fpf.tools
Library / First Principles Framework (FPF) - Core Conceptual Specification
Jump to passage
In this reading
Source · lines 56713–56741

Link to current text

Published source confirmed at last check

Source changed 2026-10-03 05:29:54 UTC · snapshot created 2026-10-03 05:30:57 UTC · last check 2026-10-03 06:25:20 UTC

C.17:8 - Evidence, uncertainty, and resistance to gaming

Every quantitative result names its evidence, calibration, uncertainty, and validity window. Keep aleatory and epistemic uncertainty separate and state any rule that combines them. A claim imported from another source or scale states the preserved meaning, lost meaning, direction, receiving use, and evidence limit; use F.9 only when an actual Bridge between reference-scheme senses is needed.

Apply these anti-Goodhart guards:

  • pair Novelty with Use-Value or ConstraintFit;
  • freeze the comparison corpus, encoder or model, and Scale edition for the result; a load-bearing change creates a new result basis;
  • keep Illumination as telemetry unless an explicit policy promotes it for a named use;
  • check delayed practical consequences, such as retention, maintenance burden, or cost-to-serve, before celebrating a proxy win;
  • connect DiversityOfSearch to the declared experimentation allowance and report overspend;
  • retain primitive coordinates and sensitivity when a composite index is used.

When a corpus, inclusion rule, model, encoder, Scale, criterion, window, or cross-source premise changes, leave a compact change account: the previous and new editions, what changed in the basis, which coordinates and aggregate conclusions reopen, whether eligibility or frontier membership changed, and which next observation can settle the difference. Latest is not a reproducible selector. Use B.4 and G.11 for the refresh; the change account explains the comparison and does not replace the new result.

Goodhart’s law is a useful historical warning, not evidence for any result. The safeguards above do the operational work.

Evidence can support a result without establishing assurance. Ordinary use stops with proportionate evidence. Enter B.3 only when an actual named assurance claim is current.

When that receiving use requires independent assessment or segregation of duties, name the Systems that performed any bearer-producing, coordinate-measurement, or overall-assessment Work and the assignments relevant to the independence claim. State the relevant conflict or independence evidence. Do not impose this assurance arrangement on an ordinary qualitative discussion that makes no independence claim.

C.17:8.1 - Cross-scale and cross-source limits

When a reading moves across scales, state which Characteristics are preserved, aggregated, projected, or lost and cite the actual aggregation, projection, transition, temporal cross-scale, or Bridge relation on which the reading depends. Keep a plain distortion note next to the projection. If the receiving use needs the named A.0 qualifier structure, use A.0:QF.2a and its optional OutcomeMapRef, TransitionRelationRef, or BridgeDistortionNote; otherwise the plain loss statement is enough.

Different projections of the same retained set or frontier may preserve different information. A useful projection is not automatically information-preserving, and one atlas-like view does not cancel the original Front, Archive, or result.

When claims come from different source schemes or corpora, do not compare by matching labels. State the actual mapping, Bridge when required, direction, loss, calibration, and receiving use. A target-use pilot may improve evidence for that use; it does not make the source and target bases identical.