Method‑SoS‑LOG — MethodFamily Evidence & Maturity
Pattern C.23 · Stable Part C - Kernel Extension Specifications
LOG (logic) for deductive shells for admissibility First use expansion: SoS‑LOG = Science‑of‑Science LOG (LEX short‑form discipline applied).
HomeContext. For this pattern, HomeContext means the U.BoundedContext where a MethodFamily is registered (LEX D.CTX).
Builds on. G.5 (MethodFamily registry/selector), G.4 (Acceptance & EvidenceProfiles), C.22 (TaskSignature S2), C.18 NQD‑CAL (QD/illumination), C.19 E/E‑LOG (emitters/policies), B.3 (Assurance lanes & R_eff), A.10 (Evidence Graph Ref), E.10 (LEX), E.18 (GateCrossing / CrossingBundle visibility). Coordinates with. G.6 (EvidenceGraph), G.8 (LOG bundling), G.9 (Parity), G.11 (Refresh).
Families of methods compete inside a CG‑Frame. The selector (G.5) must admit, degrade, or abstain per family without universal scores, using typed problem descriptors and auditable evidence. Maturity of a family (how far it has travelled from “clever idea” to “run‑safe”) must be visible to LOG rules yet separate from thresholds (which live only in AcceptanceClauses, G.4).
Keywords
- MethodFamily
- evidence
- maturity
- SoS-LOG
- admit
- degrade
- abstain
- selector.
Relations
Content
Problem frame
Families of methods compete inside a CG‑Frame. The selector (G.5) must admit, degrade, or abstain per family without universal scores, using typed problem descriptors and auditable evidence. Maturity of a family (how far it has travelled from “clever idea” to “run‑safe”) must be visible to LOG rules yet separate from thresholds (which live only in AcceptanceClauses, G.4).
Problem
Unstructured “readiness” stories and undisciplined evidence lead to:
- (i) Illicit scalarisation across mixed scale types,
- (ii) Prose‑only gating that a dispatcher cannot execute,
- (iii) Cross‑Context reuse without Bridges/CL, and
- (iv) Immature families leaking into production.
We need a notation‑independent LOG layer that turns TaskSignature (S2) + EvidenceProfiles into executable rules for admit / degrade / abstain, routing any CL penalties to
R_effonly (never mutating F/G).
Forces
- Pluralism vs. dispatchability. Competing Traditions expose different invariants; selection must compare without semantic flattening.
- Maturity vs. opportunity. Open‑ended exploration (E/E‑LOG) must coexist with run‑safe exploitation; immature ≠ forbidden → provide safe degrade paths.
- Unknowns (tri‑state). Missing or
unknownS2 fields must propagate explicitly to degrade/abstain/sandbox; no silent coercions. - Lexical discipline. Head‑anchoring, I/D/S separation, Bridge hygiene; no tool names in Core.
Solution — Method‑SoS‑LOG: deductive shells over Eligibility & Evidence
Objects & heads (LEX/I‑D‑S)
Tech heads; Plain twins are published via UTS.
MethodFamily (registered in G.5) carries Eligibility and artefact identity; MaturityCard (this pattern) carries evidence‑aware maturity; SoS‑LOG.Rule (this pattern) is an executable rule schema that returns one of {Admit | Degrade(mode) | Abstain} for a (TaskSignature, MethodFamily) pair. Descriptions live as …Description; when harnessed they become …Spec.
Rule schema (normative)
For each MethodFamily f, author an executable rule set:
with the following branch obligations:
R0 — CG‑Spec gate (precondition, HomeContext). Before R1–R3, verify CG‑Spec.MinimalEvidence for every CHR characteristic referenced by f’s Acceptance/Flows in the home Context; failure ⇒ Abstain with reasons (no silent sandbox). Publish the CG‑Spec ids consulted.
Rationale: selector legality requires the CG‑Spec gate to be explicit, not implicit in prose. Publish associated ReferencePlane notes alongside the consulted ids.
R0.QD — QD/OEE pre‑gates (if applicable). If S2 declares BehaviorSpaceRef/ArchiveConfig/EmitterPolicyRef or PortfolioMode=Archive, verify:
(i) CharacteristicSpaceRef characteristics are CHR‑typed, d≥2, ReferencePlane per characteristic declared;
(ii) ArchiveConfig is lawful (topology, resolution, K>0, InsertionPolicyRef, DistanceDef with edition id and declared metric/pseudometric status);
(iii) EmitterPolicyRef present (with edition id);
(iv) DominanceRegime present; if absent, default= ParetoOnly.
