WM-07: Give parameter learning a specific object

First authored design, 2026-09-15. This elaborates exactly WM-07. It is a reviewable design, not implementation, independent acceptance or machine completion. The existing build loop continues under claude-2; these documents do not dispatch new jobs or override recorded decisions.

All WM cascades · Checklist item · Machine-readable cascade

The problem and existing evidence

WM-07 concerns uncertainty about the model, distinct from uncertainty about the work’s current state. A finite registered-hypothesis calculation is existing bounded evidence. It does not establish continuous Dirichlet inference, learned production parameters or a consumer that uses information gain.

Governing checklist snapshot

WM-07 — Parameter prior/posterior and information gain have machine meaning. PARTIAL: actual model inputs and observations, refresh of registrations after A/model changes, actual epistemic consumer and evidence. Continuous Dirichlet claims cannot be inferred from a finite-hypothesis witness. [W11,W14,L] Built evidence: Row 11 finite registered-hypothesis module and Bayes/reference checks exist.

The preceding evidence is the checklist’s dated account. It has not been re-audited as a current deployment census for this design.

Pattern applications

WM-07-C1 — Name the unknown parameter

Pattern: wm-prediction/name-the-uncertainty-to-learn.

Describe what theta means and what observation could distinguish its values. Separate an incomplete handoff from an unreliable procedure for judging handoffs. Output the parameter family and scope, with unresolved choices visible. A generic confidence score is not a parameter model.

WM-07-C2 — Declare the learning experiment

Pattern: aif/declare-the-conditioning.

Bind current data cutoff, action/policy, model revision, parameter prior and prospective observation channel. Identify which data supplied the prior and which later observation updates it. Output an inference context that prevents counting the same observation twice or using future data to justify an earlier decision.

WM-07-C3 — Construct compatible inference

Pattern: wm-prediction/outcome-meaning-across-models.

Supply the actual prior, observation likelihood and conditional posterior on the same parameter domain. For a finite hypothesis family, retain the hypotheses and every likelihood source. If a continuous family is needed, implement its own inference and correspondence rather than cite the finite witness. Impossible evidence is an explicit domain case, not automatic uniform posterior.

WM-07-C4 — Derive the information quantity

Pattern: wm-prediction/probability-laws-with-runtime-meaning.

Calculate expected parameter information from those distributions under the accepted definition, with supports, zero cases and numeric semantics explicit. Distinguish expected information before an observation from realized posterior change after it. Do not substitute a concentration count or state entropy change. Retain an independently reproducible calculation.

WM-07-C5 — Revise and register the model

Pattern: wm-connected-learning/proposal-earns-later-use.

When A, B or parameter family changes, retain successor registration and invalidate incompatible inference artifacts. Preserve original observation identities and old predictions. A new registry entry is not evidence that learning improved prediction; admission and subsequent use remain separate.

WM-07-C6 — Connect the epistemic consumer

Pattern: wm-connected-learning/observation-to-next-choice.

Name the actual consumer that uses this parameter information and its authorized relationship to the objective. WM-10’s state mutual information is a different quantity; do not add this term again without a governing objective. Show actual calls and later observations, not only a standalone estimator.

Composition and revision

These are design input/constraint relationships; they do not assert automatic runtime firing. Each contribution constrains the next actual artifact.

Build interfaces

Parameter declaration and prior producer → inference context. Admitted observation likelihood plus context → counterfactual/observed posterior. Information calculator → quantity with definition and numeric receipt. Registration manager → compatible model identity. Epistemic adapter → actual consumer binding. All five interfaces retain data cutoff and observation identity.

Each envelope must retain its schema, content identity, source/model revisions, task/occurrence and information boundary. Typed missing inputs remain missing; do not replace them with success-shaped defaults. Implementers may reuse existing compatible components, but must identify the real producer and consumer before crediting integration.

Worked situation

A failed handoff may reflect missing content or a reviewer who routinely overlooks it. One failure does not identify the latter parameter. The model must explain which independent observation discriminates the explanations, and how the resulting posterior affects a later decision. Relabelling low state confidence as learned reviewer reliability fails this design.

This situation is illustrative, not a measured production result or an adopted probability table.

Build order and acceptance

1. Resolve the first two applications against the real work and exact authority. Preserve the unresolved decisions below; bell design questions before choosing dependent behavior.
2. Implement the middle producer/checker interfaces with explicit source and numeric contracts. Retain a meaningful failing example before repair and independent review after it.
3. Connect the final consumer and collect the particular evidence described above. Unit tests, source acceptance and ordinary use receive separate credit.

