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). "Golden Horizon: AI-Mediated Compassionate Access Through an Autonomous Agent Architecture." Technical Whitepaper WP-006. DeepSensi Medical OS.
Background. Patients with terminal or treatment-refractory conditions face a cruel information asymmetry: experimental therapies that could extend or save their lives may exist in trial registries across the globe, but no systematic mechanism connects these patients to those trials in real time. Existing trial matching relies on physician awareness, patient self-search, or commercial recruitment services — all geographically limited, financially motivated, and structurally biased toward commercially sponsored studies. Compassionate-use and expanded-access programs, the last therapeutic option for many patients, remain underutilized (fewer than ~1,500 US applications per year) due to regulatory complexity and the absence of proactive advocacy.
Methods. We describe Golden Horizon: an AI-mediated compassionate-access architecture that autonomously scans global clinical-trial registries, matches anonymized patient profiles to experimental therapies using anti-hallucination-verified "therapeutic urgency vector" extraction, and deploys an autonomous advocacy agent that manages the entire patient-to-provider lifecycle through a formally defined seven-state finite-state machine. The system is free for patients by architectural mandate — a zero-cost constraint enforced at the code level, not as a policy promise. Privacy is preserved through automated PII anonymization across 15+ pattern categories with jurisdiction-specific validation, k-anonymity enforcement (k ≥ 5) for cohort queries, and a cryptographic consent escrow that prevents de-anonymization without explicit patient and physician authorization. The architecture integrates bidirectionally with the Hyper Consilium diagnostic pipeline (WP-003) and, on a structurally independent track, with the PharmaLink commercial trial-matching infrastructure.
Conclusions. Golden Horizon is, to our knowledge, the first clinical AI system to combine borderless registry scanning, autonomous patient advocacy through a formal state-machine agent, anti-hallucination-verified matching, and an architecturally enforced zero-cost patient model within a privacy-preserving, consent-gated framework. It transforms compassionate access from a passive, physician-initiated process into an active, AI-mediated patient right. Prospective evaluation is underway.
Keywords: compassionate use; expanded access; clinical-trial matching; autonomous agents; patient advocacy; privacy-preserving AI; state-machine architecture; DeepSensi Standard
For patients who have exhausted standard-of-care, Phase 1 trials, expanded-access programs, and compassionate-use protocols represent the boundary between exhausted options and potential survival. Global registries — ClinicalTrials.gov (450,000+ trials), the EU Clinical Trials Register (EudraCT), J-STAGE, KoreaMed, IndMED, and the WHO ICTRP — collectively catalog hundreds of thousands of active studies. Yet the mechanisms connecting patients to them are fragmented, geographically constrained, and structurally passive: they require physician awareness of specific trials, patient capacity for self-search, or engagement with commercial recruiters that prioritize sponsor-funded studies.
Existing platforms operate on a model in which sponsors pay for patient identification and enrollment. This creates three structural biases: (1) only commercially sponsored trials receive recruitment support — compassionate-use and investigator-initiated trials are systematically excluded; (2) recruitment effort is proportional to sponsor payment, not patient need; and (3) geographic coverage is limited to profitable markets. The result is a system in which the patients with the greatest therapeutic need are the least likely to be connected to appropriate trials, because their profiles often align with early-phase, low-enrollment studies that offer insufficient commercial incentive.
This paper describes the Golden Horizon architecture: the borderless registry-scanning engine, the autonomous advocacy agent that manages the patient–provider lifecycle, the privacy-preserving matching infrastructure, and the zero-cost patient model enforced at the architectural level rather than as a policy promise.
Automated trial matching has been explored through NLP of eligibility criteria and structured EHR matching, but typically within a single registry and focused on eligibility rather than end-to-end advocacy. Commercial platforms provide matching as a service to sponsors, serving sponsor interests and covering primarily commercially viable trials. Regulatory frameworks for compassionate use vary widely (FDA Expanded Access, 21 CFR Part 312 Subpart I; EMA compassionate use, Article 83 of Regulation (EC) 726/2004; national programs in Japan and elsewhere), yet utilization remains low due to physician unfamiliarity and process complexity. To our knowledge, the deployment of an autonomous agent with a formal state-machine architecture specifically for compassionate-access advocacy — including proactive provider outreach and structured follow-up — has not previously been described.
The engine queries global trial registries in parallel through dedicated connectors. Individual registry failures are logged but do not block the overall search (fault-tolerant parallel querying) — essential for a system serving patients with urgent needs.
| Registry | Coverage | Connector |
|---|---|---|
| ClinicalTrials.gov | 450,000+ US/international trials | Structured API |
| EU Clinical Trials Register (EudraCT) | European clinical trials | Web extraction |
| J-STAGE | Japanese novel treatments (oncology, gastroenterology) | Academic API |
| KoreaMed / IndMED | Korean and Indian population-specific trials | Regional APIs |
| Veterinary Translational | Animal-to-human translational research | Custom extraction |
WHO ICTRP is used for cross-index reconciliation. The inclusion of veterinary translational research is a deliberate architectural choice: treatments showing efficacy in veterinary oncology or rare-disease models may represent translational opportunities invisible to human-only searches. Veterinary results are flagged with a translational-relevance score and presented with explicit caveats.
