Reach-assessment

Type: kb/types/definition.md · Tags: foundations, computational-model, self-improving-systems

Reach-assessment is the capability of an evaluation process to judge whether a candidate commitment's claimed explanatory-reach is genuine rather than adaptive fit presented as explanation. It evaluates the content of the claim: whether the proposed mechanism or invariant supports the generality claimed. Checking that a new case matches the commitment's stated condition checks only the condition, not whether that condition names the real boundary.

For example, retaining “route prompts under 500 tokens to a cheaper model” after several successes does not establish that length explains task difficulty. Reach-assessment must distinguish that mechanism from a correlation in the fitting cases, such as by varying complexity while holding length fixed.

Scope

The assessment route must match the representational form carrying the commitment:

Mixed artifacts may compose these routes, but an assessment of one form does not automatically warrant the others. Proof reaches only across its axiomatized domain; observed predictive fit reaches only across the tested shift class.

Reflectivity supplies addressability, not this judgment. A reflective process may represent, match, reject, or rescope a commitment without being able to tell whether its stated scope is honest.

Exclusions

  • Reach-assessment is not boundary matching, metadata checking, confidence, or the structural capacity to rewrite a commitment.
  • Empirical success on the fitting cases does not establish the mechanism or its reach beyond them.
  • Causal vocabulary, a proof engine, or a learned world model does not by itself supply reach-assessment. The evaluation process must exercise the relevant causal, proof, intervention, or shift test.
  • Observational causal discovery is not assumption-free; its warrant remains conditional on its discovery assumptions.

Provenance and open boundary

The causal, proof, and predictive-world-model routes are established external machinery. Commonplace's contribution is to group them by the role they play when an improvement pathway evaluates a retained commitment. The natural-language route remains unexplained: LLM-mediated evaluators appear able to judge prose generalizations with some reliability, but this definition names that missing capability rather than supplying a theory of it.

The hyphenated compound reach-assessment carries this technical sense, and explanatory-reach names the property it judges. Bare reach remains ordinary English. The oracle-side limit is oracle domain and the generator-side limit is search range.

Misuse Cases

  • Assuming that a reflective system has reach-assessment because it can revise a stored scope statement.
  • Treating a passing test suite or accurate predictor as evidence of reach without testing the claimed mechanism, intervention, or shift boundary.
  • Citing this definition to claim that reach-assessment is unique to LLMs; causal inference and proof provide older formal routes.

Relevant Notes:

Derived into: