Phase II — Capability Building

Years 5–15. Where individual capabilities scale to clinical relevance and cross-bucket integration begins in earnest. The central scientific question of the entire study — whether causal disease modelling will yield — gets its first real test in this phase.

Years 5–15

Risk profile Moderate

Decision gate B3 viability

Strategic objective

Phase II scales the Phase I demonstrations into clinically meaningful capabilities and begins serious cross-bucket integration. By end-of-phase, partial AIHS demonstrations for narrow indications should be in clinical use. The capability envelope is roughly: comprehensive diagnostics for major chronic diseases, interpretation with patient-specific predictions for well-characterised conditions, and coordinated multi-target therapeutic interventions for selected applications.

The most important question Phase II answers is whether the central scientific challenge — causal disease modelling — will yield to current approaches. This is the question on which the maximum eventual AIHS capability hinges, and it gates the Phase II decision review.

Bucket A · Diagnostic priorities

A1 (in vivo single-cell readout) moves from animal proof-of-concept to first human clinical demonstrations. Initial applications likely concentrate on accessible tissues — gut, skin, oral mucosa, retinal imaging extensions — before tackling internal organs. By end-of-phase, in vivo cellular profiling at clinical scale exists for at least two human tissue types.

A2 (whole-body molecular mapping) sees its first serious work in Phase II. The Phase II target is non-human primate whole-body cellular-resolution mapping — proof that the modality is physically achievable in a primate-sized body. Human applications remain a Phase III target. This is the longest-horizon bucket-A advance and the one most likely to slip.

A3 (continuous temporal sampling) matures fully in this phase. Multi-parameter implantable sensor arrays become a standard component of intensive care. Continuous metabolomic monitoring extends from glucose to a broader panel of metabolites. The Phase II deliverable is whole-body sensor networks with edge computing for real-time multi-modal monitoring.

A4 (integrated state representation) reaches deployment readiness. Multimodal medical foundation models trained across imaging, lab values, clinical notes, and genomic data become routine clinical tools. The Phase II open question is whether these models can produce uncertainty estimates well-calibrated enough for autonomous use, which is critical for safe AIHS operation.

A5 (distributed pathogen and damage detection) develops in this phase. Engineered sentinel cells move into clinical trial for specific applications — likely starting with cancer surveillance and infection detection. Comprehensive pathogen detection remains a Phase III target.

Bucket B · Interpretation priorities

B3 (causal disease modelling) is the defining work of Phase II. The Phase I foundation should now support large-scale perturbation biology programmes producing the kind of interventional data that causal modelling needs. By mid-phase, the question of whether B3 will yield should become answerable.

The candidate scenarios at this point: B3 yields to scaled perturbation biology plus foundation models (best case — clinical decision-support with counterfactual capability emerges); B3 yields partially in well-characterised disease classes but not generally (most likely case — sufficient for partial AIHS); B3 does not yield to current approaches and requires fundamental theoretical advances (worst case — partial AIHS becomes the realistic ceiling for the foreseeable future).

B1 (complete variant effect prediction) reaches comprehensive coverage for coding variants and tractable progress on non-coding regulatory variants. By end-of-phase, variant interpretation for clinical use should be largely solved for well-characterised genes and statistically tractable for novel variants with explicit uncertainty.

B4 (digital twins) work concentrates on patient-specific digital twins for common chronic diseases — cardiovascular twins, metabolic twins, and increasingly integrated multi-organ twins. The Phase II deliverable is digital twins capable of intervention simulation for at least three major disease categories.

B5 (continuous learning across patients) moves from demonstration to widespread deployment. Federated medical AI infrastructure becomes routine. Privacy-preserving training across international jurisdictions becomes operationally feasible.

B2 (real-time epigenomic state inference) remains gated on Bucket A progress. Liquid-biopsy-based epigenomic monitoring matures into routine cancer surveillance. In vivo epigenomic profiling remains a Phase III target.

Bucket C · Therapeutic priorities

C1 (multi-target coordinated intervention) matures in this phase. Phase II delivers approved multi-target coordinated therapies for at least three specific complex diseases — likely candidates include certain cancers, autoimmune conditions, and metabolic syndromes. The principle of platform-orchestrated rather than clinician-orchestrated combinations becomes accepted clinical practice.

