Case packet

Neutral case identifier: case-a628538d6d8332

The possible directed relationship from Artifact A to Artifact B is under review.

Artifact A

Reach-assessment

Reach-assessment is the capability of a system's evaluation process to judge whether a candidate commitment's claimed [explanatory-reach] — that the pattern it names holds beyond the evidence that produced it — is genuine, rather than adaptive fit dressed up as an explanation. This is a judgment about the content of a claim, not a check on its form. For example: a coding agent that retains "route prompts under 500 tokens to a cheaper model" after early tasks succeed has an adaptive-fit-shaped claim until something checks whether token count is actually why those tasks were easy, or just what they happened to share.

[Reflectivity] does not supply this for free. A reflective system can represent a commitment's stated applicability condition, read it, and rewrite it — [rejecting or rescoping a prior commitment] is a structural operation reflection makes available. None of that requires judging whether the commitment's stated boundary is honest. A process can mechanically compare a new case against a stated condition and act on the match — without ever assessing whether the underlying explanation captures an invariant, or is dressed-up curve-fitting that happened to generalize once. A boundary-matching process checks only whether the new prompt is under 500 tokens and swaps in the cheap model, and misapplies the commitment the first time a short prompt encodes a hard multi-step task — token count was never the mechanism, only something that happened to correlate with it in the cases that produced the commitment.

Scope

  • Reach-assessment operates on the content of a candidate commitment — its causal or explanatory structure — not on metadata about it. Checking that a new case satisfies a stated condition clause is boundary-matching, [the discriminator memory formation already uses]; reach-assessment is what would be needed to judge whether that stated condition is itself trustworthy.
  • What it would need to do is apply something like [the four-part negative test] — can a load-bearing premise be varied and the conclusion predictably change; does the claim reach into an untested domain; can it be criticized in a specific way; does it survive contact with observed fit — to a system's own candidate commitments inside its improvement pathway, not to a human author's KB note. Applied to the routing commitment: does varying task complexity while holding length fixed change the outcome predictably; does the claim reach to prompt types not yet tried; can a specific failure case be named in advance; does it hold up against the tasks that actually produced it.
  • Which route is available depends on the candidate commitment's [representational form]. In this KB's artifact vocabulary, reach-assessment can be mediated by all three retained forms: natural-language artifacts need semantic judgment such as the [first-principles explanatory-reach test]; symbolic artifacts can use explicit causal or proof obligations; distributed-parametric artifacts can use learned predictive world models. Mixed artifacts can compose those routes, but the route must match the form that actually carries the commitment.
  • The symbolic route has two forms. Causal theories: a system has reach-assessment for a proposed mechanism when its acceptance process tests that mechanism against interventions, counterfactuals, or cross-environment invariance, rather than merely fitting observed correlations. Proof, which can wrap the causal route when causal inference itself is axiomatized: where a commitment can be expressed as a typed claim, formally specified invariant, or target theorem, proof search can establish that it holds across its stated domain — the [Gödel machine]'s proof-gated self-rewrite acceptance is the worked case. See [Formal symbolic systems assess explanatory-reach only through causal and proof obligations] for the developed argument, mechanics, and source grounding.
  • A learned predictive world model is the canonical distributed-parametric route: reach-assessment means testing whether a learned predictor's action-conditioned predictions remain valid across the interventions or environment shifts a commitment claims to cover. See [World models assess explanatory-reach through action-conditioned prediction] for the developed argument and LeCun/JEPA grounding.
  • Where a commitment remains in natural-language form — an explanation not yet reduced to a checkable formal claim — no comparable formal apparatus exists yet. There, reach-assessment is observed only in LLM-mediated evaluation: an agent asked to judge whether a proposed lesson generalizes correctly appears able to do so with some reliability, and Commonplace has no theory for why. See [Provenance and departures].

