DIVERGE-PCJ1: Pairwise Counterfactual Judge

Status: frozen before any PCJ1 model score on 2026-08-08.

Question

CVG1 judged each complete solution independently and selected only 35.211% of its holdout, below the stronger QPT1 lineage at 38.850%. Its coherent oracle was 43.545%, so complementary answers existed but the independent scores did not compare the competing explanations reliably.

PCJ1 asks the narrower, structurally different question: can a model that sees the problem and both complete solutions jointly identify enough B1-only wins without sacrificing QPT1 wins?

Frozen Mechanism

The judge input contains only:

1. problem text;
2. complete Candidate A solution;
3. complete Candidate B solution.

It receives no task, benchmark, gold answer, correctness, execution result, or lineage identity. A three-class head predicts A better, tie, or B better from the final hidden state. Training presents deterministic identity-derived A/B orders. Evaluation scores every example in both orders.

Verdicts are mapped back to semantic lineages. PCJ1 commits to B1 only when both orders consistently select B1. Ties, inconsistent orders, and consistent QPT1 verdicts all default to QPT1. There is no tunable threshold.

Immutable Corpus And Split

The source is the completed CVG1 corpus:

• rows: 8,392;
• pair SHA-256: 45f1d66ce5e87dc2a1f4c3594bdde2bae26e9417e879d16eb4eddb228b696afe;
• source report SHA-256: d38e67908ca3f19c81d08b1772c4fd2d0af6d1d113a3a82058f5bc9c5128e4fd;
• content: 4,096 MATH-500-style, 4,096 BBH/science-logic, and 200 independently execution-verified MBPP-style examples.

PCJ1 uses a newly frozen identity partition with seed 2026080811:

Rule: sha256(seed + NUL + identity)[:8] mod 10000; buckets [0,7000), [7000,8500), and [8500,10000) are train, development, and holdout. All splits contain every task and outcome class.

The holdout has 81 B1-only, 208 QPT1-only, 263 both-correct, and 727 both-wrong rows. Thus B1/QPT1/oracle solve 344/471/552 of 1,279.

Receipt correction: the prelaunch audit command that printed the first table escaped the delimiter as two literal characters (\\0) while the frozen rule, implementation, and jobs used one NUL byte (\0). The wrong printed counts were discovered only after immutable reports closed. The corrected table above is reproduced independently by both reports. This is a custody-report error, not a data, seed, model, threshold, or score change.

Predeclared Arms

Two independent one-H100 arms use identical data, updates, labels, head, and gate:

1. exact Qwen3.5-4B B1 host, the scale-matched model-owned control;
2. exact Qwen3.5-9B B1 host, a stronger contemporary judge and explicit system-scale control.

The candidate completions remain the same frozen 4B B1/QPT1 lineages in both arms. A 9B pass is therefore a stronger-host system result, not a claim that the 4B architecture alone acquired the capability.

Frozen optimization: 256 AdamW updates, 8 presentations/update, balanced task/outcome strata, head width 512, context 3,072, LoRA LR 2e-6, head LR 2e-4, seed 2026080812, cosine decay, and no score-dependent retry.

Holdout Gate

An arm passes only if all conditions hold on the new holdout:

• A/B semantic order consistency >= 90%;
• B1 commit rate in [2%, 50%];
• disagreement selection accuracy >= 80%;
• selected exact accuracy at least 2 percentage points above QPT1.

The last condition requires at least 497/1,279 correct, compared with QPT1 at 471/1,279. Failure closes that host arm without threshold, seed, width, duration, renderer, or loss variants.

Only a passing holdout arm may score the already-preserved 568-example development board. Its unchanged product gate requires code >=30/40, at least +3 five-domain macro points and +15 solved answers over the strongest single lineage, at least three improved domains, and no domain regression over two points.

Cost And Successor

Expected charge before launch is 2--4 H100-hours total for both fits and their conditional applications. If both comparators fail, PCJ1 closes and the next structural mechanism is verifier-grounded counterexample revision of one whole lineage, not another classifier or threshold sweep.

Result

Both arms complete all 256 updates with zero train/development/holdout truncations and unchanged protected hosts. The conditional applications fail closed before scoring the 568-row product board.

The 9B arm is a meaningful near-miss but remains below the fixed 497 correct threshold by two answers and far below the 90% order-consistency requirement. Development agrees: 4B/9B select 465/472 versus QPT1 453. No public score is assigned.

Exact 4B/9B report SHA-256 values are 5a02855ff52c70246e08d7cd5a100b510c1b34b2f11fd60e041058f5a13018eb and 6773fb0387c113e7e6a410d93381f1565ff3e1e612163b981c969096abea960e. Judge SHA-256 values are 2c55960b769ab92f6013cfb33ad4b3d55bc1245b97a22e2846c17d64eaf6ad55 and 28fa4728b4a141d3af28fb8648d9c84163e27c1337791a19a85a872fc102651c. Total charged wall time is 0.8717 H100-hours.