# R12 Sparse-Law Constraint-Intersection Preregistration

## Diagnosis

Direct attention, learned generators, and supervised microcode all compress a
set of individually necessary demonstrations into an action summary before
predicting a law. The sparse board was constructed so that every retained
record eliminates at least one otherwise-valid hypothesis. Mean compression
does not preserve that logical conjunction.

## Treatment

The treatment keeps every record as an independent compatibility factor. A
learned byte encoder predicts only whether the first or second number is the
source. A fixed internal program bank evaluates every observation against
every finite-domain program in parallel. Program energy is the sum of exact
record compatibilities; a high-temperature posterior emits a complete
transition distribution. Query execution occurs only from the sealed packet
after source bytes are discarded.

This is architecture-native tensor computation: candidate inference calls no
host parser, callback, solver, oracle, search routine, or verifier. It is also
an exhaustive fixed-ontology prior equivalent in semantics to finite
version-space elimination. A positive result is therefore bounded law
identification, not open-ended program discovery or general reasoning.

## Frozen Controls

The board, map partition, optimization budget, source-deletion boundary, and
query executor remain unchanged. Same-weight controls negate source direction,
shift observation values, or zero all observation compatibility.

## Gates

Continue only if one frozen seed:

- reaches at least 95% development transition accuracy;
- produces at least 57/60 complete unseen maps;
- reaches at least 54/60 exact queries;
- beats every same-weight control by at least 45 exact queries;
- has zero training/development action-map overlap; and
- keeps the complete conceptual system below 200M parameters.

A pass authorizes an open-ontology falsifier. It does not authorize a native
reasoning claim because the complete operation library is built into the
architecture.

## Frozen Outcome

Job `704761` reached 60/60 complete unseen maps and 60/60 exact queries with
100% transition accuracy and zero action-map overlap. Shifted targets scored
1/60 and zeroed observations 0/60. Direction negation scored 20/60, so the
frozen 45-query margin misses by five.

An exact inverse-law audit reproduces all 20 direction-control answers. The
maximum possible margin on this frozen board is therefore 40, making the
preregistered 45 margin unattainable even for perfect treatment. Preserve the
gate miss; classify the experiment as an exact bounded capability result with
a defective promotion criterion. Full evidence:
`R12_SPARSE_LAW_CONSTRAINT_INTERSECTION_RESULT.md`.