TMC1: Typed Microcode Graph Compiler

Status: CPU mechanics passed; neural development gate failed; exact TMC1 closed

Date: 2026-08-10

Causal hypothesis

NMC1 failed before arithmetic: free-text generation lost register causality, commit identity, and stack grammar. TMC1 tests whether semantic planning is present but cannot be serialized reliably. It replaces textual program output with a model-owned typed computation graph. A generic executor alone enforces grammar and causality; it does not select operations, operands, constants, or answers.

Each graph instruction contains one operation (ADD, SUB, MUL, DIV, unary NEG, or identity COPY) and typed operands. COPY only materializes identity/alias equations as state and performs no semantic repair. An operand is either a pointer set over equal numeric spans in the source, one strictly prior state pointer, or a signed rational literal represented by fixed digit fields. Repeated equal source mentions remain an equivalence set rather than receiving an arbitrary owner. The final answer is a pointer to one computed state. No instruction contains an intermediate result or final-answer label.

Stage 0: exact CPU gate

Lower every immutable NMC1 gold register program into binary typed instructions. The train-derived maxima for source spans, instruction count, and literal digit widths become the only neural schema. Development may not increase them. The gate requires:

1. exact Fraction execution on all 6,333 train and 666 development rows;
2. exact frozen LAM1 execution on every row;
3. every final owner is a computed state;
4. zero missing source/state/literal owners; and
5. development fits the train-derived schema without truncation.

A miss closes this representation before GPU training.

Conditional neural canary

Job 749894 passed in seven seconds. Exact Fraction and learned-LAM execution are 6333/6333 train and 666/666 development. Train maxima are 15 source spans, 15 instructions, ten numerator digits, and three denominator digits; development maxima are 11/12/5/3. Report SHA-256 is f2b71afd...ab08.

The semantic owner is the immutable matched direct-CoT checkpoint from NMC1 (8a2b6550...0b53), which scored 267/666. Its Qwen3.5-0.8B weights and LoRA remain frozen. A 24,864,055-parameter causal typed slot decoder uses width 512, two source-encoder layers, four causal decoder layers, and eight heads. It cross-attends the frozen source token states and emits instruction count, operation, source-equivalence/state/literal references, and literal digits. Hard masks make forward state references and invalid instruction shapes impossible. There is no teacher-forced instruction stream; development is fully autonomous. Frozen LAM1 executes the graph.

The sole fit uses 4,096 updates, batch 32, AdamW (betas=(0.9,0.95), weight decay 0.01), learning rate 2e-4, maximum source prompt 512 tokens, seed 2026081061, and data seed 2026081062. Length, operation, left reference, right reference, and aggregate literal-digit losses each contribute one normalized component. No result/value/answer loss exists.

The first fit is bounded to one architecture, seed, and schedule frozen after the CPU receipt and before neural output. Development must satisfy all of:

• at least 45% exact answers and at least +5 absolute points over the frozen 40.090% direct owner;
• at least 80% operation and operand-owner accuracy;
• at most 10% exact answers under same-geometry source shuffle;
• at least 10-point carry-reset loss on multi-digit normal-correct rows;
• at least 30-point opcode-permutation loss; and
• zero invalid graphs, causal-reference violations, or slot exhaustion.

Failure closes exact TMC1 without width, layer, duration, seed, loss, pointer, or schema variants. A pass opens one unchanged public GSM8K test evaluation.

Claim boundary

TMC1 is a structured program compiler test, not a broad-reasoning claim. The generic grammar mask and executor are explicit scaffolds. A pass would show that model-owned natural-language semantics can populate a result-free causal graph that learned arithmetic microcode executes. It would not make typed decoding, pointer networks, or program execution novel.

Neural result

Fit 749895 completed all 4,096 updates in 859.33 seconds over 131,072 presentations / 12,651,233 source tokens. Final aggregate loss was 0.8836; checkpoint SHA-256 is a8742fca...1336. Frozen evaluations 749923/749924 completed in 22/20 seconds.

Normal development reaches 44/666 = 6.6066% answers, 26/666 = 3.9039% exact graphs, 290/666 = 43.5435% instruction counts, 1203/2646 = 45.4649% operations, and 1735/5285 = 32.8288% operand owners. One graph produces an invalid normal execution. Same-geometry source shuffle falls to 5/666, opcode permutation to 2/666, and carry reset retains only 8/36 multi-digit normal-correct rows. Thus the typed compiler is source- and executor-causal, but remains far below the frozen direct owner at 267/666 = 40.0901%.

The public test remains unopened. Exact TMC1 closes without width, layer, duration, seed, loss, pointer, or schema variants. The evidence localizes the next change: grammar is no longer the bottleneck; a static question-only hidden state does not expose the autoregressive semantic plan. A successor may use an exact model-owned draft trajectory as a second source stream, with aligned, shuffled-draft, and source-shuffled controls.