Failure of any ⇒ Abstain with reasons.
R1 — Admit. Admit IFF
(a) S2 satisfies Eligibility predicates of f (tri‑state aware),
(b) EvidenceProfile minima referenced by Acceptance/Flows for f are met (lanes/anchors/freshness) in the home Context (post R0),
(c) all relevant CAL.AcceptanceClauses (G.4) evaluate to true under lawful CHR comparisons,
(d) any maturity gating (e.g., a floor on Maturity rungs) is expressed as an AcceptanceClause and referenced here by id (no thresholds inside LOG).
LOG never sets thresholds; it only executes and cites Acceptance verdicts.
R2 — Degrade. If (a) holds but (b) or (c) is partially satisfied or unknown, return Degrade(mode) where mode ∈ {scope‑narrow | sandbox | probe‑only} and emit scope notes (USM Scope(G), Γ_time). Record which S2 unknowns or Evidence minima caused the degrade. LOG‑Degrade never changes CHR scales or planes; it narrows Scope (G) or execution mode.
Note (CAL vs LOG). CAL‑level degrade.order (fall‑back to order‑only comparisons) is governed by G.4/CG‑Spec and is not a LOG mode. SoS‑LOG never overrides CAL outcomes; a LOG branch only narrows Scope(G) or execution mode (e.g., sandbox, probe‑only), it does not alter CHR scales or admissible orders.
probe‑only MUST cite an E/E‑LOG policy id (exploration budget) and Acceptance‑bound guards.
R3 — Abstain. If S2 violates Eligibility or R0 fails, return Abstain (with reasons). Abstain is mandatory on illegal CHR operations (e.g., ordinal means) and when Bridge/CL requirements are unmet.
R4 — CL routing. Any cross‑Context/plane reuse must cite Bridge ids (with loss notes). Apply Φ(CL) and (if planes differ) Φ_plane that are monotone, bounded, table‑backed; publish policy‑ids in the SCR; penalties reduce R_eff only; F/G must remain invariant.
R5 — Proof hooks. Every branch MUST cite Evidence Graph Ref (A.10), lane tags (TA/VA/LA), freshness windows, and (if bridged) Bridge ids + loss notes; the decision is SCR‑visible. When G.6 EvidenceGraph is present, also publish EvidenceGraph path id(s) for the branch (admit/degrade/abstain). No self‑evidence.
R6 — QD portfolio semantics (if applicable). If PortfolioMode=Archive, Admit may return a QD archive (per ArchiveConfig) instead of only a Pareto set. Unless CAL authorises DominanceRegime=ParetoPlusIllumination (policy‑id recorded in SCR), IlluminationSummary is a report‑only telemetry summary and any coverage/regret are telemetry metrics (reported) that do not affect dominance.
R7 — GeneratorFamily branches (open‑ended). If S2 includes GeneratorIntent, SoS‑LOG MUST:
(i) verify EnvironmentValidityRegion is declared and lawful;
(ii) verify TransferRulesRef exists; if unknown ⇒ Degrade(scope‑narrow) or Abstain per family policy;
(iii) treat the selection surface as pairs {environment, method}; publish coverage/regret and IlluminationSummary as report‑only telemetry (IlluminationSummary = telemetry summary; coverage/regret = telemetry metrics); dominance participation per R6.
R8 — Telemetry & Refresh hooks. On any illumination increase or archive change, publish edition increments for CharacteristicSpaceRef/DistanceDefRef and the policy‑id (Emitter/Acceptance) that caused the change; expose PathSliceId for refresh/decay in SCR.
Aphorism. “Admit on lawfulness and sufficiency; degrade on uncertainty; abstain on illegality.”
Maturity ladder (poset, not a scalar; Description, not Spec)
Publish a MethodFamily.MaturityCardDescription@Context (UTS enum ids; Scale kind = ordinal; ReferencePlane declared). Do not embed thresholds here; any maturity floors used for admission are authored as G.4 AcceptanceClause and referenced by id from R1.
- L0 — Anecdotal. Claims exist; lanes sparse; examples ad‑hoc.
- L1 — Worked‑Examples. Multiple worked examples with lane tags and Scope slices declared; no replication yet.
- L2 — Replicated. Independent replication(s) in distinct Context slices (publish D.CTX + UTS rows), lane separation observed, decay windows explicit.
