# R12 Episodic-Generator Constraint Preregistration

**Post-result disposition:** invalidated as an architecture-native Shohin
reasoning gate. The experiment did not load Shohin, used exact host parsing,
exhaustive word enumeration, host sealing/execution, and failed the frozen
abstract target-program holdout. Retain this document as historical protocol.

## Question

Can the sparse-law mechanism remove its fixed global operation library and
construct a temporary hypothesis space from unfamiliar generators supplied
inside the episode?

## Board

Every episode contains four opaque actions:

1. two complete support-generator transition tables; and
2. two sparse target actions that are compositions of those generators.

The target records are inclusion-minimal for a unique map inside the
episode-local closure. Every late-query target transition is absent from the
source. Source and query render through six grammars.

Training uses cyclic/dihedral and bitwise generator systems. Development
holds out:

- target programs and maps;
- longer target-program depths;
- cardinality 16;
- the passive renderer;
- a random-permutation generator family; and
- joint combinations of those shifts.

The frozen board must independently audit exact source-deleted mechanics,
record necessity, hidden query steps, renderer invariance, target-map
disjointness, action-order causality, and target-law causality before fitting.

## Candidate

The byte reader predicts source-versus-target direction. Actions with one
record per state become episode-local generators. A tied tensor recurrence
constructs all generator words through depth six. Sparse action records remain
separate factors and intersect that temporary program space. The resulting
maps are sealed; source tensors are deleted before categorical query
execution.

The architecture contains no global law bank and receives no family,
generator-word, or target-program label. Candidate inference calls no host
parser, callback, search routine, oracle, solver, or verifier.

## Controls

All causal controls use the same trained weights:

- source direction negated;
- sparse target observations shifted;
- all observations zeroed;
- one support generator semantically deranged;
- one necessary target witness deleted;
- record order permuted; and
- support action names and physical record order jointly recoded.

Collision-prone controls are compared against exact mechanics ceilings frozen
before neural scoring. Record permutation and consistent opaque-key recoding
must be exactly invariant.

## Advancement Gates

One mechanics canary advances only if:

- treatment reaches at least 95% complete target maps and exact queries;
- every development cell and family reaches at least 90%;
- every query step is hidden from sparse target records;
- target-map training/development overlap is zero;
- shifted, zeroed, deranged-support, and witness-deletion controls remain
  below their frozen ceilings;
- record permutation and opaque-key recoding are exactly invariant;
- every deployed packet excludes source bytes; and
- the complete conceptual system remains below 200M parameters.

A pass establishes bounded episodic construction and use of a new finite
program space. Generic closure, finite depth, exact equality, sealing, and
execution remain fixed algorithms. It is not unrestricted language reasoning
or an open-ended task ontology.