Permission changes
Compare granted, prohibited, withdrawn and unresolved authority under agreed rules.
Special engagement / Independent Agent-Control Pilot
Put human control
to the test.
Examine a defined model-and-agent configuration. Find actionable control failures, understand useful performance and check whether a correction worked.
Why now
Turn commitments into observable behavior.
As AI agents take on more consequential tasks, teams need inspectable evidence of how permissions, instructions and operating conditions shape their behavior. This pilot is designed to answer a defined engineering question before the next deployment decision.
Dated methods context, reviewed 15 September 2026: Google DeepMind’s external-evaluation pilot (27 August) and Microsoft AI’s six-week consultation opened 14 September. Editorial review due 13 October 2026. These references do not imply affiliation, endorsement or implementation by ATBAE of those organizations’ methods.
Untrusted instructions
Examine whether third-party content improperly changes the task or its permitted scope.
Interruption and recovery
Observe the agreed stop mechanism and recovery behavior in the selected environment.
Work within limits
Report completion of authorized work separately. Blanket refusal does not demonstrate useful performance.
What the engagement entails
Declare. Examine. Explain. Retest.
- Declare the scope: identify the system, intended use, permissions, control families and decision the evidence must support.
- Confirm readiness: verify the agreed access, configuration and test apparatus before collecting subject results.
- Run the bounded examination: collect the agreed cases and variants, retaining failed collections and incomplete observations.
- Deliver and discuss: provide the report, methods annex, evidence package and one engineering findings meeting.
- Check one correction: retest one revised configuration against the agreed findings and budget, preserving original observations.
A standalone special engagement
US$15,000 pre-tax. No automatic enterprise multiplier. The scope excludes additional models, extra families, unrestricted penetration testing, ongoing monitoring and a certification opinion. Broader needs are commissioned through the Catalog. Any commercial adjustment belongs in the SoW.
Included / Methods and Evidence
A report you can interrogate.
How the result was obtained
Engagement-specific protocol, tested conditions, configuration identities, sample units, judging criteria, exclusions, handling of missing observations and limits on inference.
What supports each finding
Relevant response and action records, selected traces, finding-to-artifact references, evidence inventory, byte hashes, collection status and agreed reproduction instructions.
Methods and Evidence are included deliverables. They do not imply a transfer of proprietary engine source, unrestricted access to the sealed corpus, or public release of answer keys. Restricted material is handled through the agreed review and access arrangements.
Client findings are confidential. The examples distinguish a redacted retained-results extract from a purpose-built fictional report.
The output report / eight main sections
From the decision to the supporting record.
| Report section | What the reader receives |
|---|---|
| 01 / Executive decision brief | The business question, principal observations, unresolved findings and recommended next action. |
| 02 / Scope and system identity | The examined model/version, tool environment, permission rules, selected families, coverage and exclusions. |
| 03 / Methods and measurement | Case design, counterfactuals, sampling unit, repetition, judging rules, missing-data handling and limitations. |
| 04 / Results by control family | Separate findings on authority, delegation, interruption and authorized work; explicit denominators and missing observations. |
| 05 / Findings and evidence | Reproduction steps, relevant transcript extracts, attempted actions versus effects, evidence IDs and impact context. |
| 06 / Correction and retest | Proposed corrections, agreed retest criteria and a separately identified retest result that preserves the baseline. |
| 07 / Evidence inventory and provenance | Artifact identities, hashes, configuration records, collection logs, access restrictions and reproduction dependencies. |
| 08 / Limitations and next examination | What the evidence supports, what remains untested, decision prerequisites and separately scoped follow-on work. |
The fee buys an examination and its evidence package, not an eight-page document. The client receives a scope-complete executive brief, technical report, methods annex, finding register, machine-readable results, evidence inventory and one bounded correction-retest report. No eight-page cap applies. The eight-section illustration is deliberately abbreviated; client report length follows the agreed scope and findings. Two distinct examples are available: a redacted extract of retained live results and a fictional report illustration. Neither is a certificate. Read the real-results extract.
Begin with the question
Bring one system.
Define the decision.
Discuss the target configuration, intended use, permitted actions and evidence your team needs. Delivery is subject to the agreed scope and readiness checks; a favorable outcome is never guaranteed.
Mention “US$15,000 Agent-Control Pilot” in your inquiry. This scoped service does not establish universal safety, regulatory approval or full production-assurance certification.