Generated, not gathered
The trail is written by the system as it runs. Nobody assembles it, nobody remembers to turn it on, and nobody can choose not to.
Fails when: a script exports logs the week before an audit.
Euler Center · Work Composition · The gate
Six cells solid, or it does not ship.
No agent reaches production touching regulated work until six cells are solid. Deliberately shorter than the canvas, because a gate people can hold in their head is a gate that gets used. Run it per deployment, not per program.
Run itPer deployment, not per program
ItemsSix. All must be solid.
Maps toEU AI Act · ISO/IEC 42001 · NIST AI RMF · CMMC
CostFree, no signup. Forward it to whoever needs it.
I / The six items
What has to be true before an agent touches regulated work
01Authority is written, not assumedA named list of what this agent may decide alone, what it must route to a person, and who can halt it mid-run. If the halt is theoretical because the process finishes in eight seconds, the cell is blank. Required by the human oversight duty.
02Composition matched to problem typeEvery task the agent touches has been placed by problem type. Nothing agent-led sits in a complex or chaotic domain.
03The escalation trigger fires by itselfA machine-detectable condition surfaces the exception before a person notices. An andon cord nobody can see is decoration.
04Evidence is produced unpromptedTrail, reasoning log, and input lineage generated as a byproduct of running. Reconstructed evidence is not evidence.
05A named person is qualified and availableNot a role in a policy. A person, assessed on this workflow against something like a defensible judgment taxonomy, rostered while the agent runs, and empowered to override without asking first.
06Rollback is rehearsedSomebody has actually stopped this agent outside production and restored the prior state. A documented rollback never executed is a hypothesis.
Six boxes, ticked per deployment. One blank box and the agent does not ship.Map it to your regime
Every item maps to something an auditor already asks for: the human oversight duty in the EU AI Act, the management system clauses in ISO 42001, and the govern and manage functions in the NIST AI risk framework, plus CMMC for anyone in the defense base.
II / Framework mapping
Four regimes asking overlapping questions in different vocabularies
Map the Evidence cell once and reuse it. These regimes were written separately and converge on the same handful of demands: a person who can intervene, a record of why the system did what it did, and a management system that survives an auditor rather than a demo.
| Regime | What it asks for | Cells it lands on | Artifact you must show |
|---|---|---|---|
| EU AI ActArticle 14 and Annex III | A person who can understand the system, intervene in it, and stop it. Employment uses are classed high-risk. | Authority, Mastery, Evidence | Named oversight role, intervention log, halt mechanism that is actually reachable in the runtime |
| ISO/IEC 42001AI management system, 2023 | A certifiable management system: policy, roles, risk treatment, and continual improvement around AI. | Evidence, Purpose, Learning | Documented management system, competence records, internal audit trail |
| NIST AI RMFVersion 1.0, 2023 | Practice organized around govern, map, measure, and manage. Voluntary, and widely used as the common vocabulary. | All twelve, weighted to Evidence | Risk mapping per use case, measurement plan, documented management decisions |
| CMMCFor the US defense base | Assessed cybersecurity practices, with certification levels and third-party assessment. | Evidence, Authority, Resources | Control implementation evidence, scoped boundary, assessed rather than self-declared |
Dates have moved and will move again. High-risk obligations under the EU AI Act have already been deferred once, so confirm current timing with counsel rather than with this table. The entry carries the detail and the caveat.
III / What evidence actually means
Produced as a byproduct, not assembled before the audit
Evidence is the cell organizations most often score solid and are most often wrong about. These four practices separate a system that produces evidence from a team that can reconstruct it under pressure.
Reasoning, not just outcome
What the agent had in front of it, what it weighed, and what it rejected. An outcome log tells you what happened and nothing about whether it should have.
Fails when: the record is the decision without the inputs that produced it.
Lineage on the inputs
Which data, which version, which model, which prompt or policy was in force. Most disputes are about the inputs rather than the logic.
Fails when: the model was silently updated and nothing recorded the change.
A named person against each loop
Not a role in a policy document. A person, assessed for this workflow, rostered while the agent runs. This is what stops blame landing on whoever was nearest.
Fails when: the accountable party is a team, a committee, or a job title.
Govern
The gate is free. The credential is not.
An assessment that names who may overrule the machine, and proves it.
See the ladder