- L3 — Benchmark‑Severe. Repeated wins or parity on community baselines or severe tests; cross‑Tradition bridges declared with loss notes.
Optional rung (for QD/OEE‑heavy families; ordinal, closed enum):
- L4 — QD‑Hardened. Archive stability under declared InsertionPolicy/DistanceDef editions; reproducible IlluminationSummary improvements under controlled budgets; OEE generators pass EnvironmentValidityRegion severe tests.
Norms. M1. The ladder is lane‑aware (TA/VA/LA) and freshness‑aware; it is not a global numeric score. Declare Scale kind=ordinal and the closed enumeration of rungs; register the enum at UTS (twin labels; editioned). M2. Transitions MUST be justified by EvidenceGraph paths (once G.6 is available) and UTS publication; missing anchors ⇒ no advance. M3. Any maturity floor used for admission (e.g., “run‑critical Context requires ≥L2”) MUST be authored as a CAL.AcceptanceClause and referenced by id from R1; SoS‑LOG does not embed thresholds. M4. Declare ReferencePlane for the MaturityCard; on ReferencePlane crossings apply published Φ_plane policy (penalty to R_eff only), with Bridge id and loss notes.
Rationale note. Treating maturity as a poset aligns with B.3’s lane calculus and avoids scalarisation across ordinal/ratio scales; assurance penalties remain on R, never F/G.
Unknowns & Shift classes (tri‑state discipline)
U1. (LEX). Enumerations for Degrade(mode) and Maturity rungs MUST be declared as closed value sets and registered at UTS (twin labels). Lexical SD (E.10) applies.
U2. Every S2 field is tri‑state; unknown MUST map to a branch (Degrade or Abstain) declared on the family (no coercions). Each branch publishes a branch‑id and (where used) a mode from a closed enum registered at UTS (LEX enum clarity).
U3. ShiftClass semantics follow C.22. If ShiftClass ∈ {covariate‑shift, concept‑drift, adversarial} or unknown, default outcome is Degrade(scope‑narrow) unless a CAL.AcceptanceClause explicitly guards the regime.
Publication & wiring
W1. Each family publishes a MaturityCardDescription@Context (UTS twin labels; ReferencePlane declared) and registers SoS‑LOG rule ids; editions are versioned and RSCR‑tested for branch‑coverage (Admit/Degrade/Abstain, unknown paths). Φ(CL)/Φ_plane policy‑ids must be present in SCR where applicable.
W2. Admissibility Ledger. Publish an AdmissibilityLedger@Context: rows = (MethodFamilyId, RuleId, MaturityRung, BranchIds, BridgeIds, ΦPolicyIds, EvidenceGraphPathIds?, DominanceRegime, PortfolioMode, Edition), UTS‑registered; this ledger is the selector‑facing export.
W3. Strategy token. Do not mint a U.Type “Strategy”; strategy remains a composition inside G.5 (Compose) under E/E‑LOG.
W4. Selector (G.5) consumes these rules; results appear in the Dispatcher Report with reasons in/out and cited anchors/bridges.
Archetypal Grounding (Tell–Show–Show)
(Plain register for pedagogy; Core remains notation‑independent per E.10/E.8.)
Show‑1 - Continuous dynamics (ODE task).
S2 excerpt. DataShape=ODE; stiff?=unknown; Size≈10^3; Objective={↓error@ratio, ↑throughput@ratio}; Constraints={safety_gate@ordinal}; Jacobian_sparsity=high; Missingness=MAR.
Families. Implicit‑BDF vs Explicit‑RK vs Symplectic.
Rules.
— Implicit‑BDF: Eligibility tolerates stiff?=unknown if Jacobian_sparsity=high (guarded precondition); MaturityCard=L3 (replicated & benchmarked). Outcome: Admit.
— Explicit‑RK: requires stiff?=false; with unknown ⇒ Degrade(sandbox) (probe).
— Symplectic: eligible only when Hamiltonian=true; here ⇒ Abstain.
Didactic anchor. This mirrors C.22’s typed‑signature discipline and CHR legality (no ordinal means; unit alignment for ratio).
Contemporary ecosystem examples of these families (post‑2015) are organised in DifferentialEquations.jl, which exposes multiple solver families under one call surface—precisely the pattern G.5 expects. (Journal of Open Research Software)
Show‑2 - Planning/scheduling (MIP task).
S2 excerpt. DataShape=MIP; NoiseModel=deterministic; Objective={↓cost@ratio, ↑service_level@ordinal}; Size≈10^5 vars; convex_relaxation=available.
