← Complete research archive
Architecture researchTheory79 lines

DIVERGE-SAG1: Counterfactual Advantage-Gated Reasoning

QPT1 adds substantial math, science, and logic capability to Qwen3.5-4B, but its residual is globally available and code falls from 30/40 to 26/40. SAG1 tests whether negative transfer can be removed without discarding the reasoning expert:

docs/research/DIVERGE_SAG1_COUNTERFACTUAL_ADVANTAGE_GATE.mdOpen original Markdown ↗

DIVERGE-SAG1: Counterfactual Advantage-Gated Reasoning

Status: frozen before any SAG1 capability score. The two-update CUDA canary passed on 2026-08-07 without opening an evaluation board.

Hypothesis

QPT1 adds substantial math, science, and logic capability to Qwen3.5-4B, but its residual is globally available and code falls from 30/40 to 26/40. SAG1 tests whether negative transfer can be removed without discarding the reasoning expert:

  1. load the qualified B1 adapter and freeze it bit-identically;
  2. train a separate query-pointer transaction expert;
  3. compare frozen-base and expert teacher-forced loss per training example;
  4. train a prompt-only router against the detached counterfactual advantage;
  5. at inference, hard-commit to either the complete base lineage or the complete expert lineage;
  6. when the router abstains, embeddings and generation are exactly B1.

There is no logit averaging, fieldwise state averaging, external solver, answer-label access, or evaluator-time task router. The router sees only the prompt representation. The mechanism differs from QPT1 by adding a protected null lineage and training the commit policy on measured within-example advantage rather than exposing the expert residual to every prompt.

Frozen Development Run

  • host: Qwen/Qwen3.5-4B revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a;
  • protected B1 checkpoint SHA-256: f7354e6a0c4311ad792b73358b4e62d9dbe0ae1bd2d41896cf55482d9ce81feb;
  • V10 data SHA-256: 2461d6f70b44a142854d56c24e1fb42d600065e5788a2c4e055ba47b12696549;
  • 256 updates, 16 examples/update, context 1,024, seed 2026080711;
  • AdamW cosine schedule, peak LR 2e-4;
  • QPT geometry: width 512, 8 source/state slots, 4 query slots, 8 tied recurrent steps, hard-pointer temperature 0.5;
  • router: prompt mean -> 256 hidden -> one logit, bias -2, threshold 0.5;
  • detached advantage margin 0.02;
  • router/risk/sparsity weights 0.20 / 0.50 / 0.01;
  • B1 host and LoRA are frozen; only the expert workspace and router train.

The original qualified B1 is a protected no-harm reference. A separate equal-exposure control starts from the same protected B1 checkpoint and runs another 256 ordinary B1 LoRA updates on the same V10 stream, data order, learning-rate schedule, context, and seed as SAG1. This control is not a new architecture candidate. It exists to prevent the second 256-update stage from being mistaken for a SAG1 gain. It is matched for examples and target-token exposure, but not parameter count or FLOPs; any development pass is therefore a product gate rather than a parameter-matched causal architecture claim.

One two-update canary may verify exact B1 identity, finite dual-path loss, checkpoint reload, and memory. It is not a capability score. A canary miss closes this implementation before the development run.

Development Gate

The existing 538-example board is now a development board because QPT1 used it. The original B1, equal-exposure B1 continuation, and SAG1 are evaluated once with identical prompts, generation settings, and scoring. SAG1 must:

  • retain at least 30/40 executable code answers; on this 40-task board even one lost answer exceeds the frozen two-percentage-point regression limit;
  • exceed the original B1 by at least three five-domain macro points and 15 solved answers, while improving at least three domains;
  • exceed the equal-exposure B1 continuation by at least three five-domain macro points and 15 solved answers, improve at least three domains, and regress no domain by more than two percentage points;
  • leave the frozen B1 parameter hash unchanged;
  • show nontrivial but non-universal expert commitment (5%--95%), measured as the mean over all fixed-interval training-trace samples. Batch size is one, so a single final row is necessarily binary and is not a valid rate.

A development pass authorizes one source-disjoint larger-board confirmation and a transplant to exact pinned Qwen3.5-9B. A miss closes exact SAG1 without router-threshold, margin, loss-weight, width, duration, seed, or task-label variants. Existing QPT1 results and controls remain closed.