Distinguishing controls: Use informative and uninformative observations, an impossible-evidence case, a changed model revision, duplicated observation input and a finite/continuous domain mismatch. Independently recompute a small admitted example. Verify the consumer reads the result and does not add the same epistemic quantity twice under different names.

Do not rerun the entire machine merely to validate a local edit. Select checks for the changed invariant; retain commits, commands and results. A future qualifying run still requires every required node and connection.

Dependencies and open decisions

WM-01 · WM-02 · WM-03 · WM-04 · WM-05 · WM-10 · WM-11

Open: the production parameter family, its prior authority and inference algorithm, evidence eligibility, and the exact consumer objective. No new weight or Dirichlet interpretation is adopted. Conditional finite mathematics remains useful without being promoted to a continuous production result.

The dependency links express interface constraints, not a topological build order. A producer and consumer can jointly constrain an interface without requiring that all other implementations be finished first. Use a bounded packet once its relevant semantics are settled.

Sources and review boundary

The companion JSON pins the governing checklist, September rulings, current cascade-policy semantics and every applied pattern. Where the semantics source distinguishes a ruling, extrapolation or open proposal, that distinction controls. No generic finite witness is promoted to a machine instance. Existing patterns are reused within the application stated above; their other examples, thresholds and historical implementation claims are not automatically adopted.

Existing AIF reference snapshots

Retained source text, not new patterns or adopted historical claims.

@flexiarg aif/declare-the-conditioning
@title Declare the Conditioning (Score the Fit, Not the Form)
@keywords conditioning, situation, context, fit, EFE, curried, referent
@audience aif implementers, pattern authors, futon developers
@tone technical
@style pattern
@references [aif/currency-before-merge aif/valuation-reads-the-paperwork]

@how aif/structured-observation-vector
@see-also aif/term-to-channel-traceability
@why no matching problem dossier node (@holds-at absent; not among the twenty dossier nodes R1-R17, R20, TRACE). Negative claim per the committed check futon2 holes labs library-loop runs l3_no_dossier_check.py, receipt runs L3-no-dossier-check.edn (enumerates every aif pattern's @holds-at against the dossier node set extracted from PROBLEMS-*.md headings). Rationale not invented (L3 rationale-backfill pilot, zai-1, 2026-09-05)
@how own mechanism: Every valuation surface must declare what slice of the situation it consumed; G(candidate) is a lambda — only G(candidate | situation) is a value. (source: this pattern's ! conclusion line; L3 rationale-backfill pilot, zai-1, 2026-09-05)
! conclusion: Every valuation surface must declare what slice of the situation it consumed; G(candidate) is a lambda — only G(candidate | situation) is a value.

  + context: A stack grows multiple valuation surfaces (blend terms, wholeness, ΔG legs, field diagrams), each labelled with the same vocabulary ("EFE", "value", "score").

  + IF:
    You want valuations to be comparable, auditable, and honestly labelled.

  + HOWEVER:
    Different surfaces silently consume different slices of the world — one eats the live observation vector, another a 160-char prose scrap, a third nothing at all — so identical labels name incomparable objects, and cross-surface comparisons are apples-to-oranges without anyone having lied.

  + THEN:
    Give every valuation surface a three-axis declaration: SENSE (which formal object it claims), SELECTION-SEMANTICS (extremum / amortized sample / bound — over what menu), and CONDITIONING (what situation-slice it consumed). Render the declaration where the number is rendered. In pattern terms: Context → (Problem → Solution) — application is modus ponens; the world supplies the antecedent; score the fit, never the form alone.

  + BECAUSE:
    Canonical expected free energy is G(π; Q(s), C) — conditioned on beliefs and preferences by definition; a policy has no value in vacuo, and Alexander says the same of form and context. Surfaces that skip the declaration re-invent the mislabeling one level up.

    + evidence: M-evaluate-policies §9.6–9.7 + exhibit "conditions on" column (futon2, 2026-07-03): the atomic-action lane conditioned on the live observation vector while the cascade lane ate a title+status scrap and two native components consumed no situation at all — one arena, incomparable conditioning. E5 measured the conditioning channel: situation-enriched ψ rebuilt 11/13 cascades. The fold contract already typed circumstance as an argument; the plumbing had starved it.

  + NEXT-STEPS:
    - Add the sense × selection-semantics × conditioning triple to each EFE-labelled surface's declaration (extends the efe-referent-declaration shape).
    - Route all lanes toward ONE situation source (the structured observation vector) where feasible.