DIVERGE-QPT1: Qwen3.5-4B Pointer-Transaction Gate

Status: complete and closed. The score gate produced a large aggregate lift but failed the conjunctive no-regression rule; controls remain unopened.

Objective

QST1 improved math, science, and logic on Qwen3.5-0.8B but failed broad promotion because its unconditional 20-token workspace prefix regressed code and produced only a +1.007 five-domain macro lift. Its learned halt also closed almost completely by the end of training, reducing effective recurrent depth.

QPT1 tests a structurally different mechanism on exact pinned Qwen/Qwen3.5-4B@851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a:

1. query-conditioned straight-through pointers select complete prompt observations into eight source slots;
2. each of eight tied recurrent steps reads one whole source slot and writes one whole state slot;
3. fixed recurrent depth replaces learned early halting;
4. four query-owned outputs are injected into existing terminal prompt tokens through a zero-initialized gated residual, preserving sequence length and positional geometry;
5. the final-four-layer rank-8 LoRA path remains identical to the matched B1 control.

The model owns every pointer, transaction, gate, and answer. There is no external solver, tool call, answer router, host arithmetic, or teacher at inference.

Frozen Development Budget

• model revision: 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a;
• V10 data SHA-256: 2461d6f70b44a142854d56c24e1fb42d600065e5788a2c4e055ba47b12696549;
• 26,387 selected rows, data seed 20260802;
• model seed 2026080702 for both QPT1 and B1;
• 256 optimizer updates, 16 logical examples/update;
• maximum sequence length 1,024;
• AdamW, cosine schedule, peak learning rate 2e-4;
• final four text layers, LoRA rank 8, alpha 16;
• QPT1 width 512, 8 source/state slots, 4 query slots, 8 recurrent steps;
• one H100 per arm; batch and accumulation may change only to fit memory while preserving 16 logical examples/update.

A two-update canary may determine memory-fit batch geometry. It is not a capability score. QPT1 and B1 then run once at the frozen 256-update budget.

Development Decision

Evaluate checkpoint 256 identically on the existing GSM8K-100, MATH-100, HumanEval-20, MBPP-20, GPQA-198, BBH-100, and AIME-30 boards. The five-domain macro treats HumanEval plus MBPP as one 40-example code domain.

Promotion requires all of:

• at least +3.0 five-domain macro points over matched B1;
• at least +15 additional solved examples across the five-domain board;
• improvement in at least three domains;
• no domain regression greater than two percentage points;
• finite training, protected-weight hash identity, and no evaluator change;
• after a score pass, packet swap and state reset must materially reduce the gained answers, while release-off must recover the matched LoRA path.

A miss closes exact QPT1 without pointer-temperature, width, duration, seed, loss-weight, or threshold variants. A pass authorizes one larger-data continuation and a fresh benchmark milestone; it does not by itself establish general reasoning.

Result

Jobs 745181/745182 completed the exact 256-update QPT1/B1 comparison with 619,734 charged targets per arm, 16 logical examples/update, finite training, and unchanged protected Qwen3.5-4B weights. QPT1 trained at 148.691 charged target tok/s. Its final checkpoint and training-report SHA-256 values are:

• checkpoint: 97351d9b572b371ff09a7f675ef4b1893c20b69739727b4a207ec5a3c9813350;
• report: 46c66df57e905947602b63137d2b99901d946913847ea0e2fa667845cf6ffd12.

QPT1 clears the macro, solved-count, and improved-domain requirements, but the 10-point code regression exceeds the frozen two-point maximum. Automatic decision job 745198 therefore records promotion_authorized=false and required_next_step=close_exact_qpt1. Aggregate SHA-256 is 4827808a5d0ec4635e8c72cbdcf23bc6b812f91ffe39b3303f29219854afaada. The exact aggregate is preserved in docs/research/DIVERGE_QPT1_QWEN4B_PRODUCT_GATE_RESULT.json.

This is a strong specialist result, not a broad model promotion. The conditional packet-swap, state-reset, and release-off evaluations were not opened because the score conjunction failed. Exact QPT1 is closed without a rescue variant.