# BioTwin Accelerator — Checklist v0.2 (typed evidence) *Supersedes v0.1. Every gate is now an instance of the shared `EvidenceType` schema in §0 — the same schema the composites accelerator imports — so the two accelerators share one machine-checkable type system instead of two prose checklists. Refusals are **derived** from the admissibility law, not hand-listed.* **Standing boundary:** generates mechanism hypotheses, ranks the next experiment, bounds its own uncertainty. Does **not** make clinical decisions; every patient-facing output defers to the treating clinician, basis visible and override-able. `authorizes_action = false` everywhere. **The one axiom (everything below is a corollary):** > An operation is **admissible** iff it is a *homomorphism* for the declared > algebra of every input's type. The result's **posture** is the **meet** of the > inputs' postures, **lowered by the functor-class** of the operation. `meet` is > `min` on the cube ladder (the tropical semiring). --- ## §0 — Shared type system (`forza/evidence_types` — imported by both accelerators) ```python # The canonical evidence type. Composites and BioTwin are INSTANCES of this, # not separate schemas. This block is the shared machine-checkable core. EvidenceType = { "id": str, "geometry": Enum("euclidean","log_ratio","circular_torus","sphere_SO3", "simplex_aitchison","sequence_free_monoid","graph", "ordinal_lattice"), "statistic": Enum("mean","median_band","lower_tail","upper_tail","moments", "quantiles","threshold","order_preserving","cardinality"), "grain": "position on a declared partial order (lattice)", # directional "rev": "resolution at which CWA/closure is valid", "conservation": str, # quantity a merge must preserve (mass, unit-norm, ...) "merge": "the value monoid (how VALUES compose)", "merge_commutative": bool, # 'mergeable sketch' == commutative monoid "merge_idempotent": bool, # A(+)A==A? set-union yes; pooled-survival NO "uncertainty": Enum("variance_add","tangent_cov","tolerance_bound_conf", "conformal_coverage","composed_lax_error"), # how BOUNDS compose "failure_mechanism": str, # UPSTREAM of statistic (mechanism selects statistic) "functor_default": Enum("strict","lax","anti"), "posture_cap": Enum("refused","unknown","candidate","certified","reconciled"), "forbidden": "DERIVED = the non-homomorphisms for (geometry, merge)", } def well_formed(t) -> bool: # internal consistency of ONE object's tuple return (statistic_consistent_with(t.failure_mechanism, t.statistic) and merge_is_homomorphism(t.geometry, t.merge) and merge_preserves(t.conservation)) def functor_class(op) -> "strict|lax|anti": # relabel/exact -> strict ; homogenize up-grain -> lax ; # disaggregate / allocate / counterfactual (manufacture finer structure) -> anti ... def admissible(op, inputs) -> bool: return (all(is_homomorphism(op, x.type) for x in inputs) # structure preserved and meet_types(inputs) is not BOTTOM) # compatible (grain/geom/stat) def result_posture(op, inputs): p = meet_posture(inputs) # = min on the cube ladder (tropical) return lower_by(p, functor_class(op)) # anti lowers hardest; strict preserves ``` `meet_types` returns `BOTTOM (⊥)` when geometries differ and neither lifts, when grains sit on incompatible lattices, or when the requested op is off the type's support (e.g. positivity fails). `⊥ ⇒ WITHHOLD`. Every refusal in this file is one of these three, not a remembered rule. --- ## §1–§10 — Gates as typed evidence objects ```yaml # Gate 1 — Molecular identity & grain type: molecular_identity geometry: ordinal_lattice statistic: n/a # identity, not a summary grain: atomistic < residue < domain < protein < complex < cell < patient conservation: entity_identity merge: exact_match merge_commutative: true; merge_idempotent: true functor_default: strict posture_cap: certified forbidden: [compare_across_grain, compare_across_binding_mechanism] # CATCH: ATP-competitive (imatinib) and allosteric-STAMP (asciminib) are # DIFFERENT TYPES -> meet = ⊥. Never share a committor/model across mechanisms. ``` ```yaml # Gate 2 — Structure & folding (emits several typed objects) type: dihedral_ensemble geometry: circular_torus # phi/psi live