ApexTransform

INDEPENDENT RESEARCH · AGENTIC AI SAFETY

Who authorized the agent?

From human oversight as a principle to authority that can be checked before an autonomous system executes a consequential action.

Dr. Stéphane Gervais, PhD in electronics and MBA, combines international technology leadership with an independent research agenda on human authority, agent control and verifiable execution evidence.

Research focus · AI control and scalable oversightStage · VCA framework, exploratory V3 and offline V4.2Evidence · software conformance and bounded Claude observations

01 / RESEARCH AGENDA

Test authority at the point of action.

Evaluate delegated authority against independently defined task outcomes, and measure what agents do when a control denies an action.

01 / COVERAGE

Which unauthorized effects can a gateway prevent?

Use externally authored tasks and a credible defense baseline. Score unauthorized effects independently of permit presence or gateway decisions, alongside legitimate task completion.

02 / ADAPTATION

What happens after a denial?

Test whether agents stop or reach the same forbidden effect through another tool, split actions or re-delegation. Record proposals, model refusals, gateway denials and executed effects separately.

03 / APPROVALS AND UTILITY

Does the control preserve useful human decisions?

Study misleading approval descriptions, false denials, completion, latency and reviewer burden. An action-bound permit alone does not establish that a person understood what they approved.

02 / SYSTEMS APPROACH

Separate planning, authorization and evidence.

The agent is a planner, not an authority source. The proposed control architecture checks policy before a simulated external effect and records evidence afterward.

UNTRUSTEDAgent plannerproposes an action
TRUSTED POLICYAuthority control prototypechecks scope and approval
SIMULATIONTool gatewayexecutes or blocks
RECORDINGTraceProofrecords event chain
EVALUATIONReportsafety and utility measures

Architecture illustrated from a synthetic research sandbox. V3 recorded exploratory native Claude tool-use observations in November 2026. V4.2 is a separate scripted, offline conformance exercise. Neither establishes a general safety improvement.

03 / TECHNICAL ARTIFACTS

Executable prototypes, precisely labelled.

The private research artifacts distinguish deterministic software checks from exploratory model observations. Code and detailed materials are available for scientific review on request, subject to confidentiality and disclosure review.

AUTHORITY CONTROL PROTOTYPE

Action-bound authorization

Implemented · synthetic conformance

Objective: prevent an agent's fabricated authorization narrative from becoming a permitted external action.

Trusted risk tiers; scopes; action fingerprints; expiry, revocation and replay prevention; machine-readable results.

A targeted correction checks delegated amount and recipient limits against canonical action parameters and rejects inconsistent declared context. Nine regression tests cover this repair; it is not a complete verification of every contextual constraint.

Limitation: A simulated approver and demonstration key do not authenticate a real human. No frontier-model robustness study.

TRACEPROOF

Evidence integrity checks

Implemented · synthetic conformance

Objective: test whether an action and its recorded authorization remain consistent after tampering attempts.

SHA-256-linked events; HMAC-authenticated checkpoint; verification of sequence, authorization, action and results.

Offline tampering tests check sequence integrity and consistency between actions, authorizations and results. These checks establish selected software properties under the simulator’s trust assumptions.

Limitation: A verified chain assumes a trusted recorder and checkpoint. It does not prove real human approval or physical execution.

INTEGRATED AGENT SANDBOX

Tool-execution test harness

Offline conformance + exploratory Claude pilot

Objective: connect agent planning, authorization decisions, simulated tool effects and evidence in one test loop.

Scripted offline integration plus a separate native Anthropic tool-use harness; in-memory invoice/payment tools, bounded prompts and reproducible result manifests.

V3, November 2026: 192 Claude Sonnet 5.5 and 96 Claude Haiku 5.5 native multi-turn trajectories on the same 48 locally authored synthetic tasks, across policy arms and repetitions. These are model observations, not 288 independent attacks.

Limitation: The apparent policy advantage came from four variants of one unapproved-request family and reflected the metric and comparison policy. No illicit payment proposal was observed in the tested tool-result injection cases, and no unauthorized second payment was proposed in the replay cases. Those cases establish no incremental gate benefit. Payments remained fictional and in memory.