Exclusions

  • Reach-assessment is not reflectivity. [Reflective system]'s five requirements — boundary, represented aspects, self-representation, processes, causal connection — are purely structural. A reflective system built entirely on exact- or fuzzy-string matching against a stated condition clause has every one of those five properties and no reach-assessment at all.
  • Reach-assessment is not causal vocabulary. A retained sentence that uses words like "cause," "mechanism," or "intervention" has no reach-assessment unless the system's evaluator can test, identify, or otherwise warrant the causal structure it is relying on.
  • Reach-assessment is not empirical testing alone. A test suite that passes on every case it checks establishes that a candidate holds on those cases, not that it holds beyond them — the explanatory-reach question stays unanswered by passing tests. Observational causal discovery is the important exception only because the observations are interpreted through causal assumptions and discovery criteria, not because observation by itself contains causality. A proof oracle is a different matter: proving a claim within an axiomatized model does establish that it holds across the model's stated domain, a genuine, non-semantic route to reach-assessment (see Scope) — but only as far as the axiomatization reaches, [which is what warranted autonomy is bounded by].
  • Reach-assessment is not a learned world model by itself. A latent predictor has reach-assessment value only when its action-conditioned predictions are tested against the intervention or shift class at issue; otherwise it may be another adaptive fit artifact.
  • Reach-assessment is not confidence. A system reporting high confidence in a rescoped commitment is not evidence the rescoping was reach-assessed; confidence and correctness come apart exactly where reach-assessment would matter most — a fluent, wrong generalization is the failure mode this capability exists to catch.

Misuse Cases

  • Assuming any reflective pathway automatically has reach-assessment because it can represent and rewrite a commitment's scope — the operation being available says nothing about whether it is exercised well.
  • Treating a match between a stated condition clause and a new case as evidence that the underlying commitment has genuine explanatory-reach — that is boundary-matching, which presupposes the stated boundary is honest rather than testing it.
  • Treating a system as having reach-assessment merely because it stores causal-shaped theories. The assessment enters through the causal learning or evaluation process: selecting mechanisms that survive the relevant intervention, counterfactual, or environment-shift tests.
  • Treating a learned world model as having reach-assessment merely because it predicts next states. The assessment enters when action-conditioned predictions are evaluated over the intended interventions or shifts.
  • Treating observational causal discovery as assumption-free induction. It can infer causal theories from observations under discovery assumptions; it does not make explanatory-reach available from raw correlations alone.
  • Treating the Gödel machine as evidence that proof search can assess the explanatory-reach of arbitrary natural-language explanations, or as evidence for causal discovery without causal axioms in the system. It supports the formal case: commitments expressible as target theorems under the machine's axioms.
  • Citing this note to claim reach-assessment is unique to LLMs. It is not: causal inference, do-calculus, and proof search are traditional, formal routes to reach-assessment for symbolic claims, and predate LLMs by decades. What is genuinely unexplained is narrower — semantic judgment of a claim still in natural-language form, where LLM-mediated evaluation is the only route currently observed.

Provenance and departures

This is Commonplace's own concept, without a single inherited definition to depart from, but its routes have different provenance.

The causal route — structural causal models, intervention semantics, do-calculus, invariant prediction, and causal discovery from observational or interventional data — is established science, not Commonplace's own. Commonplace's terminology only names the role that machinery plays when it evaluates whether a retained mechanism really reaches beyond its fitting cases.

The proof route is also external. The [Gödel machine] is the worked self-modification case already in this cluster, grounded in Schmidhuber's target-theorem construction rather than in Commonplace's vocabulary.

The learned-world-model route comes from predictive representation-learning work such as LeCun's JEPA line, where a latent predictor supports planning by anticipating how the world changes under candidate actions.

A vocabulary registration, [following the write-time collision rule]: the load-bearing senses here ride on rare compounds — reach-assessment for this capability, explanatory-reach for the judged property (already registered in the review-explanatory-reach criterion). Bare reach carries no technical sense anywhere in the KB — the compound carries the sense corpus-wide, not only in this cluster; the oracle-side limit is named oracle domain ([warranted autonomy]) and the generator-side limit search range ([the proposal-selection loop]).

Commonplace's own gap is the natural-language route. The four-part negative test it builds on is adapted from Deutsch by [first-principles-reasoning-selects-for-explanatory-reach-over.md], and the observation that stating a boundary is itself a judgment — not a mechanical derivation — is already made by [abstract-an-experience-only-when-you-can-state-the-boundary.md]. That leaves who or what can reliably make that judgment, for a claim still in natural-language form, as an open problem: this note names the missing capability, it does not close the open problem, and why LLM-mediated evaluators appear capable of it is not theorized here.