on a torus statistic: moments # von Mises grain: residue conservation: none merge: circular_mean merge_commutative: true; merge_idempotent: true(set) / false(weighted) uncertainty: tangent_cov # concentration, not scalar variance functor_default: lax posture_cap: certified forbidden: [arithmetic_mean, euclidean_interpolation] # DERIVED: not homs on S1 --- type: backbone_orientation geometry: sphere_SO3 merge: orientation_tensor forbidden: [arithmetic_mean, euclidean_interpolation] --- type: contact_map geometry: graph statistic: threshold merge: graph_union # CATCH: Euclidean-averaging angles is the SAME launder as averaging fiber # orientation in composites. pLDDT/PAE ride along as a declared-loss field. ``` ```yaml # Gate 3 — Binding & kinetics (TST) type: binding_deltaG_rate geometry: log_ratio # k = exp(-dG/RT) is multiplicative statistic: lower_tail # safety is tail-governed grain: complex conservation: detailed_balance # k_f/k_r = K_eq (metamorphic oracle) merge: log_space_pool uncertainty: tolerance_bound_conf failure_mechanism: barrier_crossing functor_default: lax posture_cap: candidate # -> certified only via kappa-ledger + RC validation forbidden: [linear_mean_of_rates] # DERIVED: not a hom on log_ratio (use DDSketch, not KLL) # GATE: spectral_gap present -> single kinetic story admissible; # absent -> report interval, escalate grain (the ensemble case is the norm). ``` ```yaml # Gate 4 — PK/PD bridge [the molecular->patient operator; MANDATORY] type: pkpd_bridge geometry: log_ratio # concentrations; compartments statistic: median_band + lower_tail(safety) grain: molecule -> patient # this operator IS the up-grain step rev: patient conservation: mass_balance # dose accounting merge: compartment_compose uncertainty: composed_lax_error functor_default: lax # bridges grain up; posture degrades with model error posture_cap: candidate # -> certified only against measured trough/response # CATCH: any molecular-grain evidence merged with patient-grain evidence has # meet = ⊥ UNLESS routed through this object. No skipping the bridge. ``` ```yaml # Gate 5 — Dynamics & network (MSM) type: conformational_msm geometry: graph # stochastic transition matrix statistic: threshold + moments # dwell times, eigenvalues grain: microstate -> metastable # PCCA+ coarse-grain merge: count_pool merge_idempotent: false # adding the SAME trajectory twice inflates counts uncertainty: composed_lax_error functor_default: lax posture_cap: candidate forbidden: [merge_without_markov_order_test] # Anderson-Goodman gate first # GATE: spectral_gap = metastability certificate = the validity of a single rate. ``` ```yaml # Gate 6 — Causal / counterfactual [THE ANTI-FUNCTOR OPERATION — see §6 below] type: counterfactual_claim geometry: n/a (an OPERATION, do(.), not a summary) functor_default: anti # manufactures a road-not-taken from one observed road posture_cap: candidate # <= posture of the SCM it rides on (Gate 4 kappa-ledger) default_verdict: withhold ``` ```yaml # Gate 7 — Patient / phenotypic bridging type: transcript_trajectory # BCR-ABL1 on the International Scale geometry: log_ratio # response is DEFINED in logs (MMR/MR4/MR4.5) statistic: lower_tail # deep response = tail grain: patient merge: log_space_pool merge_idempotent: false # pooling the SAME blood draw twice falsely tightens CI uncertainty: tolerance_bound_conf -> conformal functor_default: lax posture_cap: certified forbidden: [linear_tdigest, double_count] # DDSketch, not KLL; idempotency guard --- type: hrv_wearable geometry: log_ratio + circular # lnRMSSD + phase evidence_tier: LOW # consumer device; noisy, device-dependent posture_cap: candidate # may not launder up to a clinical signal --- privacy: single_patient_twin IS PHI merge_cross_patient: requires differential_privacy # else leakage functor fires ``` ```yaml # Gate 8 — Evidence type & validation tier_lattice: clinical_assay > cell_line > in_vitro > wearable merge_rule: meet floors to the WEAKER tier validation: analytical_validation (reconstruction) vs clinical_validation (outcome) ladder: in_silico < in_vitro < cell < (animal) < retrospective_match < prospective_n_of_1 claim_rule: an object