MLTC1: Monotonic Lexical Transduction Compiler

Status: closed development failure; holdout sealed

Date: 2026-08-10

Predecessor: PSTC1 closed at 91.8305% exact complete skeleton

Holdout: sealed

Hypothesis

PSTC1 learned a causal stack but its free-running controller lost source position as programs became longer and scopes nested. MLTC1 removes global action generation. A generic lexical candidate extractor exposes every source-owned number span and every + - * / ( ) character in monotonically increasing source order. A neural transducer assigns each candidate exactly one role:

IGNORE, NUMBER, NEGATE, ADD, SUB, MUL, DIV, LPAREN, RPAREN.

A fixed shunting-yard executor consumes only those predicted roles and copied number pointers, emitting the same typed PUSH/NEGATE/APPLY/STOP program used by PSTC1. The executor cannot inspect an answer, call a verifier, repair a prediction, or infer a role from source text. Raw source access is limited to copying the span selected by a predicted NUMBER role.

This is structurally different from PSTC1: source traversal is monotonic and one prediction is bound to one source candidate. The learned module no longer has to remember which source symbol it should emit next. Scope is materialized by the generic precedence stack over model-selected lexical roles.

Frozen model and budget

• pinned frozen Qwen3.5-0.8B source encoder;
• one width-384, four-block bidirectional candidate encoder;
• source-span pooling, candidate-surface embedding, monotonic position embedding, and one role head with surface-valid hard masks;
• no recurrent action decoder and no learned arithmetic executor;
• exactly 1,024 updates, batch 32, AdamW LR 2e-4, betas (0.9,0.95), weight decay 0.01, gradient clip 1, one seed;
• candidate-role cross entropy with IGNORE weighted 0.25 and every selected role weighted 1.0;
• 32,768 charged examples, identical to FSTC1/PSTC1;
• fewer than 20M trainable parameters.

The independent CPU builder must reproduce every PSTC1 gold action program exactly from gold lexical roles before a GPU fit opens. Any row mismatch is fatal.

Controls

1. same-family/binary-depth source shuffle under unchanged weights;
2. identical predicted roles with a flat left-to-right executor that removes precedence and parenthesis state;
3. within-batch candidate-state permutation while candidate surfaces and source positions remain fixed;
4. frozen PSTC1 and FSTC1 references.

Development gate

The gate is conjunctive:

• lexical role sequence exact >=99%;
• selected lexical sequence exact >=99%;
• generic-executor valid program >=99.5%;
• exact materialized operation skeleton >=97%;
• every-family exact skeleton >=95%;
• mixed-precedence, unary-group, and three-plus-parenthesis exact skeleton each >=90%;
• source-shuffled exact skeleton <=25% and aligned margin >=70 points;
• flat execution loses >=35 points on hierarchical rows;
• candidate-state permutation loses >=35 points overall;
• zero invalid source pointers, overlap, truncation, or decode fallback.

One development pass opens exactly one sealed holdout. Failure closes MLTC1 without width, depth, duration, seed, LR, role-vocabulary, tokenizer, loss, or threshold variants. Arithmetic transition learning remains closed until this compiler gate passes.

Claim boundary

A pass establishes accurate model-owned lexical selection plus deterministic hierarchical program materialization. It does not establish learned arithmetic execution, broad language reasoning, or novelty of shunting-yard parsing.

CPU admission

Job 749675 admitted all 75,935 training and all 3,917 development rows with exact extensional parity to PSTC1. Maximum candidate counts are 36 and 30, below the frozen 64-slot bound. Train/development SHA-256 are 33b8eb7bf154f3e938de00f06aebd7adb720eb4a1b6bd3f81554f9eb0b11bf3f and 8e867e7cdcc47015096979314450790cd86c804fd841b34d3df411a236b33679; report SHA-256 is 8bcb025f376a2d3d481cbe8b00e878c09baa850e5f71683d424bae4e8a0d7e0f.

Development result

Mechanics job 749678, full fit 749679, and evaluations 749680--749683 completed cleanly. The compiler has 7,522,569 trainable parameters and trained for 1,024 updates / 32,768 examples in 156.0398 seconds (209.9978 examples/s), with 2,267,000,832 peak GPU bytes. Loss fell from 1.03134 to 5.07e-9. Checkpoint SHA-256 is 62eafc90b81c62c5834a9d139307a805c33c75ba080c52b70b10fe6625fe8f60.

Normal development is 3,917/3,917 = 100% for lexical-role sequence, selected-lexeme sequence, action sequence, executor validity, and exact skeleton. Every family and every frozen difficult group is 100%. Source shuffle is zero. Flat execution loses 62.2244 points on hierarchical rows, showing that the precedence stack is essential.

The conjunctive gate nevertheless fails. Permuting contextual candidate states while retaining surface-type and monotonic-position embeddings leaves 3,337/3,917 = 85.19% exact, only a 14.8073-point loss rather than the frozen 35-point minimum. Thus the candidate extractor and surface metadata carry most of the program. The perfect capability score is an engineered compiler upper bound, not evidence for a sufficiently model-owned source compiler. Holdout remains sealed and MLTC1 closes without nearby variants.

Result and training-report SHA-256 are c3a8b4bce23989802213572ef18d18322442c5ca4513bdd178851e4c5d52bebc and a5ca9639d4deafb932bb3257a319ded453381a1e675213f25b566862fd36e82d. The next admissible compiler must consume the complete raw byte tape and emit per-byte lexical ownership directly, without pre-extracted candidates or surface-type metadata.