WM-05: bootstrap the predictor through a pattern cascade

First authored design, 2026-09-15. Target: exactly WM-05, constructing Q(s|π) and Q(o|π) from machine inputs over the full interpreted policy, and delivering those predictions to their actual consumer. Joe has authorized creating this cascade from our seeded patterns. This is the design artifact that results; it is not an executed cascade or a claim that WM-05 is complete.

The problem this arrangement addresses

WM-05 has generic construction modules and scoped reference evidence. The unresolved problem is whether the actual belief, interpreted dynamics, observed likelihood and candidate policy compose into a prediction about the same work. A missing target row is one possible break. Repairing that lookup alone could leave the policy, observation domain and consumer disconnected.

The cascade must therefore prevent a locally successful repair from becoming an overall false prediction claim. Its central pattern is Carry the Outcome Meaning Through Prediction, one of WM-05's two original seeds. Declare the Conditioning, the other seed, gives that central pattern a particular situation to work on. The supporting seeds answer questions that those two cannot resolve alone.

Here there are two different objects. This authoring cascade organizes how we design and justify WM-05. π, in the predictor we are designing, is a full interpreted candidate policy supplied by the machine. Authoring this document does not make it that machine policy, nor establish a probability distribution over its own success. We are bootstrapping an explicit design before we have the validated predictor needed to score such designs numerically.

The eight pattern applications

C1 — Declare the Conditioning

Seed: aif/declare-the-conditioning.

Apply it to the actual WM-05 boundary: identify the target, current belief source, model revision, candidate source and intended prediction consumer. The current evidence includes source-level constructors and historical traces; it does not include a qualifying connected prediction run. Keep these different scopes visible. C1 produces a named situation for the rest of the design, not a value for G.

Without it, later nodes could choose convenient inputs from unrelated examples and still produce plausible numbers.

C2 — Explain the Difference Before Choosing the Carrier

Seed: wm-prediction/meaning-before-carrier.

Within C1's situation, identify the conditions of the work that different candidate policies could change. Explain what each proposed state means and which observations could contradict it. The proposed seven-status mission/ticket mapping remains unaccepted; a registry's status is not automatically the hidden state the predictor needs.

C2 contributes a state interpretation with explicit unresolved distinctions. It prevents C1's target identifier from being mistaken for an already justified model of that target.

C3 — Preserve the Whole Policy Through Comparison and Choice

Seed: wm-pattern-design/same-policy-through-comparison.

Identify the complete interpreted candidate at the prediction boundary: pattern identities, arrangement, relevant guards and model bindings. Distinguish two candidates with the same work target. Only the input-identity part of this seed is applied here; this cascade does not implement WM-11's posterior or selection law.

C3 contributes a policy description that the transition model can actually condition on. Without it, C2's careful state interpretation could still be propagated under a generic action unrelated to the chosen design.

C4 — Let the Observation Contradict the Belief

Seed: wm-prediction/observation-can-disagree.

Ask what independent observation would distinguish C2's possible states and which observation domain the prediction must expose. Identify the observation procedure, origin and limitations before calling an available record a sample. A report of insufficient evidence remains insufficient; it cannot fill an A row. C4 identifies the required observation authority and missing likelihood evidence, rather than producing invented likelihood values.

This constrains C2 in return: a state distinction with no justified relation to observations may need revision. That revision is a design feedback relation, not evidence that a runtime transition occurred.

C5 — Match Interpretation as Well as Identity

Seed: wm-connected-learning/same-interpretation-predicted-and-enacted.

Use C2's state meaning and C3's candidate description to explain what the interpreted actions do. Establish the applicable controlled transition law, including execution history where it matters. The known once-only versus repeatedly-enabled discrepancy remains an explicit question; this design does not resolve it by picking the implementation that happens to be easiest.

C5 contributes the proposed B meaning and the correspondence obligation. A documented THEN interpretation can supply declared dynamics, but it still needs independent validation of applicability. It is not a measurement that the effect happened.

C6 — Carry the Outcome Meaning Through Prediction

Seed: wm-prediction/outcome-meaning-across-models — the central composition.

Bring together C2's belief/state meaning, C3's policy identity, C5's transitions and C4's observation likelihood. Now the constructor can be described without silently changing subjects between components. The output must retain full policy/model identity and the predicted state and observation distribution at each required horizon step.

