Author: Tomasz Jan Gomoła · ORCID 0009-0001-5222-6154 Affiliation: DeepSensi Medical OS — Cognitive Infrastructure Division, DeepSensi PBC, Dover, DE, USA Version: 3.0 — July 2026 Status: Prepared for submission — medRxiv preprint; NEJM AI / npj Digital Medicine Recommended citation: Gomoła, T.J. (2026). "Autonomous Bayesian Hypothesis Generation and Adversarial Validation for Medical Research: The AutoResearcher-HYPO Architecture." Technical Whitepaper WP-005. DeepSensi Medical OS.
Background. Medical research operates under a fundamental bottleneck: the time from initial observation to clinical application averages 10–17 years. Current AI systems in medicine are predominantly reactive — they analyze data presented to them — and do not autonomously generate, validate, and refine novel research hypotheses. No existing system combines autonomous hypothesis generation with formal Bayesian convergence guarantees and adversarial multi-agent validation in a continuous loop.
Methods. We describe the AutoResearcher-HYPO architecture: an autonomous, iterative hypothesis generation and validation engine for medical research. The system generates novel hypotheses via a set of structured creativity protocols, validates them via an adversarial multi-agent swarm operating against real-time biomedical evidence, and mathematically updates beliefs using Bayesian tracking over competing research vectors. Breakthrough findings are published to a Hypothesis Marketplace (HYPO), where they undergo federated cross-institutional peer review and human Sentinel (physician) review within a Hyper Consilium (WP-003). Before assimilation into the system's self-learning knowledge base, hypotheses undergo counterfactual stress-testing to automatically generate clinical exclusion criteria — closing the loop from raw observation to verified, safely bounded clinical knowledge.
Deployment status. The engine is being exercised in ongoing, open research missions aligned with the public-benefit charter of DeepSensi PBC — currently including autism spectrum disorder and pediatric epilepsy. These are active research programs, not completed studies: no validated results, convergence signals, or causal claims are reported in this paper. The contribution here is the architecture and its safety properties, not any specific biomedical finding.
Conclusions. AutoResearcher-HYPO is, to our knowledge, the first autonomous medical hypothesis engine with formal Bayesian convergence guarantees and adversarial validation. It operates within the DeepSensi Standard (DSS-001) verification framework, ensuring that every generated hypothesis is subjected to the same multi-layer safety checks applied to clinical diagnostic output.
Keywords: autonomous research; hypothesis generation; Bayesian inference; adversarial validation; medical AI; drug repurposing; evidence synthesis; DeepSensi Standard
The gap between discovery and clinical application spans 10–17 years, with attrition exceeding 90% at each major stage. The bottleneck is not primarily execution capacity but hypothesis quality: most research programs fail because the initial hypothesis was insufficiently informed by the full breadth of available evidence. The problem is compounded by scale — over 3 million peer-reviewed articles are published annually across the biomedical sciences, and cross-domain insights that might accelerate one field often lie undiscovered in the literature of another.
Current applications are predominantly reactive and narrow: drug-discovery systems evaluate candidates against known targets but do not generate mechanistic hypotheses; matching systems identify patients for existing protocols but propose no new directions; review tools summarize but do not synthesize cross-domain connections into testable predictions. Large language models articulate hypotheses well but their application to autonomous research has been limited by three constraints: (1) absence of formal convergence guarantees — without Bayesian tracking, iterative exploration produces insights that cannot be mathematically evaluated for convergence; (2) absence of adversarial validation — hypotheses from a single model inherit that model's systematic biases; and (3) absence of evidence-quality assessment — current systems treat all retrieved evidence as equally trustworthy.
This paper introduces an autonomous, infinite-loop hypothesis generation and validation engine that addresses these three constraints through Bayesian belief tracking with formal convergence guarantees, an adversarial validation swarm that subjects every hypothesis to structured attack before acceptance, and a tiered evidence-quality framework — all operating within the DSS verification framework, so that autonomous research output meets the same safety criteria as clinical diagnostic AI.
AI-assisted discovery has produced notable successes in constrained domains (protein structure prediction; antibiotic screening), but these optimize within a defined search space rather than generating open-ended mechanistic hypotheses de novo. Recent autonomous research agents have demonstrated iterative AI-driven loops in computer science and mathematics, but without formal convergence guarantees, adversarial validation, or the evidence-quality assessment essential where false hypotheses risk misdirected trials and patient harm. Bayesian inference is well established in trial design and diagnostic reasoning; its application to research-direction selection — treating competing hypotheses as a probability distribution updated on accumulated evidence — has not, to our knowledge, been previously described in the autonomous-research context. The adversarial multi-agent architecture here (dedicated agents systematically attacking each other's conclusions) differs fundamentally from ensemble voting.
