# R12 Variable-Topology Semantic Compiler Result

## Decision

The learned global semantic-partition compiler passes the frozen
variable-topology qualification:

`global_semantic_partition_passes_variable_topology_gate`

This is the strongest controlled reasoning result in the Shohin campaign. It
demonstrates systematic compilation and execution of complete anonymous
finite machines across unseen laws, compositions, topologies, renderers, and
three held-out generator families. It does **not** demonstrate Shohin-native
general reasoning.

## Claim-Bearing Result

Newton CPU array `704729` completed all 15 preregistered
five-seed/leave-one-family-out folds. The independent audit passed without an
exception.

| Arm | Exact | Rate | Margin below treatment |
|---|---:|---:|---:|
| Treatment | **360/360** | **100.0000%** | - |
| Same weights, source direction swapped | 152/360 | 42.2222% | 57.7778 points |
| Same weights, global key scores negated | 0/360 | 0.0000% | 100 points |
| Same weights, query roles swapped | 0/360 | 0.0000% | 100 points |

Every one of the 15 treatment folds scored 24/24. Every treatment-control
direction was positive. Every treatment fold also passed:

- collision: 8/8;
- joint collision plus held-out renderer: 4/4;
- unseen composition: 4/4;
- unseen law: 4/4;
- held-out renderer: 4/4; and
- unseen topology: 4/4.

Per held-out family, treatment scored 120/120. The direction control scored
70/120 for affine modular, 43/120 for bitwise rotate/xor, and 39/120 for
permutation. Negating either the learned global key partition or learned
query-role channel caused all 360 seals/executions to fail closed.

## Candidate

- protected Shohin parameters: 125,081,664, frozen and unchanged;
- learned sidecar compiler parameters: 60,613;
- conceptual complete system: 125,142,277;
- complete-system limit: 200,000,000;
- optimizer updates: 100 per fold;
- fitting episodes: 40 per fold;
- counterfactual renderer episodes: 160 per fold;
- trained models: one per fold;
- control updates: zero, because all controls intervene on the same weights;
- candidate-time oracle/search/verifier calls: 0/0/0; and
- preparation exact/source/query parser calls: 600 each across 15 folds.

The compiler learns three separable semantic channels:

1. source/target direction for each transition record;
2. one episode-global partition of opaque keys into states and actions; and
3. start/action roles in the late query.

It seals a discrete anonymous transition packet, round-trips that packet
through the deployed wire representation, and executes only from the sealed
packet and late query. Raw source bytes are absent from the deployed wire.

## Why The Controls Are Causal

The earlier state/action channel-swap control was rejected because aggregated
occurrence evidence could cancel and leave the selected top-k key partition
unchanged. The final protocol directly negates the episode-global learned key
scores. This changes the selected semantic partition or makes record typing
invalid on every evaluated case.

The direction and query controls also modify only inference semantics under
the treatment weights. They consume no extra samples, parameters, or
optimizer updates. A positive result therefore cannot be attributed to easier
control optimization or a different initialization.

## Scope Boundary

This result establishes a bounded form of systematic reasoning:

- laws are unseen;
- compositions are longer;
- cardinality varies over 4, 8, and 16;
- action count varies from 2 through 5;
- state/action frequency is deliberately ambiguous in collision cells;
- source and query renderers are compositionally held out; and
- one generator family is excluded from each fit.

It does not establish general reasoning because:

- every episode provides the complete transition table;
- all tasks reduce to one finite-machine ontology;
- supported geometry and action quota are part of the compiler contract;
- packet execution is a fixed discrete executor;
- source deletion is enforced at the deployed-wire interface, not by a
  separately attested fresh process; and
- the 60,613 learned parameters are a sidecar, not integrated into or emitted
  by the 125M Shohin transformer.

The result should be described as **causal variable-topology anonymous-machine
compilation**, not unrestricted natural-language, mathematical, or general
reasoning.

## Evidence

- audit:
  `artifacts/r12/source_deleted_variable_topology_board_v1/variable_topology_curriculum_audit.json`
- audit SHA-256:
  `049bbbd398e6f1456d6f1809bebb701351053302e0ce44147e70735d5b155fa2`
- fold reports: 15 files under
  `artifacts/r12/source_deleted_variable_topology_board_v1/`
- five independently recomputed board manifests: five distinct hashes
- focused verification: 23 tests passed, clean Ruff, byte compilation, shell
  syntax, and Git diff checks

## Final Recommendation

Stop optimizing this solved complete-table board. The next justified gate is
**sparse latent-law induction**:

1. reveal only a strict subset of transitions;
2. require prediction and composition of withheld transitions;
3. vary the rule ontology rather than only the machine instance;
4. preserve source deletion and same-weight semantic controls;
5. include a separately attested executor process; and
6. integrate the compiler into Shohin only after the sparse-law gate passes.

Continuation pretraining remains held by user instruction. No experiment in
this result modifies the protected checkpoint.