Claim-to-Evidence Map¶
This page tells a reviewer where to verify each important statement. A claim is strong only when its implementation, test, output and scientific basis are kept separate.
Core project claims¶
| Claim | Implementation | Verification or output | Boundary |
|---|---|---|---|
| The SDK simulates virtual diabetes scenarios | src/iints/core/simulator.py, src/iints/core/patient/ |
Core, scenario and preset tests; results.csv |
Research approximation |
| Candidate actions pass a separate safety layer | src/iints/core/supervisor.py, src/iints/core/safety/ |
Safety tests, audit events and safety report | Not proof of clinical safety |
| Physiology is deterministic code, not LLM output | src/iints/core/formula_registry.py, patient model _ode methods |
Formula registry tests and numeric-authority checks | Parameter validity still matters |
| CGM can differ from latent glucose | src/iints/core/devices/models.py |
Sensor tests and separate result columns | Generic, not vendor-equivalent |
| Runs produce inspectable evidence | src/iints/highlevel.py, src/iints/analysis/, src/iints/validation/ |
CSV, metadata, manifests, reports | Artifact set varies by command |
| Data can be contract-checked | src/iints/data/ |
MDMP certificate and quality report | Certification is not clinical approval |
| AI is advisory in explanation mode | src/iints/ai/, docs/NUMERIC_AUTHORITY.md |
Prompt/policy tests and separate AI artifacts | Human review remains required |
| Desktop app calls the same SDK | apps/iints-tauri/, src/iints_desktop/ |
Rust checks, bridge tests and desktop smoke tests | Packaging/security need maintenance |
| Edge and FPGA paths are bench-only | src/iints/live_patient/, src/iints/jetson/ |
Mock, protocol and hardware-adapter tests | No medication actuation |
Formula evidence map¶
Runtime paths in this table are relative to src/iints/core/.
| ID | Runtime authority | Scientific basis |
|---|---|---|
| F01 Bergman glucose | patient/bergman_model.py |
Bergman 1979 plus declared extensions |
| F02 remote insulin | patient/bergman_model.py |
Bergman 1979 |
| F03 plasma insulin | patient/bergman_model.py |
Bergman balance plus disabled-by-default research secretion |
| F04 subcutaneous insulin | Bergman and Hovorka patient models | Hovorka 2004 and PK abstraction |
| F05 meal absorption | Bergman and Hovorka patient models | Published Hovorka two-compartment chain; explicitly adapted three-stage Bergman branch |
| F06 Hovorka glucose mass | patient/hovorka_model.py |
Hovorka 2004 plus declared extensions |
| F07 insulin-action channels | patient/hovorka_model.py |
Hovorka 2004 action channels plus heuristic molecular/tissue sensitivity scalars |
| F08 stress/exercise | patient/hovorka_model.py |
Research pseudo-hormone abstraction |
| F09 GLUT4/NIMGU | patient/hovorka_model.py |
Exercise/GLUT4 physiology context |
| F10 circadian EGP | patient/hovorka_model.py |
Dawn-phenomenon context, gated approximation |
| F11 hypo rescue | Bergman and Hovorka patient models | Counterregulation and HAAF context |
| F12 HAAF memory | Bergman and Hovorka patient models | Cryer 2013, experimental memory state |
| F13 glucagon PK/PD | Bergman and Hovorka patient models | Published exogenous-glucagon context |
| F14 renal clearance | patient/physiology.py and both ODE models |
Renal threshold/splay context |
| F15 CGM observation | devices/models.py, SensorModel.read |
Blood-to-ISF lag and sensor context |
Canonical source:
src/iints/core/formula_registry.py. Generated human reference:
docs/FORMULA_REGISTRY.md.
Benchmark evidence map¶
Paths in this table are relative to research/eucys_pack/, except where noted.
