← Complete research archive
Architecture researchPreregistered78 lines

DIVERGE-QTE1: Qwen Transaction Entailment

PQI1 and GTI1 show that unconditional role classification and transaction emission memorize renderer layouts at 125M/135M. QTE1 changes both capability floor and semantic computation:

docs/research/DIVERGE_QTE1_QWEN_TRANSACTION_ENTAILMENT.mdOpen original Markdown ↗

DIVERGE-QTE1: Qwen Transaction Entailment

Status: closed after one sealed confirmation. No training, prompt repair, or QTE1 variant is authorized.

Hypothesis and boundary

PQI1 and GTI1 show that unconditional role classification and transaction emission memorize renderer layouts at 125M/135M. QTE1 changes both capability floor and semantic computation:

A pretrained 0.8B language model can ground a complete candidate transaction by judging whether a candidate-specific semantic claim is entailed by the source instruction. Whole-candidate entailment, not unconditional role emission, supplies the QUERY transaction.

The pinned parent is Qwen/Qwen3.5-0.8B@2fc06364715b967f1860aea9cf38778875588b17, loaded from the existing hash-verified Newton artifact. It receives the canonical anonymous query plus one of two complete claims: candidate A is the requested source and candidate B the distractor, or vice versa. It scores complete YES and NO conditional likelihoods and selects the candidate with larger entailment log odds. No target label, answer, state value, executor output, teacher, tool, search, update, or fallback enters inference.

QTE1 is a capability-floor component result, not a sub-200M Shohin result and not an open-domain reasoning claim. If qualified, its logits may later become distillation targets for a smaller model-owned semantic owner.

Fixed development evidence

Read-only job 744554 was implemented and committed before its score was known. On the opened 768-query development board it reaches 768/768, every mode 256/256, and every renderer 128/128. Context scrub falls to 384/768; mapped mention swap remains 768/768. The result SHA-256 is cbd3dbdd26bc894d5f8ff337ecc2b4e6996d09b390a300abfe245e35981bc269.

Sealed confirmation gate

The independent 256-episode PQI1 confirmation board remains sealed at SHA-256 27f198680cc7bcd7e0203949fe4dee1658fc057fe75302b7e2d43d74321201b8. One direct run of the unchanged scorer is admitted. Promotion requires all of:

  • QUERY at least 765/768;
  • every mode at least 254/256;
  • every renderer at least 127/128;
  • context scrub at most 430/768 and at least 250 below normal;
  • mapped mention swap at least 765/768;
  • exact pinned model/data hashes and only legal complete candidate claims.

A pass qualifies one zero-training composition with protected TOL3 WORLD, protected NVE1 EVIDENCE, exact factorized execution, and QTE1 QUERY. That composite must pass the existing end-to-end semantic floors before natural PL1 can run. A confirmation miss closes QTE1 without prompt, wording, label, normalization, model, renderer, or threshold variants.

Sealed result and attribution

Job 744558 scored QUERY 768/768, every mode 256/256, every renderer 128/128, and mapped mention swap 768/768. Context scrub scored 512/768, above the frozen 430/768 maximum, so QTE1 formally fails and dependency-held composition 744559 canceled without allocation. Result SHA-256 is ee55580e3aa0b140990ac983543b0eacd8657987e128b18d434f6e5749570b6e.

The one read-only accounting attribution explains the failed control without changing the decision. Every scrubbed source is exactly alpha then beta. The sealed board contains 512 transaction-0 targets and 256 transaction-1 targets, and QTE1 predicts transaction 0 on all 768 scrubbed rows. Its 512/768 is therefore exactly the majority-class baseline, not retained source semantics. Attribution SHA-256 is 7475c231956d0d5c5107c77344c06996eac7738170c3116bdad826c544e269d8.

QTE1 remains useful evidence that candidate-wise entailment at the 0.8B capability floor can read all held semantic renderers. It is not promoted as a causally qualified transaction owner because the frozen conjunctive gate failed. No QTE1 retry is allowed; the next mechanism is the independently frozen outcome-grounded CGL1 lane.