The matching algorithm begins with "therapeutic urgency vector extraction" — an AI-assisted analysis of the patient's anonymized profile that identifies the most therapy-resistant, lethal, or rare components of their condition and generates targeted search queries beyond simple condition-name matching (specific molecular targets and mutations; treatment-resistance patterns; rare phenotypes; cross-specialty targets such as checkpoint inhibitors repurposed across domains). Critically, every extracted vector is verified through the anti-hallucination architecture (WP-001) before being used as a query, preventing searches for non-existent mutations or therapeutically implausible combinations. Vectors that fail verification are discarded with audit logging; if all vectors fail, the system falls back to the original set as a safety measure, ensuring no patient is denied access due to over-aggressive filtering.
Raw results undergo multi-criteria filtering: phase prioritization (Phase 1, expanded access, compassionate use), oncology and rare-disease prioritization, translational relevance, cross-registry deduplication, and volume limiting to prevent information overload. Every qualified trial is tagged with a hardcoded zero-payout flag, enforcing the zero-commercial-payout principle at the data level. Matched trials then undergo deep verification through a structured adversarial adjudication modeled on the Hyper Consilium's triadic architecture (WP-003): a Sceptic identifies reasons the patient might fail screening; an Advocate identifies why the patient is an optimal phenotype match; a Judge issues a final verdict with reasoning. This ensures matches are clinically defensible alignments, not keyword associations.
The Golden Horizon agent is a formally defined autonomous agent that manages the full patient-to-provider lifecycle through a seven-state finite-state machine. State transitions are validated against a formal transition table; invalid transitions are rejected, ensuring the agent cannot skip required processing steps or enter inconsistent states. Every transition is recorded in the immutable audit trail with timestamp, source and target state, and transition reason.
| Current state | Valid next states |
|---|---|
| INTAKE | SCANNING |
| SCANNING | MATCHING, RESOLVED |
| MATCHING | OUTREACH, RESOLVED |
| OUTREACH | CONNECTING, FOLLOW-UP, RESOLVED |
| CONNECTING | RESOLVED, FOLLOW-UP |
| FOLLOW-UP | SCANNING, OUTREACH, RESOLVED |
| RESOLVED | ∅ (terminal) |
Autonomous patient communication. At every transition the agent proactively communicates through the patient's preferred channel (Telegram, WhatsApp, SMS, or email) with automatic failover. Messages are empathetic and action-free: the patient is explicitly told they need do nothing and that the agent will handle all coordination. Patients in crisis should not be burdened with administrative work — the system acts as an advocate, not a portal.
Proactive provider outreach. When matches are identified, the agent autonomously contacts trial providers (CROs, principal investigators, sponsors) through multi-channel outreach. Registered providers are connected automatically; unregistered providers receive a structured invitation. Outreach urgency is dynamically calculated from the patient's clinical acuity (terminal status, palliative indicators, number of active conditions); high-urgency cases receive escalated outreach including automated voice calls.
Structured follow-up. Unresponsive providers receive follow-up on a defined schedule (24 h, 48 h, 7 days). If all attempts are exhausted, the agent automatically re-scans every registry for new trials, reflecting the reality that the trial landscape changes continuously. Even in the RESOLVED-with-no-match state, the agent schedules periodic re-scans. This persistence is the defining characteristic: unlike passive systems that present results once, the agent continues advocating until a connection is established or all options are exhausted.
Consent. A mandatory three-step consent protocol is enforced before any case processing: (1) data-sharing consent; (2) experimental-awareness consent; (3) contact-authorization consent. It is architecturally impossible to create a Golden Horizon case without all three completed.
Automated PII anonymization. All clinical text transmitted to AI models or external registries is processed through an anonymization engine detecting 15+ categories of personally identifiable information (email, international phone, national identifiers with checksum validation, IBAN, medical-record numbers, dates of birth, IP addresses, passport numbers, payment-card numbers, context-gated tax/business identifiers, and person names). National-ID detection validates the appropriate checksum rather than matching any numeric string of the right length, avoiding false positives on legitimate clinical values; tax/business identifiers are matched only when preceded by their keyword. A safe-list of 200+ medical eponyms (Parkinson, Alzheimer, Hashimoto, Cushing, …) prevents false-positive name redaction — a critical distinction for medical text where proper nouns are frequently clinical descriptors.
k-anonymity. For pharmaceutical cohort queries, the system enforces k-anonymity with k ≥ 5: if fewer than five patients share the same quasi-identifier combination (age band, sex, ICD-category diagnosis, geographic region), no results are returned, preventing re-identification through unique attribute combinations.
