BTT1: Raw Byte-Tape Transduction Compiler

Status: closed development failure; holdout sealed

Date: 2026-08-10

Predecessor: MLTC1 closed despite 100% normal because pre-extracted candidate surface/position metadata retained 85.19% under contextual-state permutation

Holdout: sealed

Hypothesis

MLTC1 proves that monotonic lexical selection plus generic precedence execution can eliminate PSTC1's source-position drift, but its candidate extractor owns too much syntax. BTT1 removes every pre-extracted number, operator, parenthesis, and surface-type input. A standalone bidirectional byte transducer receives only the complete raw ASCII question and absolute tape position. It labels every byte as:

IGNORE, NUM_BEGIN, NUM_CONT, NEGATE, ADD, SUB, MUL, DIV, LPAREN, RPAREN.

A generic executor copies bytes from each predicted numeric span and applies shunting-yard precedence to the predicted operator/parenthesis labels. It cannot inspect an answer, invoke a verifier, infer symbols from raw bytes, or repair malformed output. A complete program is therefore impossible unless the model locates expression ownership, number boundaries, unary signs, operators, and scope on the raw source tape.

Frozen model and budget

• byte IDs 0..255, one padding ID, and learned absolute positions;
• width 256, six bidirectional Transformer blocks, eight heads, FFN width 1024, no pretrained backbone and fewer than 10M trainable parameters;
• maximum tape length admitted by CPU audit, hard-capped at 512;
• per-byte cross entropy with IGNORE weight 0.1 and every selected role weight 1.0;
• 1,024 updates, batch 64, AdamW LR 3e-4, betas (0.9,0.95), weight decay 0.01, gradient clip 1, one seed;
• exactly 65,536 charged examples; no recurrence, decoding search, answer labels, arithmetic values, or verifier feedback.

Controls

1. same-family/depth full-source shuffle under unchanged weights;
2. all active input byte IDs zeroed while tape length/positions and original executor copy source remain fixed;
3. identical predicted byte roles with precedence and parentheses removed by a flat left-to-right executor;
4. frozen MLTC1/PSTC1/FSTC1 references.

Development gate

• complete byte-role sequence exact >=99%;
• selected byte-role sequence exact >=99%;
• generic-executor valid program >=99.5%;
• exact materialized operation skeleton >=97%;
• every family >=95%;
• mixed precedence, unary groups, and three-plus-parenthesis each >=90%;
• source shuffle and zero-byte controls each <=25% exact with aligned margin >=70 points;
• flat execution loses >=35 points on hierarchical rows;
• zero byte truncation, malformed numeric copies, invalid programs, or fallback.

One development pass opens exactly one sealed holdout. Failure closes BTT1 without width, depth, duration, seed, LR, role vocabulary, loss, or threshold variants. A pass opens arithmetic-state learning over the sealed compiled program but does not by itself establish arithmetic reasoning.

CPU admission

Job 749685 admitted all 75,935 training and all 3,917 development rows with exact byte-role-to-program parity and no exclusions. The maximum raw tape length is 174 training bytes / 172 development bytes, below the frozen 512-byte cap. Train/development SHA-256 are 8deabd1ad1cdd93466cf9a53bde3012886fd63ff327914897bd70cbd9e13bc6e and de16d5dc8dd3676f4c3f4a69c306ec19a0ea8a42e9750bf93a2aaf16f4be0bf8; report SHA-256 is a69c5dd401fb91bf6da057cd38c4c1834c834de79da9f51594fbf27bbf80993d.

Development result

Mechanics 749700, fit 749701, and evaluations 749702--749705 completed. The 4,938,506-parameter compiler trained for 1,024 updates / 65,536 examples in 17.9827 seconds (3,644.3841 examples/s) and peaked at 832,687,616 GPU bytes. Checkpoint SHA-256 is 2283c86b1a640c9d5c02ffbc70b4646a0a1ad3538c4e1d4ab4ea3b164363e278.

Normal development is 3,872/3,917 = 98.8512% exact for complete byte roles, selected lexemes, action sequence, validity, and operation skeleton. Source shuffle and zero-byte controls both score zero exact skeletons. Removing precedence and parentheses loses 59.5190 points on hierarchical rows. The compiler is therefore strongly source-causal and no longer depends on MLTC1's candidate metadata.

The frozen conjunctive gate fails. Byte-role/selected sequence miss the 99% floors, validity misses 99.5%, and three-plus-parenthesis exact is 81.8182%. Every one of the 45 failures is an invalid parse rather than a valid but wrong program. Thirty-nine occur at binary depth five; exact by depths one through five is 100%, 99.85%, 100%, 99.21%, and 93.64%. This identifies a sequence-consistency defect at deep nesting, not source ownership or semantic operation selection.

BTT1 closes without nearby model, role, loss, duration, seed, or threshold variants; holdout remains sealed. Result and training-report SHA-256 are eeaff62f41d4240ede25292d7f9bb718afe58fec1c96a2c8fa6863b1247175cf and ee34845f85b9d811ca9716de3df3603a5bb9a910b65069ce06c71b373e0a8957. The admissible successor is a single weighted grammar-constrained projection over frozen BTT1 role logits, not another fit.