A tamper-proof receipt for
every AI decision.

Your AI agents approve loans, triage patients, and settle claims. OIA proves each decision was made correctly — or flags it wasn’t.

$20K
per violation, per consumer
Colorado AI Act · Jun 2026
300K
claims auto-denied, no proof
UnitedHealth class action · 2024
136 days
until EU AI Act enforcement
Art. 12 mandatory logging
Without OIA
  • ×AI made a decision — you have a log line
  • ×Regulator asks for proof — you have a spreadsheet
  • ×Lawsuit hits — you have “we followed our process”
With OIA
  • ✓Every decision has a settlement receipt
  • ✓Cryptographic proof: what it saw, decided, and why
  • ✓Every decision is SETTLED or explicitly UNSETTLED
terminal
# Settle an AI decision in one call
$ curl -X POST https://api.oia-lab.com/v1/settle \
-H "Authorization: Bearer $OIA_KEY" \
-d '{"agent": "claims-bot", "decision": "approve", "evidence": [...]}'
# Response:
{"status": "SETTLED", "receipt": "ST-2026-0319-A7F3",
 "delta1": {"c1": true, "c2": true, "c3": true},
 "hash": "sha256:9f86d0...a7f3e2"}

3 lines. One receipt. Tamper-proof. Independently verifiable.

Try the Demo View Pricing

Start free. Settlement receipts from $49/mo.

Everyone logs.
Nobody settles.

Observability tools record that your AI did something. They don’t prove it was done right. A log is a diary. A settlement receipt is a legal document.

A dashcam that keeps secrets.

AI autopilots will make millions of decisions a day. Three things must be true about every one.

Protect

Your data is governed at the architectural layer, not the policy layer. No single vendor — and no single internal component — sees enough to reconstruct full intent.

Record

Every decision — what the AI saw, what it decided, and how it was verified. Complete chain.

Seal

Cryptographic proof. Tamper-proof. Independently verifiable. Every decision is settled or explicitly unsettled.

That’s the bar. Now here’s what we found when we tested whether anyone actually meets it.

Empirical Evidence

We tested the “Protect” layer.

Defense effectiveness under adversarial stress testing — 18,232 trials across 6 frontier LLMs.

Finding 1

The most widely deployed AI data protection layer performed worse than no defense in our benchmark.

Token-level redaction scored below the undefended baseline across every condition tested.

Finding 2

Different data classes require different defense classes.

Explicit identifiers can be neutralized to zero extraction under proper defense. Other signal classes require architecturally distinct treatment — covered by a separate layer in our stack.

Finding 3

A different defense approach exceeded the baseline — the only method that passed adversarial stress testing.

Details available under NDA for enterprise evaluation.

18,232 trials  |  6 frontier model families  |  3 regions  |  24 conditions

Class Action · 2024

It’s Already Happening

An AI system auto-denied ~300,000 claims. Override rate ~90%. Class action filed.

UnitedHealth had AI. They had processes. They did not have decision settlement.

What was missing:

No evidence trail per decision
No closure verification on denials
No adversarial-tested defense layer
Regulatory Deadlines

The Clock Is Running

EU AI Act Art. 12

Mandatory event logging for high-risk AI. Extraterritorial. August 2, 2026.

Colorado AI Act

$20,000 per violation per consumer. Effective June 30, 2026.

D&O Insurance Exclusions

WR Berkley: absolute AI exclusion. Verisk CGL exclusion forms filed. Directors are personally exposed.

No evidence infrastructure → no compliance → no insurance → personal liability.

Listen · 90 sec
Settlement in 90 seconds
0:00

DTCC settles $2.5 quadrillion in financial trades per year. Nobody settles AI decisions.

One receipt per decision.

Who decided. Based on what. What happened. Whether responsibility is closed.

Every agent decision either reaches settlement finality — or stays explicitly unsettled. Try the interactive demo →

settlement-viewer
SETTLEMENT RECEIPT
────────────────────────────────────────────
Settlement ID ST-2026-0208-A7F3 Session sess_8f3a2b1c4d5e Agent procurement-agent-03 Timestamp 2026-02-08T14:23:07.442Z TSA RFC 3161 (pluggable)
Settlement Trace (12 steps, hash-chained)
#01 tool_call vendor_api.search(category="raw_materials") ✓
#02 inference  compare_quotes(n=4, criteria=[price, lead_time]) ✓
#03 tool_call compliance.check(vendor_id="V-8821") ✓
#04 tool_call approval.request(amount=42500, approver="cfo") ✓
... (8 more steps)
#12 action    purchase_order.submit(po="PO-2026-1847") ✓
Δ1 Settlement Check
✓ C1 Evidence Consumed 12/12 steps sealed
✓ C2 Intent Isolated verified isolated
✓ C3 Settlement Signed authorized by cfo@acme
Hash: sha256:9f86d0...a7f3e2
Δ1 = C1 ∧ C2 ∧ C3 SETTLED

You can’t cryptographically seal a maybe.

Deployed. Benchmarked. Documented.

This is running infrastructure, not a roadmap. Full methodology published — we’d rather you verify than take our word for it.

12 patents pending
10 published papers
18,232 adversarial trials
Infrastructure
  • 3 Docker services, 23 REST endpoints
    On-premises deployment with your AI keys
  • Air-gapped deployment option
    For environments where data cannot leave the building
  • Benchmarked across 5 regulated verticals
    Finance, healthcare, legal, government, insurance
Architecture →
Research
  • 10 published research papers
    AI decision accountability, settlement theory, defense validation
  • FLEET Benchmark: 18,232 trials
    6 frontier AI models, reproducible methodology
  • 354+ patent claims across 6 layers
    Full settlement stack coverage, priority dates locked
Publications →
Category Gap
  • “Decision Settlement”
    No directly comparable settlement architecture identified in our patent review. Priority dates locked 2026.
  • “Cognitive Leakage”
    Foundational definition published — no prior art found in patent database review
  • NIST AI RMF Contributor
    OWASP Agentic AI Top 10 mapped
Compliance mapping →

Category Scope

Settlement became mandatory infrastructure
once before.

After the 1987 crash, unresolved financial obligations became unacceptable. DTCC was built. Settlement became infrastructure. Every trade goes through it. AI decisions are following the same path — but faster, because regulators aren’t waiting for a crash.

Forcing Function
Regulatory deadlines
EU AI Act + Colorado AI Act create legal obligation to prove each high-risk decision was made correctly. The obligation exists with or without infrastructure.
Scale of Exposure
Every regulated enterprise
Finance, healthcare, insurance, legal. Any company using AI for consequential decisions in the EU or US is in scope. Extraterritorial reach confirmed.
Infrastructure Gap
No settlement layer exists
Observability tools log events. Governance tools set policies. No existing product closes the loop on whether a specific decision was properly settled.

Regulations define obligations.
We define done.

Try the interactive demo. Review the architecture. Reach out when you’re ready.

Try the Demo Request Pilot Access

Designed for regulated industries — finance, healthcare, and legal.

or email yc@oia-lab.com

YC Chang Founder & CEO Jack Cheng Co-Founder · Co-Founder of NIO (NYSE: NIO)