Each iteration executes: (1) Strategist — a creative generation agent reads the current belief state, prior results, accumulated dead-ends, and breakthrough candidates to generate the next hypothesis, selecting from structured creativity protocols (§3.2); (2) Adversarial validation — the hypothesis is submitted to the validation swarm (§3.4) evaluating it against real-time evidence from 22+ biomedical databases; (3) Evaluation — a research evaluator measures knowledge delta and assigns a Frontier Confidence Score (FCS, 0–100); (4) Bayesian belief update — even discarded hypotheses update the distribution, because negative evidence is evidence; (5) Verdict and HYPO publication — each iteration receives KEEP, DISCARD, BREAKTHROUGH (high FCS with high-tier evidence), or CRASH; breakthroughs are automatically published to HYPO; (6) Federated and Sentinel peer review — on HYPO, a local adversarial red team and the longitudinal digital-twin verification engine challenge findings while a federated aggregator broadcasts the hypothesis to cross-institutional nodes for zero-PII validation against independent cohorts, and human Sentinels can be summoned within a Hyper Consilium for definitive clinical review; (7) Counterfactual stress-testing — before assimilation into the knowledge base, a counterfactual engine simulates clinical edge-case perturbations and automatically generates precise exclusion criteria, structurally preventing feedback-loop bias; (8) Loop — the system returns to step 1 with an updated belief state, running indefinitely until paused or a convergence threshold is reached.
To avoid the convergence trap (a system that only generates hypotheses similar to prior successes misses novel directions), the strategist employs structured creativity protocols forcing exploration of different regions of the hypothesis space: Inversion (what if a held assumption is wrong?), Cross-Domain Analogy (mechanisms from veterinary medicine, botany, or physics), Research Archaeology (studied 50+ years ago but abandoned for technological, not efficacy, reasons), Meta-Pattern Analysis (what do multiple dead-ends share?), Combination, Simplicity (Occam's razor), Outlier Analysis (unexpected recoverers), Absence Signal (conspicuously missing data), Mechanism-First, Serendipity, Patient Voice (mechanisms behind dismissed anecdotes), and Drug Repurposing. Protocol selection follows a deterministic schedule that forces meta-pattern analysis periodically and triggers serendipity after consecutive discards to escape local minima.
The module maintains a probability distribution over research vectors — competing directions that might explain the phenomenon. After each iteration k producing evidence Eₖ, the posterior of each vector Vᵢ updates by Bayes' rule:
P(Vᵢ | E₁…ₖ) ∝ P(Eₖ | Vᵢ) × P(Vᵢ | E₁…ₖ₋₁)
The likelihood function maps iteration outcomes to evidence strength (breakthroughs raise the likelihood ratio; discards apply a weak negative signal; crashes are neutral), with additional components for evidence-tier weighting and mechanism plausibility. The exact likelihood values are calibration parameters and are proprietary.
Anti-hallucination anchoring of the likelihood. A critical defense against a model inflating the likelihood ratio with fabricated reasoning is a deterministic mechanistic-bridge validation gate: before the likelihood ratio is permitted to rise above neutral, the proposed mechanistic pathway (e.g., A → C) is queried against the system's genomic/pathway analysis stack (curated biological network databases such as Reactome, KEGG, and protein-interaction graphs). If the mechanistic bridge cannot be structurally mapped to established biological networks, the evidence-integrity score is zeroed and the hypothesis is pushed to the Null Discovery Protocol. Bayesian convergence is thus anchored in deterministic molecular graphs, structurally preventing "Bayesian nonsense proving."
