Six filings.
One standard.

Abstract

A system and method for capturing cryptographically sealed evidence records of AI agent actions that render into multiple regulatory formats — SOC 2 dossiers, EU AI Act Article 12 logs, FRE 902(13)/(14) certification packets, and insurer underwriting submissions — without rework. One sealed record satisfies simultaneous compliance with auditing, regulatory, evidentiary, and insurance standards.

In plain English

Verdict captures what autonomous AI agents do at runtime and seals it cryptographically so the evidence cannot be modified. The sealed record is smart enough to render into whatever format regulators, courts, insurers, and auditors need — without rework.

Core invention pillars

• ◈Closed taxonomy of canonical AI agent event types
• ◈Content-addressed policy bundle versioning via SHA-256
• ◈RFC 6962 Merkle trees + Sigstore Rekor anchoring + Ed25519 signing
• ◈Hash-preserving redaction (GDPR Article 17 compliance without breaking proof)
• ◈Evidentiary Completeness Score (ECS) → FRE 902(13)/(14) certification map

Abstract

Pre-execution verification of autonomous agent intent against a hierarchical policy tree, enforcing graduated gatekeeping — allow, require human review, escalate, deny — at the atomic action level, with cryptographic evidence sealing of every policy evaluation. Non-provisional conversion deadline May 6, 2027.

In plain English

Before an agent acts, its stated intent is decomposed into atomic actions and checked against a policy tree. The check itself is sealed, so the record shows not only what the agent did but what it was permitted to do and who decided that.

Core invention pillars

• ◈Constrained decomposition of agent intent into schema-bounded atomic action primitives
• ◈Hierarchical DAG policy tree evaluated pre-execution, not post-hoc
• ◈Four-state graduated gatekeeping at the atomic-action level
• ◈Every policy evaluation sealed as evidence, including the allows

Abstract

Detecting behavioral drift in autonomous agents at runtime using distance analysis over output embeddings, triggering graduated intervention — alert, fallback, credential revocation, halt — with state preservation across degradation and cryptographic sealing of every intervention. Non-provisional conversion deadline May 7, 2027.

In plain English

When an agent starts behaving unlike itself, something has to catch it in flight — and prove afterwards that it caught it. This is the breaker: it degrades the agent in stages instead of killing it, preserves the work in progress, and seals what it did.

Core invention pillars

• ◈Drift detection via KL / Wasserstein distance over output embeddings
• ◈Four-tier graduated intervention ladder rather than a binary kill switch
• ◈State preservation and resumption during degradation
• ◈Multi-agent cascade isolation with sub-100ms policy-matrix propagation

Abstract

A system in which a human authorization is sealed into a tamper-evident authority record, sub-agent delegation is constrained so that authority can only narrow and never broaden along a chain, and the credential that actually executes the action is derived from the commitment of the sealed authorization itself — single-use, and mintable by no issuer. Non-provisional conversion deadline August 17, 2027.

In plain English

Every other layer records what an agent did. This filing covers binding who stood behind it. The permission is sealed before the action, it can only get smaller as it is handed down, and the key that runs the action is derived from the sealed permission — so there is no authorization server holding a secret that could mint it again.

Core invention pillars

• ◈Sealed Human Authority Record binding a named human decision to a scoped grant
• ◈Pre-execution evidence gate — no authority record, no execution
• ◈Provably narrowing delegation: authority can only contract along a sub-agent chain
• ◈Single-use execution credential derived from the sealed authorization commitment
• ◈Issuer-negative property: no authorization server or credential issuer holds a secret sufficient to mint a credential valid for the delegated action
• ◈Revocation that produces evidence rather than silence

Abstract

A certificate that no autonomous-agent action of a prohibited class occurred, checked against a population of public transparency-log entries that an authority record defined in advance. Non-provisional conversion deadline October 12, 2027.

In plain English

Most evidence proves that something happened. This filing covers proving that something did not: that no action of a forbidden kind ran, measured against a set of log entries fixed before the question was asked.

Core invention pillars

• ◈Certificate of non-occurrence for a prohibited class of agent action
• ◈Checked against a population of entries on a public transparency log
• ◈Population defined in advance by a sealed authority record

Abstract

Selecting which autonomous-agent step records are retained by a public beacon draw, against retention probabilities committed in an anchored record before the draw. Non-provisional conversion deadline October 13, 2027.

In plain English

Keeping every agent step forever is expensive, and letting the operator choose what to keep invites a fight about what was left out. This filing covers a draw: the odds for each step are committed and anchored first, then a public beacon decides what is kept.

Core invention pillars

• ◈Retention probabilities committed in an anchored record before the draw
• ◈Selection by a public beacon draw
• ◈Step records retained according to the draw