Relevant Notes:

Derived into:

  • [Formal symbolic systems assess explanatory-reach only through causal and proof obligations] — the causal and proof routes worked out from this definition's representational-form split
  • [World models assess explanatory-reach through action-conditioned prediction] — the distributed-parametric route worked out from this definition's representational-form split

Artifact B

First-principles reasoning selects for explanatory-reach over adaptive fit

Commonplace's first-principles methodology is valuable because it selects for explanations with explanatory-reach: claims that keep working outside the case that produced them because they capture why a pattern works, not just that it worked. This note adapts David Deutsch's adaptive-vs-explanatory distinction (his term is bare "reach") for KB design, treating adaptive fit and explanatory-reach as a polarity rather than a hard binary.

Adaptive information helps a system cope with the world. A genome, a neural network's weights, or a local rule of thumb can encode something useful without explaining why it works or where it stops working.

Explanatory knowledge gives a criticizable account of why the pattern works. It can be deliberately varied: change a load-bearing premise, and the explanation should constrain what changes in the conclusion. That variation is what gives an explanation its explanatory-reach.

Partial cases sit between the poles. A rule can transfer across a narrow family of cases because it captures shared structure, while still falling short of a full generative model. The point is not to demote every local observation; local fit is the evidence a later explanation must organize, predict, and improve.

Why this matters for the KB

When a note derives a design pattern from inherited constraints — finite context, no scoping mechanism, text-in/text-out interpretation — the derivation can be explanatory: it says why the pattern works and predicts where it will fail. Change the constraint, and the conclusion should change with it.

That "can be" matters. First-principles stories become post-hoc rationalization when broad premises can justify several rival practices equally well. A useful derivation should rule out at least one plausible alternative, name the constraint that does the ruling out, or predict a failure that later use can confirm.

The [computational-model] area is an explanatory-reach bet under audit. Programming-language concepts such as scoping, partial evaluation, and scheduling were developed for compilers, but they reach into KB design when the shared invariant is explicit: bounded processors compose text under constraints, and unscoped composition lets distant bindings interfere. [LLM context is composed without scoping] therefore works only if the dynamic-scoping comparison predicts real context failures and useful remedies, not merely because the analogy is elegant.

The negative test

This adapted distinction provides a quality check orthogonal to the KB's type system. A well-formed note can pass every structural check while still recording a pattern without explaining the mechanism. The test:

  1. Can you vary a load-bearing premise? If changing one premise lets you predict a constrained change in the conclusion, the note is exposing causal structure. If any premise can move while the conclusion stays rhetorically intact, the derivation may be decorative.
  2. Does it reach? Would the insight apply in a domain you have not considered, and can you say which invariant carries it there? If yes, the mechanism is deeper than the original case. If no, the note may be context-fitted.
  3. Can it be criticized? Is there a specific way the explanation could be wrong, not just incomplete? The [falsifier blocks] practice operationalizes this.
  4. Does observed fit discipline the explanation? Local success is not second-class residue; it is evidence. If the explanation cannot account for where the pattern actually works, fails, or costs too much to maintain, it has verbal explanatory-reach without operational grip.

The first three tests map to the three depths in [discovery]: shared feature (adaptive), shared structure (partially explanatory), generative model (fully explanatory, with explanatory-reach). The fourth is not a depth — it holds an explanation at any depth accountable to the observed fit it must organize.

Scope

Explanatory-reach is the quality goal for theoretical notes, not the only kind of KB value. Descriptions need economy, instructions need precision, and logs may preserve local observations before the mechanism is understood. The explanatory-reach filter says when an observation is ready to become a transferable claim; it does not replace the capture layer that supplies the observations.

Open Questions

  • Where in the KB are notes that are well-formed but merely adaptive? Those are candidates for deepening.
  • Which first-principles derivations currently rule out a rival practice, and which only explain an already-preferred practice after the fact?
  • Should this note keep a direct Deutsch source, or is the adapted distinction enough if the KB-specific test stands on its own?

Relevant Notes:

Derived into:

  • [review-explanatory-reach] — the four-part negative test restated as a recurring review procedure
  • [COLLECTION.md] — the four-part negative test condensed into the "Tests for explanatory-reach" authoring block

Under-review context phrase

the four-part negative test this note asks a system's own evaluator to apply, not only a human author