Convergence behavior (theoretical). The structure ensures that vectors consistently supported by evidence accumulate posterior probability given sufficient iterations and informative likelihood ratios — even weak but consistent signals compound. The table below illustrates theoretical posterior growth from a uniform prior of 10% across 10 vectors, under varying per-iteration signal strengths:
| Iterations | 1% signal/iter | 5% signal/iter | 10% signal/iter |
|---|---|---|---|
| 10 | 10.9% | 15.3% | 22.4% |
| 50 | 15.5% | 56.0% | 92.9% |
| 100 | 23.1% | 93.6% | 99.9% |
| 200 | 44.8% | ~100% | ~100% |
| 500 | 94.1% | ~100% | ~100% |
Theoretical only; assumes i.i.d. evidence signals and a uniform prior. When evidence is entirely uninformative (all likelihood ratios neutral), the system correctly maintains high entropy, signaling that available data are insufficient for convergence. System uncertainty is continuously measured via Shannon entropy H = −Σ P(Vᵢ) log₂ P(Vᵢ); the strategist uses entropy to balance exploration (high entropy) against exploitation (low entropy). A pathway-convergence graph additionally detects when multiple independent hypotheses implicate the same biological mechanism (for example, a shared signaling pathway or metabolic node) — treated as particularly strong evidence, since it is unlikely to arise from bias in a single research direction.
Every hypothesis is validated by an adversarial swarm operating within a protected innovation zone (evidence with honest tier labeling is permitted, but factual claims must survive attack): an Anomaly Hunter (cross-modal anomalies — values individually normal but jointly unusual, temporal mismatches, conspicuous absences); a Mechanism Linker (biologically plausible mechanisms grounded in real biochemistry, not narrative convenience); a Literature Archaeologist (searching 22+ real-time evidence databases, classifying every source by evidence tier); a Devil's Advocate (systematically attacking findings across attack vectors — anomaly validity, mechanism plausibility, evidence reality, simpler explanation, spurious correlation, confirmation bias, fabrication risk — with a survival verdict per finding); and a Synthesis Judge (final rulings, FCS scores, evidence passports, operating without access to raw patient data to prevent bias). A Null Discovery Protocol ensures the absence of findings is itself a valid, valuable output: a null finding means existing understanding is robust.
Evidence quality is classified into five tiers:
| Tier | Category | Weight |
|---|---|---|
| 1 | Human RCT / meta-analysis | 1.0 |
| 2 | Human observational | 0.8 |
| 3 | Case report / series | 0.5 |
| 4 | Animal / in-vitro | 0.3 |
| 5 | Theoretical / mechanistic | 0.15 |
The FCS integrates evidence tier, adversarial survival, mechanism plausibility, and replication potential into a composite 0–100 metric with defined thresholds (auto-discard, speculative, notable, significant, potential-breakthrough). Threshold values were empirically calibrated during pilot development with clinical domain experts to align with EBM confidence hierarchies, and are configurable for domain-specific deployments; they represent conservative defaults optimized to minimize false-positive breakthrough classification.
The architecture operates independently of any language-model provider: a universal inference gateway routes generation and validation through a configurable provider chain with automatic failover. This yields no single-vendor dependency (findings are not artifacts of one model's training distribution), cross-vendor verification (key hypotheses re-evaluated by a model from a different architectural lineage — agreement across lineages is stronger evidence than intra-family agreement), and regulatory resilience (missions survive provider changes without data loss).
Extended missions (hundreds to thousands of iterations) are made computationally sustainable by a prior-art injection mechanism: before each model call, a clinical-trajectory database is queried for semantically similar prior analyses, and when a match is found a compressed representation of the prior reasoning is injected so the model focuses on new deviations rather than re-deriving established conclusions. Optimization is disabled for high-stakes calls (final synthesis, adversarial challenge) where full context is always required. Resource consumption is metered per iteration, enabling budget-constrained missions that can pause and resume. For reproducibility, an offline replay mode reuses cached evidence responses, enabling deterministic re-execution by independent reviewers to verify convergence behavior. Specific similarity thresholds and injection policies are proprietary.
The architecture is being exercised in ongoing, open research missions aligned with the public-benefit charter of DeepSensi PBC, currently including autism spectrum disorder and pediatric epilepsy. These are stated here only to describe where the engine is running; they are active research programs, not completed studies.
We deliberately report no preliminary results, no convergence signals, and no single-case observations from these missions. This restraint is a design choice, not an omission: system-generated hypotheses are, by construction, candidate research directions that require independent clinical validation, formal IRB review, and prospective study design before any clinical interpretation — and, given the sensitive history of speculative hypotheses in some of these conditions, presenting interim signals as findings would be exactly the failure mode this architecture exists to prevent. Each mission runs under the same DSS verification framework (§5) and immutable audit trail as clinical diagnostic output, so that every hypothesis generated is traceable and every discarded direction is permanently recorded. The claim of this paper is therefore about the engine — its operational behavior, convergence mathematics, and safety architecture — and not about any specific biomedical conclusion in autism, epilepsy, or any other condition. Results from these missions, if and when they reach the evidentiary bar, will be reported in dedicated, independently reviewed publications.
