Framework overview
PACTE · Integrity assessment methodv0.1 · Draft for public comment · 2 October 2026

PACTE — did the agent get there the right way?

Most evaluations ask whether an agent produced the right result. PACTE asks how it got there: what it was trying to do, what it was allowed to do, what it could do, how it went about it, and what evidence supports the account.

A correct result does not establish compliant operation.

The five questions

Purpose · Authority · Capability · Tactics · Evidence

P
Purpose

What objective was the agent assigned or rewarded for?

Stated vs inferred objective

A
Authority

What was it permitted to access or change?

Declared vs effective authority

C
Capability

What could it actually do?

Claimed vs demonstrated capability

T
Tactics

How did it pursue the objective?

Expected vs observed tactics

E
Evidence

What supports the account of what happened?

Evidence quality and uncertainty

Output

An assessment report, not a badge

A PACTE assessment records findings, gaps, uncertainty and residual risk. It does not produce a score, a pass or a certificate, and it should never be described as an approval. The template below is a draft for review.

FieldWhat the assessor records
System & deploymentName, operator, model, harness, and AIH-12 horizon for this deployment.
PurposeObjective as assigned. Objective as rewarded or measured. Any divergence.
AuthorityDeclared permissions vs the permissions its credentials actually granted. Gaps.
CapabilityCapabilities claimed vs those demonstrated in this assessment.
TacticsActions expected vs actions observed. Shortcuts, workarounds or concealment.
EvidenceSources relied on and their evidence class. What could not be checked.
UncertaintyWhat the assessor does not know, and how it affects the findings.
Residual riskWhat remains after controls, and who carries it.
Findings & conditionsPlain-language findings and any conditions attached. No pass mark.
Defined in AIE-12 v0.1

Grading the evidence

PACTE asks assessors to label each source by how far it has been tested. The six classes are defined in AIE-12 Autonomy & Impact Evidence, a v0.1 draft for public comment.

  1. 1Assertion
  2. 2Demonstration
  3. 3Documented case
  4. 4Reproducible test
  5. 5Controlled pilot
  6. 6Independent review

Weakest to strongest. An assertion is a claim; an independent review is a check by someone with no stake in the result.

Relationship to the other instruments
AIH-12 tells you how much scrutiny a deployment warrants. F-ACT checks whether the GUARD controls exist before the agent acts. PACTE examines what actually happened when it did. They are designed to be used together, but each can be read and applied on its own.