Substrate-First Clinical AI (SFClinAI)

A position note arguing that clinical AI's safety properties must be enforced by a symbolic substrate, not by the behavior of learned components — with a coupled two-layer (patient + institutional) architecture, a temporal-evolution axis, six worked examples, and prior-art mapping.

This repository holds the prose of that position, plus the clinlat Rust kernel (v0.2.0-alpha.0; v0.1.0 on crates.io): a symbolic substrate for clinical hypothesis refinement with sound deduction operators, ontology-bounded atoms, and auditable provenance.

A working draft (NOTE.md, v0.12.0-draft) and accompanying Mermaid diagrams (ARCHITECTURE.md, v0.2.0-draft) staking out a falsifiable diagnostic claim: clinical AI is bottlenecked less by model size than by the absence of explicit, auditable, reversible belief-state architectures that can update medical hypotheses under hard safety constraints (NOTE.md §3).

The architecture rests on eighteen load-bearing principles — five for the patient-state substrate (§4A), five for the institutional-state substrate (§4B), three for the interaction semantics between them (§4C), and five for temporal evolution and substrate currency (§4D). Each is stated separately so it can be attacked separately.

Not a product, startup thesis, or published paper. The note is explicitly a working draft for scrutiny, not for citation (NOTE.md header). The clinlat kernel is an early research substrate — it discharges patient-substrate proof obligations in code, but it is not a clinical tool: no deployed system, no patient data, no regulatory engagement. The work is a research program with a 2–4 year horizon for narrow applications and 5–8+ years for broader diagnostic reasoning (§7). Originality claims carry an explicit obsolescence window and are re-verified quarterly; the current assessment is valid through 2026-08-24 (NOTE.md header; §6 closing).

1. NOTE.md §TL;DR through §3 — the problem, why scaling doesn't fix it, the principle stated plainly.
2. ARCHITECTURE.md — five Mermaid diagrams. Read these alongside §4 of the note.
3. NOTE.md §4 — the eighteen load-bearing principles (4A patient ×5, 4B institutional ×5, 4C interaction ×3, 4D temporal ×5).
4. NOTE.md §7E.1–7E.6 — six worked examples: Sepsis-3 prognostic stratification, KDIGO AKI staging, Wells/PE with sequential testing, CURB-65 disposition, chronic depression as an anti-example, and the live SSC 2021 → 2026 guideline transition through the substrate.
5. NOTE.md §6 and §8 — prior art, originality narrowing after substantial verification, and honest limitations.

Short on time: stopping after step 3 still holds the load-bearing claim.

Formal-methods readers: SPEC.md sits between the note and the diagrams in the document stack (NOTE → SPEC → ARCHITECTURE). It carries the typed definitions, invariants, and proof obligations behind the eighteen principles, each annotated [formalizes: NOTE.md §X.Y]. The clinlat kernel is where the §2 patient-substrate obligations are discharged in code.

SFClinAI/
├── NOTE.md             # Position note, v0.12.0-draft
├── ARCHITECTURE.md     # Five Mermaid diagrams, v0.2.0-draft
├── SPEC.md             # Engineering formalization, v0.3.0-draft
├── clinlat/            # Rust substrate kernel, v0.2.0-alpha.0
│   ├── Cargo.toml
│   ├── src/
│   ├── README.md       # Kernel documentation
│   └── docs/           # Soundness arguments
├── README.md           # This file
└── LICENSE             # CC BY 4.0 (verbatim legal code)

Two substrates, coupled, evolving in time.

The patient-state substrate carries a refinable lattice of clinical hypotheses with sound deduction operators, ontology-bounded candidates, first-class abstention, and auditable provenance; the learned component is a constrained refinement proposer, not a decision-maker (§4A.1–4A.5).

The institutional-state substrate mirrors this for cross-patient resource allocation — beds, formulary, lab cycles, on-call coverage — with capacity-update operators, allocation abstention, and capacity-learned components as constrained proposers (§4B.1–4B.5).