Families. MILP (branch‑and‑bound), Constraint‑Programming, Heuristic meta‑search.
Rules.
— MILP: Eligibility requires convex_relaxation=available; MaturityCard=L3 in the home Context ⇒ Admit.
— Constraint‑Programming: MaturityCard=L2; Acceptance demands service_level≥B (ordinal predicate). With B met but baseline parity unknown ⇒ Degrade(scope‑narrow).
— Heuristic meta‑search: MaturityCard=L1 ⇒ Degrade(sandbox) or Abstain depending on RSCR parity policy.
Didactic anchor. Selector returns a Pareto set (no cross‑ordinal weighting), as required by G.5.
Contemporary “single call / many solvers” packaging that motivates MethodFamily rows is exemplified by JuMP (2017–2022), which cleanly separates model description from solver choice. (Miles Lubin)
— DifferentialEquations.jl illustrates family‑based solver packaging (multi‑method under one interface), 2017–2024 ecosystem. (Journal of Open Research Software) — JuMP illustrates model/solver separation and registry‑like selection (2021–2022 papers, site). (Miles Lubin) — Science of Science review (2018) supports the emphasis on replication/benchmarks in maturity assessment. (Science)
Show‑3 - QD archive (policy search).
S2 excerpt. PortfolioMode=Archive; CharacteristicSpaceRef(d=2); ArchiveConfig(CVT, res=1k cells, K=1, DistanceDefRef.edition=v2, InsertionPolicyRef=dyn‑elite); EmitterPolicyRef=v3; DominanceRegime=ParetoOnly.
Rules. Admit returns an archive; illumination reported; changes to DistanceDef/Emitter editioned in SCR; dominance remains ParetoOnly.
Show‑4 - Open‑ended GeneratorFamily (POET‑class).
S2 excerpt. GeneratorIntent{GeneratorFamilyRef=GF‑01, EnvironmentValidityRegion=EVR‑A, TransferRulesRef=TR‑A, CoverageMetric=…}; PortfolioMode=Archive.
Rules. Admit yields portfolios over {environment, method}; Degrade(scope‑narrow) if TransferRules=unknown; telemetry publishes coverage/regret and IlluminationSummary with edition/policy‑id on improvements.
Bias‑Annotation
Principle‑taxonomy lenses. Universality (trans‑discipline), Didactic primacy (Tell–Show–Show), Open‑ended evolution (refresh‑ready), Lexical firewall (no tool names in Core), Notation independence. Limits: Worked examples reference widely‑used ecosystems in Plain register only.
Conformance Checklist (normative)
Consequences
- Explainable admission. Every Admit/Degrade/Abstain is backed by anchored evidence and explicit unknown handling (selector reports are SCR‑linked).
- Run‑safe pluralism. Multiple families can co‑exist with policy‑governed exploration (E/E‑LOG) and maturity‑aware gating.
- Portable governance. Bridge hygiene and CL routing make cross‑Tradition reuse deliberate and costed (penalties to R only).
Rationale
The ladder and LOG shells align with FPF’s Assurance calculus: F (form) is governed by artefact kind, G (scope) by USM slices, and R (reliability) accumulates via WLNK then Φ(CL) penalties. Treating maturity as evidence‑typed rungs—rather than a “score”—avoids illegal arithmetic and lets design/run remain separate via DesignRunTag discipline (A.4) and explicit GateCrossings. This mirrors contemporary science‑of‑science insights: replication, benchmarking, and field health indicators are the currency of maturity, not anecdote. (Science)
Relations
Builds on: G.5 (selector consumes these rules), G.4 (Acceptance & EvidenceProfiles), C.22 (S2 typing), C.18 NQD‑CAL, C.19 E/E‑LOG, B.3 (Assurance tuple & WLNK).
Publishes to: UTS (MaturityCards, rule ids), SCR/RSCR (branch coverage; parity hooks).
Constrains: G.8 (LOG Bundling must cite MaturityCards), G.9 (parity harness draws baselines per rung), G.11 (refresh windows per rung & decay), G.5 (Open‑Ended Family mode for GeneratorFamily).
Outcome. The pattern introduces new content (LOG shells + maturity poset + degrade modes + publication Standard) and does not duplicate CG‑Spec legality rules, CHR guard‑macros, or CAL acceptance mechanics; it integrates them into admissibility logic for MethodFamilies.