B.5.2.1:4 - Solution — NQD-Generate and its C.18 records
Method name (Plain/Unified Tech). NQD‑Generate — a U.Method that, given (i) a HypothesisSpace and (ii) a CharacteristicSpace with a CoverageGrid, returns a finite candidate package: a front over the declared DominanceSet plus the retained archive/tie-break telemetry needed to keep diversity and novelty reviewable without making them default dominance dimensions.
Minimal signature.
-
Inputs (declared in MethodDescription):
HypothesisSpace,CharacteristicSpace,Seeds?,Budget (time/compute),EmitterPolicy(E/E-LOG policy id),QualityMeasures (Q components),NoveltyMetric?(required when generation, comparison or retention uses novelty),CoverageGrid/Granularity,CellCapacity K? (default=1),EpsilonDominance ε? (default=0),TieBreakPolicy? (S/I),DedupThreshold?,Policy(TimeWindow),DeterminismSeed? -
Outputs: CandidateSet = {h_i: (desc_i, Q_i, N_i?, D_i?:=ΔDiversity_P(h_i | Pool), S_i?, UseValue_i?), genealogy_i?, provenance_i (including policyId from E/E-LOG and, when a DHC coordinate is consumed, DHCMethodRef.edition and the other active fields of C.21’s
DHCReplayBasis)} whereQ_iis a vector andprovenance_icaptures generator settings and evaluation sources.
The coordinate entries cite the constituted results used. When illumination is reported, carry IlluminationSummary separately as a report about its declared retained set. Unused optional readings need not be produced. If Use-Value is present, identify its objective or acceptance criterion and scope. For a gain claim, also identify the baseline and comparison or counterfactual Method; for a predicted reading, identify the model edition and assumptions under C.17.
N, D, S, and I are archive, tie-break, telemetry, or policy-promoted signals by default; only the declared DominanceSet enters the front. Use-Value is decision-side/supporting unless the Context declares it inside the active Q tuple / DominanceSet; when it is only recorded as a side measure, keep it outside dominance.
Strategy (notation‑neutral).
- Seeding and constructors. Declare the explanatory question, hypothesis space and permitted variation operators in the MethodDescription. Initialize with admissible seeds: known explanations, random draws under that grammar or prior hypothesis-bearing epistemes with their grounds and limits. Name how an operator constructs a changed explanatory claim, for example by replacing a proposed mechanism or combining two mechanisms under a stated interaction assumption. When there are no seeds, use the declared finite enumeration or sampling rule.
- Construct and assess. Apply the selected variation operators to the seeds or retained candidates within the budget; record each resulting hypothesis and its parent/operator provenance. Then test candidate must-constraints for comparison and evaluate each required Q component. Keep an excluded candidate and its reason distinguishable from a candidate not yet generated. Maintain up to K retained entries per declared cell; obtain only the N, ΔDiversity_P and Surprise readings used under C.17, and any required retained-set illumination report. Deduplicate under the declared rule and threshold.
- Budget‑bounded loop. Iterate until the budget or declared coverage-convergence condition is reached, or the applicable C.19 policy ends active exploration; return the (ε‑)Pareto front over the declared
DominanceSet. When the Context consumes the ordinary default, that means the declaredQcomponents underDefaultId.DominanceRegime. KeepN,D=ΔDiversity_P,Surprise, andIlluminationSummaryas archive/tie-break/telemetry signals unless one Context policy explicitly promotes one of them into dominance and records the policy id.Use-Valueenters dominance only when the Context declares it inside the activeQtuple; otherwise it may appear as one decision-side/supporting side note. - Traceability. Keep the generator settings, grids/metrics versions, seeds, policy,
TimeWindow, filled cells and each exclusion’s reason recoverable in the generation and comparison account. Use C.18 for the archive and front returns. State how the final set was produced, includingε,Kand deduplication. Keep any decision rationale required by the use in that account; reuse an adequate account. - Algorithmic freedom (informative). Implementations MAY use MAP‑Elites/illumination, novelty search with local competition, Bayesian/surrogate‑assisted search, or deterministic enumerations; ε‑dominance or knee‑point thinning MAY be used after recording the full front in provenance.