GAIIB™

GoverneveryAIdecisionbefore it ships.

GAIIB™ checks the inputs, the model output and the final decision of your AI system against the regulations and SOPs you map. Every call returns a typed outcome and leaves evidence an auditor can verify.

Start developer sandbox Book an architecture review
  • Fail-closed by default
  • One REST call to integrate
  • Evidence for every decision
  • Human review built in
POST /api/v1/verify
RequestPRE_COMMIT
{
  "system": "Retail Lending Model",
  "payload": {
    "creditScore": 742,
    "monthlyIncome": 9200,
    "loanAmount": 180000,
    "bankDecision": "APPROVE"
  }
}
Evaluation
  1. L0 · Evidence cache
  2. L1 · Policy context
  3. Deterministic rules
  4. L2 · Risk classifier
  5. L3 · Statutory audit
  6. L4 · Consensus check
  7. Evidence sealed
Evaluating pre-commit checkpoint…

Illustrative simulation of the decision pipeline with sample data.

Map decisions to
EU AI Act (2024/1689)DPDP Act 2023RBI Digital Lending DirectionsISO/IEC 42001:2023NIST AI RMF 1.0GDPRISO/IEC 27001SOC 2ECOA / Reg BYour internal SOPsEU AI Act (2024/1689)DPDP Act 2023RBI Digital Lending DirectionsISO/IEC 42001:2023NIST AI RMF 1.0GDPRISO/IEC 27001SOC 2ECOA / Reg BYour internal SOPs
Why decision governance

Automated decisions need more than a content filter.

Credit, insurance and hiring models make consequential calls about people. The hard part is proving each one followed the rules you are accountable for.

Decisions nobody can explain

A model declines an applicant and the only record is a score. Regulators ask for the reason, the rule and the data used.

Silent approvals on failure

When a provider times out or a policy is missing, many gateways fall through to "allow". In lending that is a booked loan.

Evidence assembled after the fact

Audit trails are rebuilt from logs weeks later, without the policy version or model that actually made the call.

Where ZeroCore sits

Three checkpoints across the decision lifecycle.

Call ZeroCore before data reaches a model, after the model answers, and right before your system commits the outcome. All three share one correlation id.

  1. Application
    Your product collects features
  2. PRE_INFERENCE
    Pre-inference
    Is this input allowed to reach a model?
  3. Your model
    Scoring, LLM or rules engine
  4. POST_INFERENCE
    Post-inference
    Is the model's output defensible?
  5. PRE_COMMIT
    Pre-commit
    May the final decision be committed?
  6. Committed
    Loan booked with an evidence trail
How checkpoints work
Capabilities

Policy first. Models second. Evidence always.

Deterministic rules run before any model is called. The LLM audit layers explain and challenge. Everything that happened is recorded with the exact policy and models used.

Deterministic lending policy

Prohibited features are denied, missing required features return insufficient evidence, sensitive attributes go to review. Instantly, with no model in the loop.

creditScoremonthlyIncomeloanAmountemploymentYearsmaritalStatusreligiondebtToIncomegendercaste
required → else INSUFFICIENT_EVIDENCEprohibited → DENYsensitive → REVIEWno model called

Decision evidence, fingerprinted

Each decision stores its outcome, checkpoint, engine version, model ids and a fingerprint of the policy that applied. Change a rule and old cache entries stop matching.

Decision evidencesample
  • INSUFFICIENT_EVIDENCE
    Gold Loan · PRE_COMMIT
    sha256 26ca023f6698975f-1f5…
  • ALLOW
    Collections Bot · PRE_INFERENCE
    sha256 04becb81-6114483238b…
  • DENY
    Retail Lending · PRE_INFERENCE
    sha256 -743c558c76a3733e-3a…
  • REVIEW
    SME Credit · POST_INFERENCE
    sha256 0eda17692dcc2dc62f2b…

Two-person exceptions

Reviews are separate, sealed records. Approving an exception needs a second auditor.

  1. Decision held
    Outcome REVIEW, original row is never edited
  2. Reviewer A
    Proposes an exception with written justification
  3. Reviewer B
    A different auditor must confirm
  4. Sealed record
    Keyed HMAC seal + signed webhook

Fail-closed errors

Outages and timeouts return typed errors. Try one:

HTTP/1.1 503
{ "outcome": "ERROR",
  "code": "DEPENDENCY_UNAVAILABLE",
  "action": "NOT_EVALUATED" }

Retryable. No decision is fabricated.

Signed webhooks

Review outcomes and alerts arrive signed with HMAC-SHA256 and a timestamp to stop replays.