C2 (universal tissue access) advances substantially. Reliable CNS delivery for gene therapies becomes routine clinical capability. Methods for crossing other privileged barriers — testis, joint cartilage, retinal — mature. By end-of-phase, no major tissue should be effectively inaccessible to therapeutic delivery.

C3 (architectural tissue reconstruction) sees substantial work but full success remains a Phase III target. The Phase II milestone is in vivo organ regeneration in non-human primates, including vascularised tissue reconstruction for at least one organ type. Human applications remain experimental.

C5 (bounded reversible agents) reaches platform maturity. By end-of-phase, every therapeutic agent in development should have built-in reversibility designed in from the start. The Phase II deliverable is platform-level standards and validated kill-switch designs across therapy classes.

C6 (multi-layered safety architecture) moves from standards to validated deployment. By end-of-phase, multi-layered safety architectures should be a regulatory expectation for autonomous medical systems, with validation methodologies established and accepted.

C4 (neural reconnection) sees Phase II progress on peripheral nerve injury and limited progress on spinal cord injury. CNS reconnection at scale remains beyond reach in this phase — this is the advance most likely to remain partial indefinitely.

C7 (accelerated healing energy supply) emerges as an active research area in this phase. The question of where the ATP comes from gets serious attention. Phase II progress is mostly conceptual — establishing that the problem exists, characterising the thermodynamic constraints, identifying candidate approaches.

Integration milestones

Partial AIHS demonstrations for narrow indications. By end-of-phase, partial AIHS implementations — meeting some of the four AIHS criteria, not all — in clinical use for specific indications. Likely starting points: sepsis management, certain cancer treatments, intensive care for trauma patients.
Regulatory pathway maturity for closed-loop autonomous systems. FDA and equivalent international regulators have established review pathways for closed-loop autonomous medical systems. Approvals proceeding routinely for narrow-scope applications.
Cross-bucket benchmarking infrastructure. Shared evaluation environments where new AIHS-relevant capabilities can be benchmarked against integrated clinical scenarios. The reference test bed from Phase I, scaled up.
Federated training infrastructure at international scale. Continuous-learning systems training across multiple jurisdictions with privacy-preserving infrastructure. Demonstrates the institutional and technical viability of population-scale medical AI.
First closed-loop demonstrations beyond single-organ scope. Stretch goal: demonstrations of closed-loop autonomous medical action coordinating across multiple organ systems for the same patient. Likely to remain partial through Phase II.

Funding allocation profile

Phase II · Approximate allocation

A · 22%

B · 30%

C · 23%

Safety · 13%

Integ · 12%

Diagnostics

Interpretation

Therapeutics

Safety architecture

Integration

Phase II allocation shifts notably toward Bucket B reflecting the increased investment in causal modelling and digital twins. The B3 push specifically should be funded at a level commensurate with its risk-adjusted importance — it is high-risk but very high-value work.

Dominant risks during Phase II

B3 failure to progress. If causal disease modelling does not yield to expected approaches, the maximum eventual AIHS capability is permanently limited. This is the dominant scientific risk of the entire programme and is concentrated in this phase.

Premature deployment. Partial AIHS systems in clinical use create patient-safety pressure to expand scope faster than the science supports. Discipline about staying within validated capability envelopes is essential and will face institutional pressure.

Capability outrunning safety architecture. Phase II is when novel autonomous capabilities first start arriving in clinical reality. If safety architecture work has not kept pace, this is when deployment errors become serious. Phase I underinvestment in safety architecture (if it occurred) bites hardest in Phase II.

Public trust events. The first highly publicised failure of an autonomous medical system could set the field back substantially regardless of how technically isolated the failure was. Communication strategy, transparent failure reporting, and proactive safety culture matter as much as technical work.

End-of-Phase II Gate · ~Year 15

B3 viability assessment

Has causal disease modelling reached the level needed for meaningful counterfactual reasoning in patient care?

This is the most consequential review moment in the entire programme. If yes, full AIHS becomes plausible by Phase IV. If no, AIHS capability will be permanently limited to mechanistically understood domains — which is still transformative but represents a different programme strategically.

If B3 has yielded

Proceed to Phase III with confidence. Scale toward full AIHS prototypes. Begin serious investment in the longest-horizon advances (C4, C3, A2).

