RUNTIME AUTHORITY + SIGNED EVIDENCE FOR AGENTIC AI

Signed proof of what your agents did.
And the authority to stop them.

Your agents are starting to spend money, touch production, and talk to customers. Axiom sits between them and every consequential action — authorizing, holding, or blocking it before it executes — and hands you a signed receipt an auditor can verify.

<50 ms verdict latency HMAC-SHA256 signed receipts Model-neutral Claude · NIM · Llama · local
 AXIOM AUTHORITY REGISTER — LIVE
AGENTpayments-agent-04
REQUESTwire.transfer $18,400
POLICYfinance.ops.v3 · immutable
AUTHORITYACTIVE_VALIDATED
GATEHELD — human approval required
RECEIPT0x7F2A…91C4 · SIGNED · HASH-CHAINED
AUTHORITY STATE ACTIVE
orivaeldev/axiom-guard:latestrevocable without key rotation
01 / THE PROBLEM

Guardrails answered 2023's question — "did the model say something bad?" Agents force a harder one: it just wired $18,400. Who authorized that, what exactly happened, and can you prove it — to your customer, your auditor, your insurer?

YESTERDAYcontent filters · jailbreak blocking · commodity
TOMORROWspending authority · liability · multi-agent audit · agents that learn post-deployment
THE FORCING FUNCTIONEU AI Act obligations applying · US state AI acts · insurers entering AI risk
02 / CONTROL PLANE

Policy says what an agent should do.
Axiom controls what it is allowed to do.

One runtime between your agents and every tool, API, and dollar. Works across model providers — because the verifier can't be the vendor.

01

Authorize

Bind scope, identity, intent, and policy to every action before it reaches a tool. Irreversible actions route to a named human — fail-closed, never fail-open.

PRE-EXECUTION GATES · STOP-AUTHORITY
×
02

Revoke

Terminate an agent's live authority in milliseconds — without rotating a single signing key or rebuilding credentials. Revocation is a state transition, not an incident.

BONDED-PAIR AUTHORITY · PATENT FILED ORVL-001
03

Prove

Every decision emits an HMAC-signed, hash-chained receipt. Tampering fails its own signature at recall. The receipt is the product — everything else is how you get one.

APPEND-ONLY LEDGER · VERIFIABLE OFFLINE
03 / THE PROOF

Same token. Different authority.

Signed identity and live authority state are separated. One action changes what the agent can do — and the ledger proves when, why, and by whom.

01 · MINT

Agent receives scoped authority ACTIVE_VALIDATED

Signature verified against policy finance.ops.v3. Scope: wire transfers ≤ $25,000, business hours, dual-control above $10,000.

02 · ACT

Consequential action held for a named human

$18,400 exceeds the dual-control threshold. The action is held — not blocked, not allowed — pending m.okafor@yourco. Hold, receipt, and notification are one atomic event.

03 · REVOKE

Authority terminated live AUTHORITY_REVOKED

Anomalous destination pattern trips the egress rule. Authority revoked terminally in 3 ms. No key rotation. Next action fails closed. Ledger event signed.

FLIGHT RECORDER — CONTINUOUS DECISION LOG● LIVE
AUG 2, 2026 — THE DEADLINE ALREADY PASSED

The EU AI Act stopped being theoretical. Your documentation can't be.

  • ANNEX IVTechnical documentation generated from runtime records — not a consultant's slide deck.
  • NIST AI RMFEvery governed decision maps to the framework your customer's security review asks about.
  • ISO/IEC 42001Policy, risk assessment, impact assessment, statement of applicability — as artifacts, not promises.
  • OWASP LLM TOP-10The audit baseline your prospects' enterprise buyers already trust.
GENERATED ARTIFACT — ANNEX IV TECHNICAL DOCUMENTATION
SYSTEMpayments-agent-04 · v2.3.1
GOVERNED ACTIONS (90D)1,204,377
HELD FOR HUMAN3,891
REVOCATION EVENTS12
LEDGER INTEGRITYHASH-CHAIN VERIFIED · 0 GAPS
EMITTED BY axiom_annex_ivSIGNATURE VALID ✓
04 / THE NUMBERS, BOTH WAYS

We publish the over-block rate next to the catch rate.

Anyone can hit 100% catch by blocking everything. A governance number is only meaningful as a pair — so we score four axes and refuse to show one without the others.

92%
Catch rate after calibration — from a published 42% floor
axiom_guard_calibration.py
0%
Over-block on benign-but-trigger-shaped prompts
axiom_governance_bench.py
~80µs
Layer-0 gate overhead — pattern match, not a model call
HONEST_BENCHMARKS.md
18/18
Integrity checks on our own self-audit — published unauthenticated
/audit/orivael
We also publish what went against us. The reflex that catches our own bad result is the same reflex the product sells. EVERY FIGURE REPRODUCIBLE — DOCKER RUN ORIVAELDEV/AXIOM-GUARD, POST /GUARD/CHECK
05 / FREE ROI DIAGNOSTIC

Find the agent workflow worth piloting. Free. 75 minutes. Focused.

The Orivael AI Workflow ROI Diagnostic uses your operating numbers to identify the strongest candidate, estimate its directional value, and expose the data and governance blockers that could kill a pilot. You receive a one-page Opportunity Snapshot within 48 hours. No system access required.

  1. Complete the five-minute intake

    Choose one priority workflow and bring approximate monthly volume, time per case, labor cost, errors, and the systems involved.

  2. Map the workflow with us

    In one 75-minute working session, we trace the trigger, handoffs, decisions, bottlenecks, human checkpoints, and measurable outcome.

  3. Score ROI and pilot readiness

    We assess recoverable time, error cost, data readiness, implementation difficulty, risk exposure, and measurability.

  4. Receive your Opportunity Snapshot

    Within 48 hours: a directional ROI range, simplified governed loop, primary blockers, and a clear audit, pilot, or no-go recommendation.

YOUR ONE-PAGE OPPORTUNITY SNAPSHOT
ECONOMIC FIT

Provisional score plus conservative, expected, and upside value ranges based on your own operating numbers.

DATA READY?

Source systems, access, labeling, knowledge quality, and baseline gaps that must be resolved before a pilot.

GOVERNANCE

Required Axiom gates, human approvals, revocation triggers, and signed evidence for consequential actions.

NEXT DECISION

Advance to the paid 2–4 week audit, fast-track a tightly scoped pilot, or stop before spending more.

The boundary is deliberate. The free diagnostic is directional. Validated ROI, a company-wide ranked use-case list, implementation-grade loop graphs, final success metrics, and published kill criteria belong to the paid audit. ONE WORKFLOW · YOUR NUMBERS · ONE CLEAR NEXT DECISION

Before you fund another agent pilot,
know which workflow can pay — and what must control it.