POST https://lender.example/hooks/zerocore
X-ZeroCore-Event: decision.review_resolved
X-ZeroCore-Signature: t=1761818400,v1=5d10491d5512e7c276ae6192-50f5527
HMAC-SHA256 verified against raw body

Your SOPs as policy context

Upload underwriting manuals and internal policies. The most relevant passages are retrieved by similarity for every decision and cited in the reasoning.

Idempotent and traceable

Retry with the same Idempotency-Key and get the stored decision back without being billed twice. Correlation ids tie every checkpoint of one loan together.

Typed outcomes

One answer your code can act on.

No ambiguous scores. Every evaluation ends in exactly one outcome, and only one of them means proceed.

ALLOWHTTP 200

A configured policy was evaluated end to end and passed. The only outcome that lets your workflow proceed.

DENYHTTP 400

A deterministic rule or the audit layers found a violation. Reason, citation and evidence are recorded.

REVIEWHTTP 400

Sensitive attributes, layer disagreement or low confidence. The case goes to the Auditor Desk for a human decision.

INSUFFICIENT_EVIDENCEHTTP 400

Required features or the model output are missing, so no decision can be justified.

POLICY_NOT_CONFIGUREDHTTP 400

No framework or SOP is mapped to the system. ZeroCore refuses rather than guessing.

ERRORHTTP 5xx

A dependency failed or timed out. A typed error is returned and nothing is ever approved by default.

Integration

Live in an afternoon, not a quarter.

ZeroCore is a REST API. No SDK lock-in and no proxy rewiring: add a call at each checkpoint and branch on the outcome.

01

Map your policy

Attach frameworks and upload SOPs to an AI system. Set prohibited, required and sensitive features.

02

Call one endpoint

Send the payload to /api/v1/verify at the checkpoint you need, with an idempotency key and your correlation id.

03

Act on the outcome

Proceed only on ALLOW. Route REVIEW to people, stop on everything else, and keep the decision id.

Read the API docsVerify webhooks
const res = await fetch("https://gaiib.com/api/v1/verify", {
  method: "POST",
  headers: {
    Authorization: `Bearer ${process.env.ZEROCORE_API_KEY}`,
    "Content-Type": "application/json",
    "Idempotency-Key": crypto.randomUUID(),
    "X-Correlation-Id": loan.id,
  },
  body: JSON.stringify({
    systemId: process.env.ZEROCORE_SYSTEM_ID,
    checkpoint: "PRE_COMMIT",
    payload: loan.features,
    output: { decision: bank.decision, reasonCodes: bank.reasonCodes },
  }),
});
const decision = await res.json();

if (res.ok && decision.outcome === "ALLOW") commit(loan, decision.decisionId);
else if (decision.outcome === "REVIEW") holdForReview(loan, decision.decisionId);
else stop(loan, decision.outcome ?? decision.code);
Public trust badge

Show customers your AI is governed, live.

An embeddable badge that reflects the system's current compliance state and links to a public verification page. It degrades honestly when the score drops.

Configure

Preview a state
HTML embed
<a href="https://gaiib.com/verify/YOUR_SYSTEM_ID" target="_blank" rel="noopener">
  <img src="https://gaiib.com/api/badge/YOUR_SYSTEM_ID"
       width="296" height="48"
       alt="Retail Lending Model governed by ZeroCore" />
</a>

Find your system id in the dashboard under AI Systems.

Badge preview
ZEROCORE™ COMPLIANTRetail Lending ModelLIVE

Compliance score at or above 95% with no open alerts.

Clicking the badge opens

A public verification page for the system with its current status and certificate, without exposing any decision data.

0
Decision checkpoints
pre-inference · post-inference · pre-commit
0
Typed outcomes
only ALLOW means proceed
0
Reviewers per exception
a different auditor must confirm
0h
Idempotency window
safe retries, no double billing
Zero approvals are ever returned without a completed evaluation.
Pricing

Start free. Scale with your decisions.

Every plan includes the full fail-closed pipeline, typed outcomes and decision evidence.

  • Same fail-closed pipeline on every plan
  • Idempotent retries are never billed twice
  • Tenant-isolated decision evidence
  • Annual billing and invoicing available

Knowledge Base

How ZeroCore™ behaves in production: integration, failure modes, evidence and human review.

zerocore · knowledge-base
Query: How does ZeroCore™ integrate with an existing AI pipeline?

Put a governance checkpoint in front of your next decision.

Create a sandbox, map a framework, and run your first governed decision today. Or walk through your architecture with our team.

Start free Book an architecture review