If B3 has not yielded

Pivot to maximising value within partial AIHS scope. Reduce investment in advances dependent on B3. Communicate revised capability ceiling publicly and honestly.

Research Questions

Open questions from the field

Detail-level questions a researcher or programme planner landing on this page would likely ask. Expand each to see the answer.

Phase II is built around the B3 question — what exactly is being tested, and how would the answer be read?

The central question Phase II answers is whether causal disease modelling (B3) can reach the level needed for genuine counterfactual reasoning in care — the "if we intervene here, what happens elsewhere?" capability. This is the hinge the whole programme's ambition turns on.

What is actually being tested is not a single experiment but an accumulating body of evidence: do causal models, trained on the perturbation data and longitudinal cohorts built in Phase I, begin to outperform pattern-matching on individualised treatment prediction? Do hybrid mechanistic-learned models extrapolate correctly to intervention regimes outside their training distribution? Does the capability generalise beyond the handful of well-characterised diseases, or stall at them?

The answer is read at the end-of-Phase-II gate, and crucially it is not binary. A strong yes points Phase III and IV toward universal capability. A clear no redirects the programme toward maximising value within mechanistically-understood domains — still transformative, just bounded. The most likely reading is partial: B3 works for some disease classes and not others, in which case the gate's job is to characterise which and set realistic scope. The phase is deliberately structured so that this assessment can be made honestly rather than retrofitted to justify continuation.

What does a 'partial AIHS for narrow indications' actually look like in clinical use by end of Phase II?

The body of this phase sets the target as partial AIHS demonstrations in clinical use for narrow indications. Concretely, that means closed-loop or near-closed-loop systems operating in well-bounded, high-value clinical contexts — not a general-purpose healing platform.

Sepsis is the recurring example: continuous multi-parameter monitoring (A3), pathogen and damage detection (A5), interpretation through a causal model of sepsis progression (partial B3), and coordinated antimicrobial-plus-supportive intervention (C1), all under the safety architecture (C6). Similar narrow systems would target specific cancers, select chronic diseases with well-understood mechanisms, and managed post-surgical recovery.

The defining features are that each operates within a validated scope, handles the common cases autonomously while routing the unusual ones to clinicians, and is reversible and overridable throughout. These are recognisably AIHS — continuous, integrated, autonomous-within-bounds — but each covers one indication rather than the universal capability the full vision implies. They are the proof that the integrated approach delivers clinical value, which is the foundation Phase III builds on.

Why is Phase II the longest and most expensive phase, and where does the money go?

Phase II is where the field attempts its hardest science (B3) at the same time as scaling its most mature engineering into clinical systems, which is an inherently expensive combination — and where the largest share of the programme's cost concentrates.

The spending divides across three fronts running in parallel. The causal-modelling push is the riskiest and least predictable: large-scale perturbation biology, the compute and data infrastructure for hybrid models, and the long iterative cycles of validation against clinical outcomes. Clinical scaling of the partial-AIHS demonstrations is expensive in the way all clinical translation is — trials, manufacturing, regulatory engagement — multiplied across several narrow indications. Integration infrastructure — the connective tissue letting the buckets operate as closed loops — is itself a major build.

The reason the study accepts this cost structure is that Phase II is where the programme's central uncertainty is resolved. Spending less, or rushing the phase, would mean arriving at the most important decision gate without the evidence to answer its question honestly. The phase is long because the B3 question cannot be answered quickly, and expensive because answering it requires building real clinical systems rather than running isolated experiments. The decision gates exist precisely so this large investment can be recalibrated as evidence accumulates rather than committed blindly.

Referenced from across the study

Topic Roadmap — → Full Phase II detail with funding allocation, risks, and decision gate Phase III — Integration and scale (Years 15–30) The third phase brings full AIHS prototypes for select clinical applications.
Advance B3 — The Phase II and Phase IV roadmaps both treat this partial-capability scenario as a successful programme outcome, not a fallback — the universal-capability scenario is the upside case, not the baseline expectation.

Continue

Next phase

Phase III — Integration and Scale

Years 15–30. Full AIHS prototypes arrive for select indications. The bucket-integration question is settled.

Back

Phase I — Foundation

The infrastructure phase that this builds on.

Sideways

Decision gates

All three review moments compared and consolidated.