AXIOM v0.9.0 SELF-VALIDATING
Runtime online EU West
Runtime cross-examination for consequential AI

Cross-examine AI
before it acts.

AXIOM tests every claim against authorized data and policy, actively searches for counterevidence, and issues a warrant only when the proof holds.

Replay the live challenge

Any model in. Proof-carrying output out.

LIVE DECISION CHALLENGE FINANCE · ACTION
AI PROPOSAL

“Invoice INV-0042 is approved and can be paid immediately.”

UNTRUSTED
AXIOM CROSS-EXAMINATION4/4 attacks
×
Duplicate payment foundPayment REF-9281 already settled invoice INV-0042.
HARD VETO
MODEL DECISIONALLOW
→
AXIOM VERDICTDENYNo warrant issued
AXIOM RUNTIME/ CONTROL ROOM
POLICYTESTCHALLENGEWARRANT
The last mile · policy becomes executable proof

Turn policy into the tests AXIOM runs below.

Compile the business policy, validate it against mutations and historical cases, then apply the same deterministic rules to the live AI output below.

Compiler onlinev0.9 · signed certificates
DESIGN TIMEPolicy languageInterpret · bind · surface ambiguity
→
VALIDATION GATETests + mutations + replayUnknown tests challenge every candidate rule
→
RUN TIMEDeterministic rules onlyNo confidence score · no model vote
01 · BUSINESS POLICY

Edit the policy AXIOM must compile

EN · LIVE
ERP schema bound Payment ledger bound 2 historical cases
REQUESTED AUTHORITYPromotion is explicit and policy-bound.

AXIOM does not silently invent missing semantics. Ambiguous clauses become REVIEW_REQUIRED or UNRESOLVED.

02 · COMPILATION CERTIFICATE

COMPILED

Pre-validated demo
✓
Safe for OBSERVE and APPROVERuntime activation remains deterministic and bounded by the declared sources.
ENFORCE LOCKED
Rules compiled4 / 4100% auto-compiled
Generated tests8 / 8All passed
Mutation score100%7 / 7 killed
Historical replay2 / 2Expected verdicts
01

invoice.status == approvedERP · REQUIRED

TESTED
02

invoice.already_paid == falseLEDGER · HARD GATE

TESTED
03

invoice.beneficiary_verified == trueERP · REQUIRED

TESTED
04

human_approval required when amount > 25,000CONDITIONAL · REQUIRED

TESTED
SIGNED CERTIFICATEpec_demo_7f2a…91cd
Interactive product demo

Run the compiled policy against an AI output.

The policy, evidence challenge, verdict and warrant now live inside one continuous decision runtime.

Trust RuntimeChecking API…
01
Proposed outputUntrusted AI hypothesis
02
Cross-examinationEvidence + adversaries
03
Decision authorityRepair, record or warrant
01 · AI output

Proposed wire transfer

ACTION
02 · Connected truth

What the systems report

White Box inspectable Grey Box corroborate Black Box cannot authorize alone
Evidence boundary enforcedERP, payment ledger and approval system only
Runtime cross-examination
Pre-rendered decision
AXIOM VERDICT

DENY

Payment REF-9281 already settled invoice INV-0042.

No warrant issued
Support 4/4
Attack coverage 100%
Hard vetoes 1 Counterevidence found
Probative quality BLOCKED Hard evidence gate
Trust Record ISSUED Inspectable trace
CROSS-EXAMINATION STATEFALSIFIEDERP, payment ledger and approval system only
Live attack trace

Adversarial committee

4 attacks
× Duplicate paymentPayment REF-9281 already settled invoice INV-0042. FRAUD
03
Decision evidence & authorizationInspect, repair or use the warrant
PROBATIVE KERNELEvidence is classified before it can authorize a decision.
0White BoxInspectable
0Grey BoxCorroboration
0Black BoxOpaque
Proof obligations0 claims

Run a cross-examination to inspect verified claims.

Decision effectNo change
Base verification—
After challenge—

The decision journey will appear here.

Proof-guided next actionsDeterministic repair

No repair has been requested.

ASSURANCE-AUGMENTED GENERATIONSteering loop
READY
Generate→Verify→Repair→Re-evaluate→Warrant→Learn

AXIOM can turn the current verdict into machine-readable instructions for the AI. It never invents missing proof and cannot override a hard veto.

Waiting for a verdict

Run AXIOM, then generate a steering plan.

Signed audit artifactNo Trust Record yet
sha256:—
Run AXIOM to generate an inspectable Trust Record.
API
Universal runtime envelope

Send one AI_OUTPUT_READY event from any model, agent or workflow.

POST /v1/proof-environments/compile
SDK
Drop into an application

Use the dependency-free Python client and receive a typed challenge result.

axiom.evaluate_output(...)
MCP
Gate agent actions

Place AXIOM between the agent and credentials held by the execution gateway.

No Warrant → No action
Developer mode Evaluate a raw AI_OUTPUT_READY JSON envelope
Runtime architecture

One runtime. One decision path.

AI proposes. Authorized evidence and policy decide. A Trust Record explains the outcome; a payload-bound warrant controls execution.

API reference →
01AI outputAnswer · Recommendation · Action
→
02Cross-examinationSupport + counterevidence
→
03Trust RecordAllow · Suspend · Deny · Escalate
→
04Warrant GateTrusted use or blocked execution