The named study runner lives in tools/research/.
| Evidence | Repository path | Review question |
|---|---|---|
| Aggregate arm table | assets/EUCYS_RESULTS_TABLE.csv |
Do arm counts and metrics match the report? |
| Algorithm figure data | assets/EUCYS_MAIN_FIGURE.csv |
Can every plotted bar be reconstructed? |
| Main figure | assets/EUCYS_MAIN_FIGURE.png |
Does visual labelling match the table? |
| Full report source | ../EUCYS_REPORT.md |
Are protocol, results and limitations stated? |
| Final workflow | ../EUCYS_FINAL_WORKFLOW.md |
Can the benchmark and report be regenerated? |
| Study runner | run_eucys_final.sh |
Is the command path explicit? |
Scientific source map¶
| Topic | Primary source used in SDK documentation |
|---|---|
| Glycaemic targets and hypoglycaemia | ADA Standards of Care 2026, DOI 10.2337/dc26-S006 |
| Time in range | Battelino et al. 2019, DOI 10.2337/dci19-0028 |
| Bergman minimal model | Bergman et al. 1979, DOI 10.1152/ajpendo.1979.236.6.E667 |
| Hovorka model | Hovorka et al. 2004, DOI 10.1088/0967-3334/25/4/010 |
| Meal absorption | Dalla Man et al. 2007, DOI 10.1109/TBME.2007.893506 |
| CGM lag | Wentholt et al. 2004, DOI 10.1089/dia.2004.6.615 |
| Exercise in T1D | Riddell et al. 2017, DOI 10.1016/S2213-8587(17)30014-1 |
| Exercise and GLUT4 | Richter and Hargreaves 2013, DOI 10.1152/physrev.00038.2012 |
| HAAF | Cryer 2013, DOI 10.1056/NEJMra1215228 |
| Renal glucose handling | Hummel et al. 2018, DOI 10.1007/s00125-018-4656-5 |
| OhioT1DM | Marling and Bunescu 2020, CEUR paper |
| AGP interpretation | International Diabetes Center guide and TIR consensus |
The complete maintained source list is in docs/EVIDENCE_BASE.md and
docs/SOURCE_LIBRARY.md.
Mermaid diagram sources¶
The diagrams in this dossier are both embedded in the Markdown pages and stored as reusable Mermaid source:
| Diagram | Source |
|---|---|
| System architecture | docs/eucys/diagrams/system-architecture.mmd |
| Simulation step | docs/eucys/diagrams/simulation-step.mmd |
| Numeric authority | docs/eucys/diagrams/numeric-authority.mmd |
| Evidence lifecycle | docs/eucys/diagrams/evidence-lifecycle.mmd |
| Desktop bridge | docs/eucys/diagrams/desktop-bridge.mmd |
| AI boundary | docs/eucys/diagrams/ai-boundary.mmd |
| Data lifecycle | docs/eucys/diagrams/data-lifecycle.mmd |
| Cross-scale evidence | docs/eucys/diagrams/cross-scale-evidence.mmd |
| Validation ladder | docs/eucys/diagrams/validation-ladder.mmd |
Reproduction commands¶
Build this dossier:
tools/research/build_eucys_dossier.sh
Build the existing evidence PDFs:
tools/research/build_eucys_pack.sh
Run the final benchmark workflow:
tools/research/run_eucys_final.sh \
--algo algorithms/example_algorithm.py \
--output-dir results/eucys_2026 \
--seeds 1,2,3,4,5,6,7,8,9,10 \
--no-prepare-ai
Build and validate the documentation:
mkdocs build --strict
Run the principal software checks:
python3 tools/ci/check_architecture_boundaries.py
python3 -m pytest tests/ -q
mypy src/iints/
Review checklist¶
A reviewer should be able to trace:
- A sentence in the report to a table or raw trace.
- A table value to deterministic metric code.
- A simulated state to a registered equation and parameter set.
- A safety intervention to a candidate, accepted action and reason.
- A dataset to a source and transformation manifest.
- An AI statement to supplied evidence.
- A biological context view to its public source and interpretation boundary.
- A release artifact to a tagged software version.