CTF1: Canonical Transaction Capability Floor

Status: closed negative on development; public test and holdout sealed

Date: 2026-08-10

Question

CTE1 taught pinned Qwen3.5-0.8B to emit executable canonical traces, but accuracy fell to 134/666 versus its direct owner at 267/666. DTC1, DTMC1, CTE1, and LTR1 jointly localize failure to semantic planning before learned execution. CTF1 changes only owner capacity: can an untouched, stronger same-family model already use the exact canonical transaction interface?

Frozen Intervention

• exact pinned Qwen/Qwen3.5-4B revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a;
• no adapter, fine-tuning, warm start, selector, verifier, or host repair;
• exact immutable 666 CTE1 development identities and prompt;
• greedy no-thinking generation, maximum 512 new tokens, seed 2026081053;
• unchanged DTC parser, typed graph, and learned LAM1 executor;
• normal and same-register-depth source-shuffled views;
• aligned state-reset and opcode-permutation interventions; and
• no holdout or public GSM8K access.

CTF1 is an interface/capability-floor test, not a parameter-matched comparison or a claim that scale is the Shohin mechanism.

Prospective Gate

Every condition is conjunctive:

• at least 600 traces compile and execute normally;
• aligned reaches at least 300/666 exact answers;
• aligned exceeds trained 0.8B CTE1 by at least 100 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 linked-correct rows;
• opcode permutation loses at least 30 points from aligned;
• zero normal execution invalidity; and
• public test remains closed.

A pass establishes that the source-to-canonical-ledger-to-learned-execution interface is viable above the observed 0.8B capacity floor and permits one separately frozen 4B post-training experiment. A miss closes this interface across both tested scales without prompt, parser, decoding, threshold, or nearby model variants.

Claim Boundary

CTF1 cannot establish that LAM improves an already-capable 4B model, because there is no matched direct-answer comparison in this read-only ceiling. It tests whether a capable model can own the semantic program consumed by the existing learned executor under causal controls.

Frozen Result

The first jobs 750098/750099 failed before generation because the new raw owner wrapper selected the adapter embedding path. Commit 80c11a0 changed only that dispatch boolean and added a guard. Immutable replays 750111 and 750112 then completed in 917/867 seconds from runtime manifest 36ec677f...fb97.

• aligned learned-execution answers: 419/666 = 62.9129%;
• source-shuffled answers: 7/666 = 1.0511%;
• compiled/executable rows: 562/562;
• linked rows: 544;
• state-reset linked-correct: 0/419;
• opcode-permuted correct: 3/666;
• normal execution invalid: zero; and
• exhausted generations: seven.

The score clears the capability, scale-margin, source-causality, state, opcode, and validity conditions, but misses the frozen >=600 coverage condition. CTF1 is therefore conjunctively FAIL; the public test remains sealed.

Read-only attribution clarifies the boundary. The untouched 4B owner's own claimed final answer is correct on 487/666. Path overlap is: 408 both direct and ledger correct, 79 direct-only, 11 ledger-only, and 168 neither, for an oracle union of 498. Of the 104 compile-invalid rows, 75 have a correct claimed final answer. Stronger scale restores semantic planning, but forcing every answer through this exact canonical ledger discards more correct model-owned answers than it repairs.

Exact CTF1 closes without prompt, parser, decoding, threshold, adjacent scale, or post-training rescue. Normal/source-shuffled/aggregate report SHA-256 values are dc4e939b...05f0, 9ce8dae6...2c63, and 8bccbe9a...5b77.