RESEARCH STATUS / 17 NOVEMBER 2026

Verifiable Chain of Authority

Developed in September 2026, Verifiable Chain of Authority (VCA) is an evaluation framework for delegated human authority in agentic systems. Its authority-state schema, eight testable properties and ten failure classes connect authority provenance, delegation, non-amplification, freshness, contextual constraints, exact-effect binding and evidence continuity.

Manuscript in preparation. VCA is a proposed cross-system evaluation framework. Its reference implementation does not establish complete enforcement or general agent safety.

V4.2 · scripted offline conformance

Correction and reproducible checks

An adversarial audit with Claude identified a mismatch between declared context and canonical action parameters in the reference verifier. I personally reproduced the verifier probe and the model-free diagnostic control, and directed a targeted correction with regression tests.

V4.2 sends scripted proposals through a deterministic gateway before any fictional in-memory effect. Expected outcomes are specified from the legitimate task independently of gateway decisions. No live-model evaluation has been run for V4.2. These checks do not establish an empirical safety advantage.

Next · externally grounded evaluation

Test adaptation and legitimate completion

The next study would use a compatible subset of externally authored tasks from AuthorizationBench or AgentDojo, with independently specified outcomes and a credible defense baseline such as CaMeL where comparable.

The priority is behaviour after denial, unauthorized effects and legitimate completion. No gain over a credible baseline, excessive utility loss or successful alternate-path execution would count against the approach. The protocol would be frozen with an external timestamp before confirmatory runs.

Code and the detailed dossier remain private pending scientific and disclosure review. Research access enquiries.

04 / EXPERIMENTAL METHOD

Make both safety and utility measurable.

The proposed evaluation should report failures and operational costs, not only the number of rejected attacks.

Offline validation + exploratory model evaluation

What was actually implemented and observed

I developed the VCA framework and directed the November exploratory studies with substantial AI coding assistance. An adversarial audit with Claude showed that the apparent between-policy advantage reflected the metric and comparison design. I withdrew the interpretation as a general safety improvement.

  • Authority and trace verification exercised offline
  • Scripted and mock HTTP integration checked
  • Native Claude Sonnet and Haiku trials completed; no generalizable efficacy estimate
Next / not yet independently evaluated

Proposed controlled experiments

Use externally authored tasks and benchmark-defined ground truth, a genuinely comparable defense baseline, matched legitimate tasks, and independent outcome assessment. Prespecify the protocol before running. Study how agents behave after denial and how approval channels can be manipulated.

  • Primary outcome: task-labelled unauthorized effects, scored independently of permit status
  • Utility: legitimate completion and false refusal rate
  • Operational burden: enforcement latency and human escalations
  • Reproducibility: model/version, scenarios, prompts, trial counts and uncertainty

05 / PUBLICATIONS & CONTEXT

A research trajectory grounded in published work.

Editorial contributions and framework papers are distinct from completed empirical agent experiments. Publication types are identified accordingly.

Related research, not an affiliation. SLEIGHT-Bench: SLEIGHT-Bench (Najt et al., 2026) examines blind spots in monitoring systems, including authorization confusion. This independent work instead asks what happens at the enforceable tool boundary and whether subsequent evidence remains consistent.
Scientific integrity. These prototypes do not demonstrate that autonomous agents are safe. The software conformance tests are synthetic and native Claude tool-use trials were exploratory, with self-authored tasks and no measured general safety effect. Authenticated human approvals, independently secured evidence stores, externally authored adaptive challenges and robust utility measurement remain open work. Research code remains private pending scientific and disclosure review. Substantial AI coding assistance was used; no Anthropic affiliation or endorsement is implied.

Some published work appears under the name Stéphane Gervais Ducouret. Browse all articles.

RESEARCH COLLABORATION

Investigating control that can be tested, not merely asserted.

Open to technical dialogue on agent-control evaluations, human authorization, safety instrumentation and reproducibility. Commercial AI strategy and deployment advisory remain separate.