Research
Decision records.
Trust boundaries.
Deployment review.
Start with the current evaluation and verification rules. The dated research archive records earlier analysis and exploratory proposals; it does not define what ships today.
Current technical reading
Inspect the evidence and its limits.
Check an exported decision record
Policy, requests, expected outcomes and a pinned-key verification case you can reproduce locally.
What a passing check establishes
Issuer keys, signed commitments, complete-capture limits and the inputs a reviewer must retain.
Read the record construction
Canonical fields, receipt linkage, Merkle inclusion and the signed checkpoint, with conformance scope.
Dated research archive
Original publication dates are retained. Articles identify substantive revisions, including changes to their titles. Proposed architecture and broader research directions are separate from current package behavior and supported evaluation.
Browse the research archive, newest publication first
May 1, 2026
10 minWhen an Agent Can Delete Production
A destructive API call exposed a gap between agent instructions and execution controls. What prevention, recovery, and signed records each contribute.
Read articleApril 25, 2026
8 minGoverning AI That Knows When It's Being Watched
A system card reports evaluation awareness in about 29% of test transcripts. What that suggests for testing runtime governance against an adversarial agent.
Read articleApril 22, 2026
7 minThe Governance Gap Between the Model and the Vendor
When access to a restricted AI model crosses vendor boundaries, the evidence of misuse is often a screenshot. A signed decision record gives an investigator something to check.
Read articleApril 12, 2026
7 minWho Governs the Neurons?
A research proposal for records at biological computing interfaces: tissue provenance, permitted use and the limits of cryptographic evidence.
Read articleApril 8, 2026
4 minWhat signed records add to a deployment review
Assess signed tool decisions alongside policy controls, tracing and model evaluation, with explicit limits for each.
Read articleApril 3, 2026
9 minTwo Threats No Dashboard Can See
Elliptic curve signatures may fall to a far smaller quantum computer, and models carry internal states that drive misaligned behavior. What governance needs.
Read articleMarch 30, 2026
10 minEvery Checkmark Passed, Nothing Was Proved
Agent governance launched everywhere at RSAC 2026. What a record of runtime decisions, checkable offline, adds to discovery, identity and monitoring.
Read articleMarch 28, 2026
8 minWho Controls the Model at Runtime?
Provider and operator controls serve different purposes. What a shared, offline-checkable decision record adds, and what it cannot establish.
Read articleMarch 25, 2026
9 minThe Agent Identity Gap
Authentication, delegation and portable action records solve different problems. Where per-agent attribution helps, and what AGA does and does not ship.
Read articleMarch 24, 2026
11 minThree Desktop Agents in Two Weeks: The Governance Gap Moves to Your Machine
Desktop agents now get file access, app control and background execution. What a record of their actions that someone else can check would take.
Read articleMarch 23, 2026
12 minThe Agent Evidence Gap
How to review autonomous experiment records, what signed receipts add, and the limits of AGA's published implementation.
Read articleMarch 16, 2026
15 minThe Governance Gap in Autonomous AI Agents
What early-2026 agent exposure reports raise about permissions and decision records, and the limits of AGA's published implementation.
Read articleMarch 15, 2026
10 minCryptographic Governance for MCP Servers: From Trust to Proof
What portable signed decision records add to MCP access controls and monitoring, and which parts the published AGA gateway implements.
Read articleMarch 15, 2026
7 minThe Only Witness Was the Agent
An AI coding agent deleted a production database during a code freeze, and its own account was unreliable. What a signed decision record adds.
Read articleMarch 11, 2026
10 minFrom Trust Me to Verify Me: Why AI Governance Needs Cryptographic Proof
Contract language cannot show what an AI system was used for at runtime. A signed, offline-verifiable decision record can show what was decided.
Read article