The two communicate through a defined event interface with joint licensing of recommendations and joint abstention as a first-class output: when patient-locally-optimal diverges from institutionally-feasible, the system produces both with explicit diff rather than silently downgrading (§4C).

All substrate components are versioned; evidence currency is architectural signal; operator-set changes are sound by construction; guideline updates propagate into active care only through clinician-mediated re-review (silent drift is structurally forbidden); provenance is evolution-aware (§4D.1–4D.5).

The composite is positioned as a route from UNDCS-class to CCS-class under the Tan et al. (2026) regulatory taxonomy, making LLM-class components evaluable under existing FDA-SaMD pathways (§5).

• NOTE.md: v0.12.0-draft. Working draft for scrutiny. Not for citation.
• SPEC.md: v0.3.0-draft. The formalization layer between the note and the diagrams: 57 named definitions (DEF-*), 23 invariants (INV-*), 17 proof obligations (OBL-*), 3 mathematical commitments (MC-*), and 12 open questions (OQ-*). Each definition is annotated [formalizes: NOTE.md §X.Y], and §8 bidirectional traceability covers all eighteen principles with no known coverage gaps. The clinlat kernel discharges the §2 patient-substrate obligations in running code.
• ARCHITECTURE.md: v0.2.0-draft. Re-anchored to SPEC.md v0.3.0-draft: principle-to-formalization map covering all eighteen principles with criticality tier (P/S/F) and SPEC.md DEF/INV/OBL anchors; Diagram 4 now renders the institutional symmetric re-review path (DEF-TE-06b) alongside the patient-side clinician-mediated path.
• clinlat: v0.2.0-alpha.0 (v0.1.0 published to crates.io; the v0.2.0 milestone M1 — patient-substrate completion — is in progress). Patient-state substrate with Hyp refinement poset, Outcome<H,A> result type, and the SOFA-3 respiratory operator. Phases 0–3 of M1 have landed: an Atom type backed by four OntologyAdapter implementations (SNOMED CT, RxNorm, LOINC, ICD-11) replacing string atom IDs; a typed Provenance carrier with version-respecting derivation chains (OBL-PS-04); and the Galois connection α_PS / γ_PS property-tested for the adjunction laws (OBL-PS-02). Remaining for the v0.2.0 cut: OperatorSet composition (Phase 4); three further operators — KDIGO AKI, Wells/PE, CURB-65 (Phase 5); and the SOFA-respiratory property-test-tier upgrade (Phase 6). See clinlat/README.md for quick start, clinlat/CHANGELOG.md for version history, and clinlat/docs/operators/sofa_resp_soundness.md for the SOFA soundness argument.
• No deployed system. No regulatory engagement.
• Originality assessment timestamped 2026-05-24, valid through 2026-08-24 (quarterly re-verification cadence, per NOTE.md header and §6 closing).

Please cite the specific version and date accessed, and treat as a working draft:

Sucandra, K. (2026). Substrate-First Clinical AI: A Position Note (v0.12.0-draft) [Working draft]. https://github.com/SHA888/SFClinAI

Author byline: Kresna Sucandra, MD. The note's header asks that the draft not be cited as a settled position; if citation is unavoidable, cite as a working draft.

The note exists to be argued with (§9). In order of preference:

1. A demonstration that the synthesis is already published and was missed in the §6 prior-art mapping.
2. A demonstration that one of the eighteen principles in §4 is wrong or unnecessary — they are stated separately so they can be attacked separately.
3. A demonstration that the substrate-first framing fails on a clinical decision the six worked examples (§7E.1–7E.6) did not cover.

Issues for short critiques; for long-form, open a PR and tag it as a long-form critique.

• Prose and diagrams (NOTE.md, ARCHITECTURE.md, SPEC.md, README.md) are licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). See LICENSE for the preamble and full legal code.
• Code (including the clinlat/ crate and any future code) is licensed under MIT OR Apache-2.0. See LICENSE-MIT and LICENSE-APACHE for the full text of each license.

Attribution: Kresna Sucandra, MD.