Skip to main content
Aurélien Vannieuwenhuyze

RESEARCH PROJECT · 001

TENSYM.

Agentic AI & Temporal Neuro-Symbolic Reasoning for Clinical Trials

What if an AI agent’s reliability also depended on what it knew — and when it knew it?

TENSYM investigates how errors spread across AI agents and whether independent temporal rules can verify their decisions in a fully simulated clinical trial.

Active research · Started October 2026
Applied research · Agentic AI · Neuro-symbolic reasoning

01 / THE RESEARCH QUESTION

Can a decision be verified if its temporal context cannot?

Information can describe an event that has already occurred yet still be unavailable when an agent makes a decision. Across multiple agents, documents and corrections, this becomes a question of causality, provenance and auditability.

What did the agent know? When? Which source and version did it use? Which rule applied?

02 / THE EXPERIMENTAL SETTING

A deliberately small scope, to isolate the mechanisms.

An initial cohort of ten entirely fictional participants and forty visits has been generated. The agent roles and scenarios below describe the intended experimental design; there are no multi-agent results at this stage.

10synthetic participants
4visits per participant
3planned agent roles
5planned scenarios S0–S4

03 / THE EXPERIMENT

Observe the decisions. Then verify the evidence.

The proposed experimental setup will have AI agents work with synthetic source documents and eCRF data. Independent temporal verification remains a hypothesis to test, not an observed result.

INPUT / 00

Synthetic events and source data

The cohort and synthetic source documents have been generated; their temporal availability will be examined in future experiments.

AGENT / 01

Research nurse AI

Will prepare and share source notes using only authorized information.

AGENT / 02

Clinical data coordinator AI

Will transcribe information into the eCRF and address queries.

AGENT / 03

Clinical research associate AI

Will review accessible data and may raise clarification queries.

INDEPENDENT CHECK / RULES

Independent temporal rule engine (planned)

Proposed component for independently checking rules, evidence availability, access permissions and provenance, without relying solely on model-generated explanations.

OUTPUT / EVIDENCE

Traces, decisions and replay

The design calls for logging agent actions and their supporting evidence for independent evaluation.

In the planned experimental design, the three roles are consolidated in single-agent conditions. Rules will be active only in configurations that enable them.

04 / WHAT THE PROJECT WILL TEST

Five scenarios. Four configurations. No predetermined outcome.

Each proposed scenario will compare observed decisions against synthetic reference facts hidden from the agents.

S0Baseline

Consistent data

Measure false alerts and unnecessary blocking.

S1Date discrepancy

Visit on Day 17, eCRF says Day 15

Test whether incorrect transcription hides a visit outside the Day 12–16 window.

S2Transcription error

154 vs 124 mmHg

Observe whether a wrongly transcribed measurement spreads or is detected.

S3Missing information

Hospitalization omitted

Check whether a clarification and simulated human review are triggered.

S4Temporal leakage

Available Day 15, decision Day 14

Test denial of access to information that is not yet available.

COMPARATIVE DESIGN

The same experiment, with or without active checks.

One independent scorer must assess all four conditions against the same criteria. No experimental results have been published yet.

#Agent organizationVerification during simulation
A1 agent · 3 combined rolesNo active rule engine
B1 agent · 3 combined rolesRule engine enabled
C3 specialized agentsNo active rule engine
D3 specialized agentsRule engine enabled

05 / VISUAL OVERVIEW

An illustration of the broader concept.

The diagram below is an early conceptual illustration. Some modules and visit windows exceed the agreed V0.1 scope; the experimental protocol described on this page is authoritative.

Conceptual TENSYM illustration featuring AI agents, a synthetic eCRF and a temporal verification engine.
Open full-size illustration

06 / RESEARCH IN PROGRESS

Current research status

The initial SYN-HTA-001 cohort is available: 10 fictional participants, 40 visits and 120 sets of systolic/diastolic blood pressure and pulse measurements. The generator underwent technical coherence and reproducibility testing, including 10,000 synthetic cohorts. This does not constitute clinical validation.

07 / RESEARCH JOURNAL

Documenting research, one milestone at a time.

Each note reports a tangible research milestone, the decisions behind it, the checks performed and the limitations. Unfinished experiments are never presented as results.

NOTE 001 · OCTOBER 29, 2026 · MILESTONE 02
Explore the research journal

SCOPE & TRANSPARENCY

TENSYM is a research simulation built exclusively on synthetic data: no real patients, no real clinical trial, and no claim of eCRF certification or regulatory compliance. It does not assess treatment efficacy or make autonomous medical decisions.

ABOUT THE RESEARCHER

Research is also an invitation to work together.

I’m Aurélien Vannieuwenhuyze, an AI Research Engineer working on temporal medical data, neuro-symbolic reasoning and agentic AI systems. I’m open to new opportunities in R&D and Applied AI, especially in healthcare and pharma — through permanent roles, consulting assignments or research collaborations.

View my CV ↗