R12 Sparse-Law Neural Microcode Preregistration
Trigger
Both architecture-generic table completers fail on hash-disjoint unseen operators:
| Candidate | Development state accuracy | Complete maps | Exact queries |
|---|---|---|---|
Direct set attention (704748) | 46.5000% | 0/60 | 4/60 |
Learned generator factorization (704750) | 15.7083% | 0/60 | 1/60 |
The factorized treatment also loses its direction-negated control in exact queries. It is closed.
Hypothesis
A 125M-scale language model may need an internal computational substrate rather than being expected to rediscover arithmetic and bitwise execution in its weights. The new treatment separates:
- a learned byte controller that orients sparse demonstrations and predicts operator-family microcode plus parameters; and
- a deterministic finite-domain ALU inside the model forward pass that executes the microcode and emits a transition distribution.
This is architecture-native execution: no host callback, parser, solver, search, or posthoc verifier runs at candidate time.
Fixed ALU
The ALU exposes three operation schemas over domains 8 and 16:
- modular affine;
- rotate/xor; and
- Gray-conjugated modular affine.
The controller predicts family, multiplier, offset, rotation, and mask distributions. The ALU evaluates their differentiable mixture. Program parameters in development are hash-disjoint from training.
Training receives preparation-only exact labels for the operation family and its relevant parameters in addition to complete-map and source-direction losses. Development receives no labels at candidate time; those labels are used only for scoring. This treatment therefore tests supervised induction into a fixed internal instruction set, not discovery of that instruction set.
This is deliberately an ontology-bearing architecture and cannot establish open-ended law discovery. Its purpose is to test whether explicit internal microcode closes the sparse-completion gap.
Frozen Canary
The board, source deletion, map partition, optimization budget, and same-weight controls are unchanged from the generator-factorization canary. Four counterfactual source orders teach the held-out relation lexemes without including the exact passive renderer.
Continue only if the treatment:
- exceeds 46.5% development transition accuracy;
- produces at least one complete unseen map;
- beats every same-weight control in exact query accuracy; and
- retains zero training/development action-map overlap.
A pass authorizes a five-seed microcode qualification, not a general-reasoning claim.
Frozen Outcome
Job 704760 completed the preregistered 2,000-update H100 canary. Treatment
reached 22.1250% development transition accuracy, 0/60 complete maps, and
0/60 exact queries. The direction-negated same-weight control reached 1/60
exact queries. Training/development action-map overlap was zero.
All continuation criteria fail. Decision:
supervised_microcode_fails_hash_disjoint_sparse_law_induction.
Full evidence: R12_SPARSE_LAW_MICROCODE_RESULT.md.