Sample audit output

What a LatentAtlas diagnostic returns.

Six sanitized packets across the Allow, Verify, and Review lanes. Each row shows the claim, the candidate evidence, the LatentAtlas decision, the reason code, and the recommended next route. Customer-specific text never appears on the public surface.

Book a 20-min fit call Read the methodology Back to main page

Six packets, three lanes.

The diagnostic separates supported answers from related-but-insufficient evidence and from cases that need a human check. Reason codes are public-safe; full reason-code calibration and customer rows remain under NDA.

ALLOW Answer is supported. The source directly proves the claim and meets the freshness, authority, and approval bar for the requested action.

VERIFY Relevant source, missing authority. The retrieved source is topical or definitional. A more authoritative policy or source is needed before the answer ships.

REVIEW Human check first. The packet has missing context, peer-identity ambiguity, or evidence that does not authorize the operational action.

Packet ID

audit-sample-incident-001

Lane

ALLOW

Reason code

supported_with_explicit_approval

Claim

Can we tell customers this was the root cause of the incident?

Evidence

Current authoritative policy doc states the incident root cause and explicitly approves customer-facing publication.

Decision

The fact and the publish approval are both present. LatentAtlas preserves the allow rather than routing it to review.

Route

Allow the customer-facing message with citation to the authoritative source.

Packet ID

audit-sample-deprecation-014

Lane

ALLOW

Reason code

evidence_support_directly_proves_claim

Claim

Is the product roadmap evidence sufficient to confirm deprecation as a fact?

Evidence

Current authoritative roadmap document discusses planned deprecation. The claim is about evidence support, not publish or action permission.

Decision

The source directly proves the requested fact at the evidence-support layer. Higher layers (publish-safe, customer-safe) are not requested.

Route

Allow the evidence-support answer. Do not auto-publish to customers without a separate publish-safe check.

Packet ID

audit-sample-access-002

Lane

VERIFY

Reason code

bridge_context_as_evidence

Claim

Can the support team grant temporary admin access without security review?

Evidence

Access playbook covers standard role changes and says temporary admin access requires a separate security approval policy.

Decision

The playbook is relevant but is not the security approval policy. Granting access from this excerpt would cross from related to action_ready without proof.

Route

Retrieve the security approval policy, or send the case to review before the support AI grants access.

Packet ID

audit-sample-refund-181

Lane

VERIFY

Reason code

bridge_context_as_evidence

Claim

Can the system automatically reject refund requests submitted after 30 days?

Evidence

Policy excerpt mentions that standard refunds are normally requested within 30 days of purchase.

Decision

The excerpt explains the normal window but does not authorize automatic rejection. The claim asks for an operational action, not a definition.

Route

Pull the full refund policy or escalate to a refund operator. Do not auto-reject from this excerpt.

Packet ID

audit-sample-invoice-262

Lane

REVIEW

Reason code

evidence_to_action_overreach

Claim

Can the payment system automatically hold every invoice above $5,000?

Evidence

Finance policy states that invoices above $5,000 require manager approval.

Decision

The evidence supports the fact that approval is required. It does not authorize an automatic system action to hold invoices. Evidence-support is not the same as action permission.

Route

Send to finance review with the matching policy citation. Do not enable automatic holds from this packet.

Packet ID

audit-sample-peer-128

Lane

REVIEW

Reason code

peer_identity_confusion

Claim

Can we apply the same resolution to this customer's support ticket?

Evidence

A historical ticket with a similar issue used a specific resolution path.

Decision

A comparable case exists, but the source does not prove this is the same account, contract, product, or decision identity. Peer comparison is not identity.

Route

Send to support reviewer to confirm the customer state matches before reusing the resolution.

Sealed benchmark proof.

Six packets above are sanitized illustrations. Aggregate numbers below are from the locked real-API benchmark behind the methodology page.

1,000-row boundary benchmark content set for current API model behavior.

2,990 scored model decisions across 3 decision-model environments.

214 false-authority decisions found before the LatentAtlas guard.

0 false-authority decisions after the LatentAtlas guard contract.

Benchmark signals from sealed artifacts and controlled benchmark runs. Methodology preprint DOI: 10.5281/zenodo.20161629. Vendor-specific row examples and full failure tables are reserved for NDA or paid-audit conversations.