fpf.tools
Library / First Principles Framework (FPF) - Core Conceptual Specification
Jump to passage
In this reading
Source · lines 115415–115442

Link to current text

Published source confirmed at last check

Source changed 2026-10-03 11:52:20 UTC · snapshot created 2026-10-03 11:53:41 UTC · last check 2026-10-03 12:15:10 UTC

G.3:5 - Archetypal Grounding

AG‑1 — ML fairness auditing (post‑2015 selective and set‑valued practice). System: a CG‑Frame for evaluating deployed classifiers across cohorts with explicit abstention/defer behavior. CHR authoring: publish DemographicParityGap and EqualizedOddsGap as Characteristics with:

  • explicit ReferencePlane, with deployment population and sampling regime recorded separately as scope and protocol conditions,
  • ObservableOf (audit protocol + uncertainty model + window),
  • interval scale (bounded; zero semantics explicit),
  • missingness semantics (cohort sparsity and label noise are typed),
  • legality surfaces and guard surfaces that forbid illicit cohort mixing and require explicit proof hooks for aggregation across cohorts.

Downstream: CAL acceptance binds thresholds and failure behavior; selector remains set‑returning under partial orders and may treat “defer/abstain” as a first‑class outcome (tri‑state semantics pinned through G.Core).

AG‑2 — Clinical diagnostics (post‑2015 evidence‑aware evaluation). System: a CG‑Frame for comparing diagnostic pipelines under evolving datasets and protocols. CHR authoring: publish Sensitivity and Specificity as ratio‑scale, dimensionless Characteristics on [0,1], with:

  • explicit ObservableOf (trial protocol, inclusion criteria, uncertainty model),
  • freshness/decay expectations (protocol drift is modelled as decay),
  • legality surfaces that forbid averaging incompatible ordinal labels (e.g., severity grades) and require explicit unit/exposure constraints for any derived rate.

Downstream: CAL acceptance governs thresholds and guard‑bands; evidence wiring is cited via Path/PathSlice to make refresh triggers actionable.

AG‑3 — Quality‑Diversity / Illumination (post‑2015 MAP‑Elites/CMA‑ME lineage). System: a CG‑Frame where selection returns archives/fronts rather than a single winner. CHR authoring: declare which Characteristics play Q/D/QD‑score roles and pin the metric definitions (descriptor map, distance definition, method editions) so archives are reproducible across runs and refresh can be triggered on edition changes. CHR does not scalarize partial orders; set‑return semantics are pinned through G.Core.