← Complete research archive
Architecture researchClosed / no-go169 lines

CTE1: Canonical Transaction Externalization

DTC1 proves that explicit model-owned arithmetic transactions are causal and execute reliably, but the ordinary direct owner emits an accepted transaction on only 257/666 development problems. CTE1 tests whether this is a post-training target defect rather than an architectural l…

docs/research/SHOHIN_CTE1_CANONICAL_TRANSACTION_EXTERNALIZATION.mdOpen original Markdown ↗

CTE1: Canonical Transaction Externalization

Status: closed negative on development; public test sealed

Date: 2026-08-10

Hypothesis

DTC1 proves that explicit model-owned arithmetic transactions are causal and execute reliably, but the ordinary direct owner emits an accepted transaction on only 257/666 development problems. CTE1 tests whether this is a post-training target defect rather than an architectural limit.

Train the same pinned Qwen3.5-0.8B backbone to externalize one compact, canonical arithmetic transaction trace directly from each word problem. A generic grammar lowers the generated trace into the unchanged typed graph and frozen learned LAM1 executes it. No learned fixed-slot graph decoder is used.

CTE1 changes the supervised representation, not model scale, data identities, optimizer budget, or executor. It is distinct from NMC1's result-free register language: the target uses the familiar GSM annotation form that the pretrained model already partially emits. It is also distinct from SLC1's broad synthetic addressable ledger and from DTMC1's nonautoregressive full-graph prediction.

Canonical Target

An independent training-only builder reads each immutable admitted gold register program and renders every causal record as:

<<fully_parenthesized_expression=exact_result>>
...
#### exact_final

PUSH operands use their exact rational surface. LOAD operands use the exact result of the referenced prior record, making the textual alias recoverable as a causal STATE link. Unary and binary operations are fully parenthesized. The final line is emitted for ordinary deployment readability, but DTC1 parsing ignores it; only the typed transaction graph reaches LAM1.

The builder must round-trip every target through the already frozen DTC1 parser and exact assessor, reproduce the immutable answer, preserve causal state ownership, and report all source/state/literal reads. It may not use the public test or alter source identities. Any row that fails exact mechanics is excluded and reported before training; admission requires all 6,333 train and 666 development rows.

Training prompt:

Emit a concise arithmetic transaction trace for the word problem. Use
<<expression=result>> for every step, then write #### followed by the final
result. Emit no other text.

PROBLEM:
{source}

Frozen Model And Budget

  • exact pinned Qwen/Qwen3.5-0.8B revision 2fc06364715b967f1860aea9cf38778875588b17;
  • fresh final-four-layer rank-8 LoRA over all linear projections, alpha 16;
  • base weights frozen, BF16, no quantization;
  • exactly 1,024 updates, batch 4, accumulation 2;
  • AdamW, LR 2e-5, existing trainer defaults and gradient clipping;
  • exact NMC1 model/data seeds 2026081051/2026081052;
  • all 6,333 existing source identities, no selection or curriculum;
  • 1,024-token training context with zero retained truncation;
  • greedy no-thinking evaluation, maximum 512 new tokens, seed 2026081053;
  • unchanged frozen learned-LAM1 checkpoint.

This exactly matches NMC1 direct/program training examples, optimizer steps, adapter geometry, learning rate, and seeds. Target token count is reported, not force-matched, because representation compactness is part of the tested intervention.

Development Controls

Evaluate exactly once on the existing 666 source-disjoint rows:

  1. aligned source;
  2. deterministic same-register-depth source shuffle, scored against the untouched target answer;
  3. state-read reset to zero on the aligned compiled graph;
  4. frozen LAM1 opcode permutation;
  5. immutable direct owner 267/666, NMC1 0/666, TMC1 44/666, DTMC1 45/666, and DTC1 108/666 as fixed references.

Prospective Gate

All conditions are conjunctive:

  • all 6,333 train and 666 development canonical targets pass exact CPU round-trip mechanics;
  • zero retained source/target truncation;
  • at least 600/666 generated traces compile and execute normally;
  • aligned reaches at least 300/666 exact answers;
  • aligned exceeds the immutable direct owner by at least 33 answers;
  • source shuffle is at most 67/666;
  • at least 300 aligned rows contain a causal state read;
  • state reset loses at least 20 points on aligned-correct linked rows;
  • opcode permutation loses at least 30 points from aligned;
  • zero normal execution invalidity among compiled rows; and
  • no public-test access.

A pass opens one separately frozen public GSM8K evaluation of CTE1 and the direct owner. A miss closes exact CTE1 without target-format, punctuation, prompt, rank, layer, update, LR, seed, decoding, parser, or threshold variants. No output fallback, verifier, selector, host repair, or answer extraction may be added after scoring.

Claim Boundary

A pass would establish that compact canonical trace post-training can connect a small pretrained language owner to causal learned execution and improve its source-disjoint arithmetic reasoning. It would not prove unrestricted general reasoning, architecture novelty for transaction notation, or a LAM1 holdout claim outside this separately defined GSM pathway.

Data And Mechanics Result

CPU job 750045 admits all 6,333 train and 666 development identities. The exact canonical corpus contains 20,678 / 2,168 transactions and 15,654 / 1,629 cross-record register loads. Every target parses and executes to the immutable terminal answer with no train/development overlap and no public-test access. Train/development SHA-256 values are 8fb68943...6625 and aff46617...eb04; report SHA-256 is abb12785...d556.

Tokenizer audit 750050 passes with zero truncation. Maximum complete lengths are 393/1024 train and 335/1024 development tokens. The train target contains 276,777 charged response tokens; development contains 29,215. Audit SHA-256 is 6f999a91...43b8.

Immutable runtime 1e21f38 has manifest SHA-256 9216c99e...48c1.

Frozen Development Result

Fit 750074 completed all 1,024 updates in 474 seconds. The checkpoint and training report SHA-256 values are fb0bae33...a1662 and 76439357...e2755. Concurrent immutable evaluations 750083/750084 completed in 653/658 seconds:

  • aligned exact answers: 134/666 = 20.12%;
  • source-shuffled exact answers: 4/666 = 0.60%;
  • immutable direct-owner reference: 267/666 = 40.09%;
  • compiled/executable traces: 599/598;
  • linked rows / linked-correct rows: 570/131;
  • state-reset linked-correct: 1/131;
  • opcode-permuted correct: 1/666;
  • normal execution invalid: one;
  • exhausted generations: 49; and
  • generated transactions: 3,294, including 2,548 state reads.

The traces are source-, state-, and opcode-causal. They are not an effective semantic planner: only 33 generated completions exactly match their canonical target, and answer accuracy decreases from 76/212 = 35.85% at gold depth two to zero across all 39 depth 6--8 rows. The owner often emits a valid but semantically incomplete or wrong arithmetic program; learned execution then faithfully computes that wrong program.

CTE1 fails the prospective capability, coverage, opcode-loss, and zero- invalidity conditions. It closes without target-format, prompt, parameter, duration, parser, or threshold variants. Public GSM8K test remains sealed. Normal/source-shuffled/aggregate report SHA-256 values are 8c75ca70...12954, 5eb863a4...a478c, and 8dbac6e1...1010b.