Its relation to C is compatibility of outcome meaning. C is needed by the downstream risk consumer, not by the mathematical marginalization that produces Q. This cascade does not invent preference members or expand WM-05 into completing WM-06/13.

C7 — Preserve Probability Laws Across Representations

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

Challenge C6's construction at the same supports and revisions. Require valid input kernels, normalized nonnegative distributions, explicit treatment of impossible or unsupported cases, and the accepted numerical correspondence criterion. A floating-point result is not made exact by printing fewer digits. A Lean law about a supplied table is not evidence that the machine supplied that table.

C7 constrains the construction; it is not a final cosmetic check. Its findings return to the particular producer or composition that violates the common law.

C8 — Follow the Finding into the Next Decision

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

Follow the actual constructed Q into its declared production consumer. Show that the consumer receives the same policy, outcome support and model identity, rather than reading a legacy proxy while the new result is merely logged. For WM-05, this applies the producer-to-consumer discipline; it does not claim the whole learning loop or canonical G selection has now been built.

A missing consumer remains a WM-05 integration gap. A mismatch feeds back to C1's boundary account and C6's output contract.

Why these are composed, rather than simply listed

The principal structure is:

C1 → {C2, C3}; C2 → {C4, C5}; C3 → C5; {C2, C3, C4, C5} → C6 → C7 → C8.

These arrows mean justified input or constraint relationships in the design. C4 also challenges C2; C7 can require revision of C6; C8 can expose a mistaken C1 boundary. The separate cascade.json records the type and reason for each relationship and pins every source pattern. Feedback edges are not interpreted as automatic repeated execution of patterns.

The interaction matters. C2 alone could yield a sophisticated state model with no independently grounded likelihood. C4 alone could gather valid observations about the wrong state distinction. C3 alone could retain perfect candidate identifiers while C5 models different execution semantics. C6 forces those contributions to concern one prediction; C7 and C8 stop a correct-looking marginal from receiving credit without mathematical and actual-consumer correspondence.

What construction C6 is meant to explain

For a declared finite Markov model, with policy/history information included wherever required, the intended composition has the form:

q_0(s) = the admitted current machine belief q_(t+1)(s') = Σ_s B_pi,t(s' | s) q_t(s) q_(t+1)(o) = Σ_s' A_(t+1)(o | s') q_(t+1)(s')

These equations describe conditional construction, not a new choice of State, A or B. B_pi,t must represent the full relevant interpreted policy at that step. If a state alone does not contain enough history, augment or otherwise model the necessary conditioning explicitly; do not suppress it in code because this shorthand omits it. An adaptive policy also requires the appropriate contingent choices to be represented. Marginals do not automatically supply a joint trajectory distribution needed by another consumer.

The finite-support assumptions and kernel laws explain why normalized inputs can yield normalized predictions. They do not justify the inputs empirically. In particular, no uniform belief, identity B, point-mass Q or legacy Gaussian proxy fills a missing producer in this design.

A concrete branch this cascade rejects

The earlier P1 proposal suggested adding mission/ticket belief rows. That proposal has not established the model interpretation or the independent observation likelihood. If we take its seven-status carrier merely because the reader accepts it, C2 remains unresolved. If we then turn registry lifecycle changes into observed ground truth, C4 fails. If the predictor records a new Q while the consumer continues to read its old action score, C8 fails.

This is why the proposed cascade is useful before any implementation: it identifies the conceptual work hidden by that seemingly small repair. It does not imply that adding target rows could never be part of the eventual solution.

Present boundary and next authoring move

C1 is applied here to delimit WM-05 and its source-versus-production gap. C2–C5 name the unresolved model interpretations; C6 specifies the conditional construction they must support; C7/C8 define the correspondence to be established. No populated machine model, constructed live Q, numerical G comparison or execution is claimed.

The next authoring move is C2 together with C4: work through what the state and observation distinctions mean for the actual selected work, with C3 keeping the candidate policy explicit. Their dependence is why choosing a carrier first and seeking observations later is premature. WM-02/03/04/06 are dependency interfaces in this WM-05 design, not newly commissioned workstreams. Existing dispatch restrictions remain in force.

Inspect the composition

Download the source-pinned cascade manifest. The relationships below are authored design proposals.

WM-05 checklist item · Five walkthroughs