Every drug reimbursement submission ends on a committee's table. A panel of experts weighs the clinical narrative, the comparators, the economics, and the precedents — then issues a recommendation that determines whether patients get access, and under what conditions. For sponsors, that deliberation has always been the single most consequential step in market access and commercialization, and the least predictable.
Our validation study, posted to medRxiv in March 2026, asks whether the deliberation itself can be simulated — and tests the answer on decisions the models could not possibly have seen. This is the study behind Loon Waters™.
medRxiv · Pre-print · Posted March 3, 2026 Monte Carlo Committee Simulation with Large Language Models for Predicting Drug Reimbursement Recommendations and Conditions: A Novel Neurosymbolic AI Approach Ghayath Janoudi, Mara Rada, Eugene Yasinov, Trevor Richter — Loon Inc., Ottawa, Ontario, Canada doi.org/10.64898/2026.03.02.26347434 · Download the PDF
Key results: 93.2% recommendation accuracy · 96.8% high-mandate accuracy · 86.3% condition accuracy · 67 temporally separated decisions
Why predicting the recommendation is not enough
Health technology assessment agencies issue reimbursement recommendations that determine patient access to new therapies. Predicting those outcomes would let sponsors optimise market access strategy before filing, and let health systems anticipate budget impacts. The obstacle is that traditional machine learning approaches require extensive manual feature extraction — and predict only categorical outcomes.
In practice, the category is the least informative part of the decision. The overwhelming majority of positive HTA recommendations arrive as "Reimburse with conditions" — and it is the conditions that shape real-world access: who may be treated, who may prescribe, when treatment must stop, what price reduction is expected, and what further evidence is required. A prediction that stops at the recommended category tells an access team almost nothing about what to prepare for.
A committee in silico
Monte Carlo Committee Simulation is a neurosymbolic multi AI system that simulates multi-panelist deliberation rather than asking a single model for a verdict. Persona-conditioned large language model panelists deliberate a submission independently; their votes are aggregated through weighted voting, with uncertainty quantification built into the process.
The system predicts two things at once:
- the recommendation category — Reimburse with Conditions, or Do Not Reimburse; and
- the five condition categories attached to it — Population Restrictions, Prescriber/Setting Requirements, Continuation Conditions, Economic Conditions, and Evidence Conditions.
A validation the models could not memorise
The standard objection to evaluating large language models on public decisions is data contamination: if the outcome was on the internet before the model's training cut-off, a correct "prediction" may be nothing more than a model's ability to memorize and recall that information.
The study was designed so that this objection cannot apply. We conducted a temporal external validation on CDA-AMC (Canada's Drug Agency) sponsor-submitted recommendations published between October 2024 and December 2025 — 67 decisions in total, all issued after the knowledge cut-off of the underlying models. Every correct prediction therefore reflects reasoning about the evidence, not memory of the answer.
What the numbers show
On submissions where the system expressed confidence (n=44), recommendation prediction achieved 93.2% accuracy (95% CI 84.1–100.0%), exceeding the 91.8% majority-class baseline. Discrimination was well above chance level (AUROC 0.817 versus 0.500), and confidence estimates were calibrated (ECE = 0.091) — the reported confidence tracks the actual probability of being right.
Condition prediction is the harder task, and the more valuable one. Getting all five condition categories right simultaneously means selecting the correct combination out of 32 possibilities — something random guessing would get right just 3.1% of the time; the system achieved 48.8% subset accuracy on that all-or-nothing standard, roughly sixteen times chance, and 86.3% Hamming accuracy across individual categories — against 25.8% for a no-conditions baseline. Per-category accuracy ranged from 68.3% (Continuation Conditions) to 97.6% (Economic Conditions), with Continuation Conditions showing the strongest discriminative ability (AUROC 0.896).
An AI that knows when it is unsure
A forecast is only usable if you know when to trust it. The study pre-specified a Strength of Mandate measure that stratifies predictions by the decisiveness of the simulated committee's vote — and accuracy tracked it faithfully, from 96.8% on high-mandate predictions down to 40.0% on weak ones.
The pattern held for the errors themselves: 83.3% of all mistakes occurred in cases the system had itself flagged as uncertain (p=0.0025). And in the five cases where the system abstained entirely, accuracy would have been just 40.0% — the abstentions were warranted. On cases where the system expressed high confidence, teams can act; where it signals uncertainty, expert judgement knows exactly where to focus.
What this means for market access teams
The Loon Waters™ Monte Carlo Committee Simulation enables a shift from reactive to proactive market access: anticipating the specific reimbursement conditions a committee is likely to attach — before the committee convenes — with calibrated confidence that identifies which predictions to trust. Submission strategy stops being a single, irreversible bet and becomes something a team can pressure-test and iterate while there is still time to change it.
The positioning matters as much as the performance: the system is a forecasting aid that complements, rather than replaces, human deliberation. In Loon Waters™, the simulation backend is paired with ex-HTA insider interpretation — structured model output read by people who have sat on the other side of the table.