomBeam grid security): a missed deadline is a system failure, not a preference violation. Every agent run is measured against a hard ceiling of 5000ms. Exceeding it fails the gate unconditionally, regardless of the per-run budget.

Override via ASSURANCE_LATENCY_HARD_GATE_MS environment variable. Current per-run budget: 60000ms.

Evaluation Criteria & NIST Control Mapping

Every agent run is scored against these named, versioned criteria. Required criteria form a hard pass/fail gate. Control IDs are NIST SP 800-53 Rev 5.

Criterion	Description	Required	Severity	NIST Controls
`output_nonempty`	Run produced a non-empty, captured output	REQUIRED	HIGH	NIST SP 800-53 SA-11 NIST SP 800-53 SI-2
`audit_bound`	Run is bound to a content hash + input hash (auditable)	REQUIRED	CRITICAL	NIST SP 800-53 AU-10 NIST SP 800-53 AU-12
`git_bound`	Run is bound to a code commit SHA (replayable against versioned code)	OPTIONAL	MEDIUM	NIST SP 800-53 CM-3 NIST SP 800-53 SA-10
`model_pinned`	Model identifier is explicitly pinned (no 'auto')	REQUIRED	HIGH	NIST SP 800-53 CM-6 NIST SP 800-53 SA-4
`temperature_bounded`	Sampling temperature <= 0.5 for high-integrity tasks	REQUIRED	HIGH	NIST SP 800-53 SA-11 NIST SP 800-53 SI-7
`latency_bounded`	Latency within declared per-run budget (bounded compute)	REQUIRED	MEDIUM	NIST SP 800-53 SA-11 NIST SP 800-53 SC-5
`latency_hard_gate`	Hard real-time deadline: latency <= 5000ms (RTOS-style, non-negotiable)	REQUIRED	CRITICAL	NIST SP 800-53 SI-6 NIST SP 800-53 SA-11
`replay_evidence_present`	Replay evidence exists when replay is possible (executor registered)	REQUIRED	HIGH	NIST SP 800-53 SA-11 NIST SP 800-53 SI-6
`determinism_min`	Replay determinism score >= 0.85	REQUIRED	HIGH	NIST SP 800-53 SA-11 NIST SP 800-53 SI-7
`confidence_min`	Self-reported confidence >= 0.7	OPTIONAL	MEDIUM	NIST SP 800-53 SA-11 NIST SP 800-53 RA-5

Agents under assurance

reference_agent
zeroclaw_tool_agent

Aggregate metrics are computed live from the assurance ledger. Detailed per-run records, replays, and evaluations are available to authorized operators via the admin controls.