AutoResearcher-HYPO operates within the DSS verification framework (DSS-001): every hypothesis passes the same deterministic safety barriers applied to clinical diagnostic output (input-coherence validation, cognitive-bias detection); cited evidence is subjected to Evidence Integrity Score assessment (provenance, funding-bias detection, population matching, norm-genealogy audit); all iteration results, including the complete Bayesian belief state, evidence chains, and adversarial-attack outcomes, are recorded in an immutable audit trail with cryptographic hash-chain integrity; and the formal reliability analysis (WP-001) provides theoretical error bounds for the adversarial-validation component. Records are Merkle-batched and anchored to an external ledger, so the provenance of every hypothesis — including discarded ones — is permanently and independently verifiable. Breakthrough hypotheses can be forwarded to a population-simulation engine that generates synthetic digital-twin cohorts (N = 10–1,000) and evaluates therapeutic hypotheses via Monte Carlo simulation (estimated response rates, numbers needed to treat, predicted adverse-event profiles), carrying a mandatory "simulation only" disclaimer and recorded in the audit trail. This means autonomous research output carries the same verifiable safety guarantees as DSS-certified clinical output — a property not shared by any other autonomous research system known to the author.
The architecture has potential implications for drug repurposing (cross-domain creativity plus evidence archaeology suited to identifying approved drugs with off-target effects, at dramatically reduced regulatory timelines), rare-disease research (autonomous generation and evaluation of hundreds of hypotheses where manual review is prohibitive), population health (hypotheses that account for pharmacogenomic variation and population-mismatched evidence), and pharmacovigilance (real-time access to adverse-event reporting, with every safety signal subjected to the same anti-hallucination and norm-genealogy verification, enabling detection of signals not yet in published literature).
Comparison. AutoResearcher-HYPO occupies a distinct category: it is the only system that generates hypotheses autonomously, validates them adversarially, tracks convergence mathematically with a formal guarantee, and reproduces its reasoning deterministically. Analytical tools (literature summarizers, citation-context systems, consensus meters) are valuable aids but are not autonomous research engines: they do not generate hypotheses, provide no convergence guarantee, and offer no adversarial validation or population simulation.
This paper reports no biomedical findings, by design: the research missions in which the engine currently runs are active and unfinished, and no clinical decisions should be based on system output without independent verification. Prospective evaluation of the architecture itself is underway, with EHR integration via FHIR R4 and HL7 v2; institutions interested in the evaluation program are invited to contact the corresponding author. Hypothesis quality depends on the underlying models; while adversarial validation and cross-vendor verification mitigate individual biases, systematic biases shared across all model families cannot be detected by intra-system validation alone. Evidence retrieval is limited to programmatically accessible databases (non-English research, paywalled studies, and unpublished "file-drawer" data may be missed). Extended missions incur significant computational cost, mitigated but not eliminated by the sustainability mechanism. The Bayesian convergence guarantee is mathematical, not empirical: it is conditional on evidence signals being at least weakly informative, and for entirely uninformative questions the system correctly reports high residual entropy rather than converging.
AutoResearcher-HYPO is, to our knowledge, the first autonomous medical hypothesis engine combining generative creativity protocols, formal Bayesian convergence guarantees, and adversarial multi-agent validation within a continuous research loop. Its mathematical foundation ensures that directions consistently supported by evidence accumulate posterior probability given sufficient iterations and informative signals, while the adversarial architecture ensures hypotheses survive structured attack before acceptance. Integration with the DSS verification framework establishes a new paradigm in which research AI is held to the same safety standards as clinical AI. The engine is running in ongoing, open research missions (currently including autism spectrum disorder and pediatric epilepsy) under the DeepSensi PBC public-benefit charter; consistent with the safety-first posture of this work, no biomedical results from those missions are claimed here, and any future findings will be reported only after independent clinical validation. The architecture, creativity protocols, Bayesian tracking methodology, and adversarial swarm are the subject of patent applications.
Tomasz Jan Gomoła is the founder and System Architect of DeepSensi PBC, which developed the AutoResearcher-HYPO architecture. Certain components are the subject of patent applications. No external funding was received.
© 2026 Tomasz Jan Gomoła / DeepSensi PBC. The Gomola Framework and DeepSensi Standard are open, royalty-free (attribution required). The reference implementation (DeepSensi Medical OS) is proprietary. Patent applications pending.