Cryptographic consent escrow. The transition from anonymized matching to de-anonymized patient contact is governed by a cryptographic consent escrow. De-anonymization requires explicit patient consent (architecturally enforced boolean gate), physician (Sentinel) authorization, valid escrow-token verification, and compliance-flag verification (HIPAA, GDPR/RODO). The event is recorded in the immutable audit trail with cryptographic anchoring — a permanent, tamper-evident record of who authorized identification, when, and under what compliance conditions. (Token construction and thresholds are proprietary.)
Hyper Consilium. Diagnostic sessions classified COMPLEX or RARE (WP-003) automatically trigger a Golden Horizon scan as part of extended output, and scan results are presented alongside diagnostic conclusions — enabling the physician to evaluate diagnosis and experimental options in a single clinical context.
PharmaLink dual-track. PharmaLink (commercial) and Golden Horizon (compassionate) operate as parallel but structurally independent tracks. The zero-payout constraint on Golden Horizon matches is not a configurable setting; it is a hardcoded architectural constant, ensuring Golden Horizon can never be repurposed as a commercial recruitment channel.
| Dimension | PharmaLink (commercial) | Golden Horizon (compassionate) |
|---|---|---|
| Revenue model | Commission per qualified lead | Zero revenue (patient-access mission) |
| Trial sources | Registered pharma-partner trials | Global registries (borderless) |
| Patient cost | Zero (pharma-sponsored) | Zero (architectural mandate) |
| Payout | Commission-based | Hardcoded zero |
| Matching depth | ICD + labs + demographics + adjudication | Urgency vectors + anti-hallucination + adversarial adjudication |
| Provider outreach | Webhook to registered CRM | Autonomous multi-channel |
| Follow-up | Manual / webhook | Autonomous 24 h / 48 h / 7 d |
| Consent | k-anonymity + consent-to-reveal | 3-step mandatory + cryptographic escrow |
| Audit | Immutable log + blockchain anchor | Immutable log + blockchain anchor |
Value for pharmaceutical partners (through PharmaLink, not Golden Horizon): accelerated recruitment via AI-mediated matching with multi-factor eligibility verification that reduces screen-failure rates; consent-gated real-world evidence for pre-trial feasibility and post-market surveillance; pharmacovigilance integration across FDA safety endpoints; and population-scale simulation of therapy-response distributions before committing to full trials.
Regulatory variation in compassionate-use pathways means a matched trial may require jurisdiction-specific navigation that the system does not automate; Golden Horizon identifies trials, but the regulatory application remains the treating physician's responsibility. Registry completeness varies; ClinicalTrials.gov offers the most comprehensive programmatic access. The veterinary translational pathway requires careful clinical interpretation — animal-model efficacy does not guarantee human relevance. Provider response rates to autonomous outreach are not yet characterized at scale. Anti-hallucination coverage for vector extraction depends on the verification knowledge base; novel mutations or newly published findings may not yet be represented, potentially flagging valid vectors as hallucinated. Prospective validation is actively underway; the architecture integrates with EHR systems via FHIR R4 and HL7 v2, and oncology centers and rare-disease networks interested in the evaluation program are invited to contact the corresponding author.
The zero-cost imperative. Enforcing zero patient cost at the architectural level — rather than as business policy — reflects a specific ethical position: compassionate access should not be contingent on ability to pay for matching services. By hardcoding zero payout, the principle survives changes in management, strategy, or market pressure.
Autonomy and informed consent. The three-step consent protocol ensures patients understand the experimental nature of matched therapies. The agent's communication design explicitly avoids false hope: patients are told matches are experimental, eligibility is not guaranteed, and the system's role is to identify and connect — not to recommend or endorse.
The veterinary translational question. Presenting animal-model results to human patients raises legitimate ethical questions. The system labels veterinary-sourced results clearly, requires the physician to evaluate translational relevance, and includes such results only above a defined human-translatability threshold. The justification: for a patient with no remaining standard-of-care options, awareness of promising translational research — clearly labeled — is preferable to ignorance.
Golden Horizon combines borderless registry scanning, autonomous patient advocacy through a formally defined state-machine agent, anti-hallucination-verified matching, and an architecturally enforced zero-cost patient model within a privacy-preserving, consent-gated framework. The autonomous agent transforms compassionate access from a passive, geographically limited, physician-initiated process into an active, borderless, persistent advocacy system in which no matched trial goes unpursued and no unresponsive provider goes unfollowed. Integrated with the DeepSensi diagnostic pipeline (WP-003), the PharmaLink infrastructure, and the anti-hallucination verification architecture (WP-001), Golden Horizon positions compassionate access as a component of a broader cognitive infrastructure for medicine — one in which the most vulnerable patients receive the most systematic advocacy. Certain components (the autonomous-agent lifecycle management, the therapeutic-urgency-vector extraction protocol, and the cryptographic consent escrow) are the subject of patent applications.
Tomasz Jan Gomoła is the founder and System Architect of DeepSensi PBC, which developed the Golden Horizon 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.