DTMC1: Draft-Conditioned Typed Microcode Compiler

Status: development gate failed; exact DTMC1 closed without public test

Date: 2026-08-10

Hypothesis

TMC1 proved that a typed source compiler is causal but recovered only 45.46% of operations and 32.83% of operands from one frozen question encoding. The same direct owner solves 40.09% only after autoregressive generation. DTMC1 tests whether that generated trajectory externalizes the missing semantic plan in a form a typed result-free compiler can use.

The system has two model-owned stages:

1. the immutable NMC1 direct-CoT owner greedily generates one exact draft from the source; and
2. a fresh typed graph compiler reads source plus that complete draft and emits the same result-free causal graph used by TMC1. Frozen learned LAM1 executes it.

At inference there is no gold chain, verifier, solver, answer label, host repair, or task router. Candidate source pointers remain restricted to the source segment; the draft supplies semantic trajectory context, not pointer answers. The direct draft may contain arithmetic results, but the typed target and compiler output contain no result field or final answer. LAM1 still owns execution.

Frozen draft corpus

Generate one greedy, no-thinking, maximum-512-token draft for each of the 6,333 training identities using direct checkpoint SHA-256 8a2b6550...0b53, pinned Qwen revision 2fc06364...8b17, seed 2026081053, and the same system/user prompt used by the frozen direct development evaluation. Eight identity-modulo shards are allowed only for wall-clock parallelism. Every row, including an exhausted or wrong draft, is retained. Merge requires exact identity coverage, zero duplicates, immutable shard hashes, and no public-test access.

Development uses the already generated direct report SHA-256 234a029a...a49; drafts may not be regenerated or selected.

Current execution status

All eight immutable generation shards completed after exact replacements for two tasks assigned to a node that exposed no CUDA device. The merged corpus contains 6,333/6,333 unique identities and 1,697,360 generated tokens. Draft answers are correct on 2,359/6,333 = 37.2493%; all 1,904 exhausted drafts are retained. Corpus SHA-256 is dfa3541f...0b1c.

The real-tokenizer audit passes every training row and all three 666-row development-control views. Maximum lengths are 756/1024 training tokens and at most 729/1024 development tokens. Numeric pointer custody is confined to the unique source prefix through the MODEL-OWNED DRAFT marker, with zero truncation and zero draft-number leakage.

Frozen fit 749998 completed all 4,096 updates in 2,741.95 seconds. It charged 131,072 examples and 49,385,260 source-plus-draft tokens at 47.80 examples/s; peak GPU allocation was 4.31 GB. Final loss was 0.8280 and checkpoint SHA-256 is 8d7ed14b...0da4.

Frozen evaluations 750031--750033 then closed the gate:

Aligned operation accuracy is 1312/2756 = 47.6052%; operand-owner accuracy is 1784/5506 = 32.4010%; exact graph is 20/666; normal validity is 664/666. Carry reset retains 14/36 normally solved multi-digit rows and opcode permutation scores 0/666. Thus both source and draft information are causal, but the draft-conditioned fixed-slot compiler adds only one answer over TMC1 and remains dramatically below its own direct owner. It fails every capability and structural floor plus zero-invalidity. Aggregate SHA-256 is 6e8ea75e...720f; SHOHIN_DTMC1_RESULT.json stores the frozen reducer. Public GSM8K test was never opened.

Frozen fit

DTMC1 reuses the exact TMC1 24,864,055-parameter compiler geometry and frozen semantic owner. Only the source representation changes from source to source + exact model-owned draft, with segment custody preserving numeric pointer masks over source only. Initialization seed, 4,096 updates, batch 32, LR 2e-4, optimizer, loss components, data order, and LAM executor remain identical to TMC1.

The immutable input envelope is PROBLEM:\n{source}\n\nMODEL-OWNED DRAFT:\n{draft} under the same system prompt and no-thinking chat template. The context ceiling is 1,024 tokens and truncation is forbidden. This is the smallest increase from TMC1's 512-token question-only ceiling that can retain the frozen 512-token drafts. Token offsets for candidate numeric owners must intersect the PROBLEM source bytes only; numeric strings in the draft are never pointer candidates.

Development controls and gate

Evaluate exactly once on all 666 existing source-disjoint identities:

1. aligned source plus its exact owner draft;
2. source plus a within-register-depth shuffled owner draft;
3. same-depth shuffled source plus that donor's aligned draft; and
4. frozen direct owner (267/666) and question-only TMC1 (44/666) as fixed references.

All gates are conjunctive:

• aligned answers at least 301/666 (45.195%), which exceeds direct by at least five absolute points;
• operation and operand-owner accuracy each at least 80%;
• aligned exceeds shuffled-draft by at least ten points;
• source+draft shuffle scores at most 10%;
• carry reset loses at least ten points on multi-digit aligned-correct rows;
• opcode permutation loses at least 30 points; and
• zero invalid graphs or executions.

Failure closes exact DTMC1 without draft count, sampling, rank, width, layer, duration, seed, loss, or prompt variants. A pass opens one unchanged public GSM8K test evaluation of direct and DTMC1.

Claim boundary

DTMC1 is a closed negative, not proof of general reasoning or a novelty claim for draft-and-revise systems. It shows that the model-owned draft carries causal information, because matched draft shuffling removes 40 of 45 solved rows, but the fixed-slot full-graph decoder cannot convert that information into useful capability. No rank, width, layer, duration, seed, loss, or prompt retry is authorized. A successor must make its supervised action explicitly draft-specific, such as pointer/edit sequence transduction with deterministic generic execution, rather than predict the same complete graph from every presentation of a source.