claims only up to its highest VALIDATED rung ``` ```yaml # Gate 9 — Conformal calibration [safety-critical] type: calibrated_interval uncertainty: conformal_coverage n1_rule: conformal_under_covariate_shift (weighted/Mondrian) # she is not the cohort posture_cap: certified # ONLY if measured coverage >= declared coverage # NOTE: the covariate-shift edge here is the SAME object as Gate 6's # transportability edge (a selection-diagram arrow). Gate 9 and Gate 6 merge. ``` ```yaml # Gate 10 — Authority boundary ladder->posture: mechanism_insight -> candidate hypothesis + next_test -> candidate (+ MEO probe attached) counterfactual -> anti-functor (see §6); default withhold subpopulation_validated -> certified (cohort only) this_patient_support -> reconciled (clinician-in-loop, basis visible, override-able) outside_validated_envelope: meet = ⊥ -> WITHHOLD # e.g. ~22y exposure extrapolation authorizes_action: false ``` --- ## §6 — Gate 6 as the anti-functor operation `do(strategy)` A counterfactual manufactures a road-not-taken (finer structure) from a single observed road, so it is an **anti-functor**: posture **floors down** by construction and is admissible only through a validated model whose `kappa-ledger` caps it. Concretely: ```yaml operation: do(strategy) # e.g. do(asciminib, dose, schedule) functor_class: anti admissible_iff: ONE of (a) validated mechanistic SCM # abduction -> action -> prediction posture <= SCM.kappa_ledger # never more certified than the mechanism (b) target_trial_emulation with assumptions CHECKED # -> target_trial_boundary.py else: REFUSE and emit n_of_1_design # the MEO/value-of-information tail time_varying_rule: # CML is chronic: dose_t -> transcript_t -> dose_{t+1} grain_time_lattice: directional (treatment-confounder feedback) ONLY_homomorphic_merge: g_method # marginal structural model / IPTW / g-formula forbidden: naive_adjustment # not a hom on the time lattice -> points WRONG way assumptions_as_refusals: # each is a meet=⊥ or a required declaration consistency_SUTVA: strategy must be a concrete dose+schedule # else ⊥ positivity_overlap: no support (e.g. rare T315I profile) -> ⊥ -> withhold exchangeability: unverifiable -> BOUND it: report E-value + negative_controls if a plausible confounder (adherence) flips the sign -> refuse transportability: covariate-shift edge declared + reweighted # == Gate 9 refutation_before_emit: # DoWhy-style refuters; survive to earn posture [placebo_treatment, random_common_cause, subset_stability, unobserved_confounder_sens] output_shape: # non-device CDS: clinician can inspect + override mechanistic_counterfactual + E_value + assumptions + confirming_n_of_1_design ``` --- ## §Overlap — one type system, two accelerators (the consolidation) The point of v0.2: composites and BioTwin are now **instances of the same types**, so the shared engines are literal, not analogical. | Shared `EvidenceType` | Composites instance | BioTwin instance | |---|---|---| | `geometry:lower_tail` + `uncertainty:tolerance/conformal` | strength -> CMH-17 A/B-basis | occupancy / toxicity -> conformal-under-shift | | `geometry:sphere_SO3`, `merge:orientation_tensor` | fiber ODF (Advani-Tucker) | backbone orientation | | `geometry:sequence_free_monoid`, order-preserving | layup schedule (CLT D-matrix) | peptide / residue sequence | | `geometry:graph`, `statistic:threshold` | connectivity / percolation | contact map / MSM | | `functor:anti` | allocation / disaggregation | **`do()` counterfactual** | | `merge_idempotent:false` guard | pooled coupon survival | pooled blood-draw / trajectory | **The one muscle both still need** is the `lower_tail`/`tolerance` type's engine: build `tail_uncertainty` once (weakest-link -> lower-tolerance-bound -> conformal), register it against `geometry:lower_tail`, and both accelerators inherit it — CMH-17 for composites, conformal-under-shift for the patient. --- **Status:** v0.2 — typed. Both files should now `import forza/evidence_types` (§0) rather than restate a checklist. **Next:** point `COMPOSITES_SKETCHING_AND_EVIDENCE_CHECKLIST.md` at the same §0, and implement `well_formed` + `meet_types` so the refusals run instead of read.