NMC1: Natural-Language Microcode Compiler

Status: development gate failed; exact NMC1 closed

Date: 2026-08-10

Objective

LAM1 proves that learned finite decimal transitions compose exactly once a correct postfix program exists. NMC1 tests the missing semantic interface: can a pretrained language model compile a natural word problem into a compact, result-free program that the frozen learned microcode executes more reliably than an equal-update ordinary chain-of-thought SFT control?

This is structurally different from BTT/WGP. The source is a natural word problem, not an explicit arithmetic expression. A pinned Qwen3.5-0.8B source compiler emits a register microprogram. The target contains operations, source constants, and references to earlier computed registers, but omits every stated equation result and the final answer. At inference the frozen LAM1 microcode executes the generated program. There is no verifier, solver, answer label, teacher, or host repair.

Frozen data

• source: immutable GSM8K train JSONL SHA-256 e0a44460964649db791a0ced18449b34c34875218b7d21b23241cbdb6ca2a104;
• overlap reference and later public benchmark: GSM8K test JSONL SHA-256 752adc99f23132275139f1f7be57126b08735a55daf0c56859ee6df70abafbac;
• accept only rows whose complete <<expression=result>> annotations parse into exact +, -, *, and / arithmetic and whose final register equals the #### answer;
• reject rows with no complete equation program or unsupported operators;
• replace an operand by the latest prior register whenever its exact value is already available;
• split admitted train identities by normalized-question SHA-256 modulo 10: remainder zero is development and all others are training;
• require zero normalized-question overlap between train, development, and public test.

The program serialization contains no computed record result and no final answer field. An independent Fraction assessor and frozen learned-digit executor must agree on every admitted gold program before GPU training.

Frozen model and fits

Model is pinned Qwen/Qwen3.5-0.8B revision 2fc06364715b967f1860aea9cf38778875588b17, locally hash-bound. Both arms use fresh final-four-layer rank-8 all-projection LoRA, alpha 16, learning rate 2e-5, batch 4, gradient accumulation 2, maximum sequence 1,024, exactly 1,024 updates, seed 2026081051, and data seed 2026081052. Token admission must prove that every complete prompt and target, including EOS, fits without truncation in both arms before either fit starts.

Arms:

1. NMC1: source to result-free register microprogram, followed by frozen LAM1 execution.
2. Direct control: same source and admitted identities, ordinary original GSM8K chain-of-thought/final response, with identical model, LoRA geometry, updates, batch, optimizer, and decoding budget.

Actual non-padding training tokens, elapsed time, peak memory, trainable parameters, and generated tokens are reported. The direct arm is allowed to consume more target tokens; NMC1 receives no compute credit for being shorter.

Data and mechanics admission

CPU job 749812 admitted 6,333 matched training identities and 666 source-disjoint development identities. They contain 22,846 computation records and 75,242 actions. Rejections were 95 rows without a complete equation program, two unsupported floor-division rows, and 377 rows whose annotation did not end in the final answer. Program/direct train SHA-256 are 71010bac117d9fbca2a28e7f9a63f24456b519a7b4bc172ddd772ff27e851512 / b78fa53081114f1b5c08f34c0f26df641dfd975a4975130a6b0c8098a0944757; development SHA-256 is 981b83016d9a895af3016b3629c5852b868c80e364dbe97e3901a8d3e8ced4bd.

Token audit job 749813 retained every row under the frozen 1,024-token limit. Maximum complete sequence is 363 tokens for NMC1 and 574 for direct CoT. Mechanics job 749814 obtains exact Fraction/LAM parity on all 6333/6333 train and 666/666 development programs with zero normal invalidity. On gold development programs, carry reset scores 116/666 and opcode permutation 13/666. Both prerequisite gates pass.

Development gate

Before any public-test score is opened, all conditions are conjunctive on the source-disjoint development split:

1. gold-program Fraction/LAM parity is 100%, with zero invalid or overflow;
2. NMC1 syntax validity is at least 90%;
3. NMC1 execution validity is at least 85%;
4. NMC1 final-answer accuracy is at least 60%;
5. NMC1 beats the matched direct control by at least 3 absolute points;
6. same-family/depth source shuffle scores at most 10%;
7. carry reset loses at least 10 points on multi-digit normal-correct rows;
8. opcode permutation loses at least 30 points overall; and
9. no admitted target serializes any stated equation result or final answer.

Failure closes exact NMC1 without prompt, rank, layer, duration, seed, split, parser, decoding, or threshold variants. A pass opens one unchanged full public-test evaluation of both arms. The test is a standard public benchmark, not a sealed project holdout, and will be labeled accordingly.

Claim boundary

A development pass would show that a natural-language model can own semantic program selection while learned recurrent microcode owns arithmetic execution. It would not establish broad mathematical reasoning, novelty of program-aided language models, or generalization beyond the admitted arithmetic language. The changed factor is the result-free register interface plus learned local execution, measured against ordinary matched SFT.

Result

Matched fits 749816/749817 completed all 1,024 updates with 901,888 trainable parameters each. Program/direct checkpoints have SHA-256 eef6cbe6...0ae1 / 8a2b6550...0b53. The first evaluation launch 749835--749837 failed closed before generation because its receipt validator looked for the top-level checkpoint update inside metadata. Commit f24a237 corrected only that custody check; it did not alter weights, data, prompts, decoding, thresholds, or the public-test state.

Frozen evaluations 749838--749840 reject NMC1. Program syntax is 40/666 = 6.006%, execution validity 19/666 = 2.853%, exact programs and final answers are both 0/666; source shuffle is also zero. The matched direct arm scores 267/666 = 40.090%. Program generation exhausts on 78 rows. Read-only parse attribution finds 335 noncausal register loads, 155 wrong final commits, 78 missing complete envelopes, and 58 other structural failures. Among 40 parsable programs, 18 have a wrong terminal stack and three underflow.

Every capability gate except the source-shuffle ceiling fails. Public GSM8K test remains unopened. Exact NMC1 closes without prompt, rank, layer, duration, seed, split, parser, decoding, or threshold variants. The result isolates the failure before arithmetic execution: free-text autoregression did not maintain typed register causality. The next admissible successor must replace text serialization with a grammar-masked typed instruction policy and generic executor, not merely train NMC1 longer.