DPR1: Model-Owned Draft Panel Revision

Status: closed negative on the conjunctive information-ceiling gate, 2026-08-09. No capability fit, holdout read, or larger-MoE work is authorized.

Rationale

MPR2's trained owner carries weak useful information but a single aligned draft cannot improve over source-only computation. Its owner+aligned oracle is only 265/1,289, below the required 286 even under perfect selection. DPR1 changes the temporal object from one premature commitment to a coherent panel of eight complete model-owned trajectories. The panel is not averaged fieldwise; every trajectory remains independently readable.

This is a standard test-time-compute family, not a novelty claim. It is the fastest falsifier for whether the pinned OLMoE owner contains enough diverse correct computation to justify training a panel-conditioned reviser.

Frozen ceiling probe

• exact pinned OLMoE and MPR1 trained source-only owner checkpoint used by MPR2;
• exact 1,289-row source-disjoint development board only;
• K=8 complete samples per source in one fixed run;
• sampling: temperature 1.0, top-p 0.95, top-k 20, no thinking, maximum 768 new tokens, seed 2026080924;
• source plus original revision envelope, with the old draft causally hidden through the full model exactly as in the owner;
• no verifier, assessor, candidate correctness, or answer label is visible to generation.

The offline assessor reports fixed-index accuracy, whole-panel oracle, domain oracle, normalized completion diversity, generated tokens, exhaustion, latency, memory, checkpoint hash, and source/runtime hashes.

Frozen gate

A panel-conditioned training campaign is allowed only if all conditions pass:

1. whole-panel oracle >=350/1,289;
2. MATH oracle >=90/623;
3. logic/science oracle >=245/637;
4. executable-code oracle >=15/29;
5. at least 25% of rows contain two or more distinct normalized completions;
6. all eight fixed candidate indices are complete, with no index selected or dropped post hoc, and total exhaustion <=80/10,312;
7. exact source/checkpoint/runtime hashes and complete reports.

If the gate passes, freeze one panel-conditioned reviser with matched shuffled panel and hidden-panel controls before scores. If it fails, close DPR1 without K, temperature, top-p, prompt, seed, or decoding variants. Holdout remains sealed in either case until a later trained development gate passes.

Result

Eight H100 shards generated all 10,312 requested trajectories in 12.3--17.9 minutes. Fixed candidate indices score only 136--162/1,289, all below the trained greedy owner's 247. Whole-panel oracle is nevertheless 542, with MATH 149, logic/science 384, and executable code 9. Every row has at least two distinct normalized completions; 15/10,312 generations exhaust 768 tokens.

Overall, math, logic, diversity, and exhaustion gates pass by wide margins. The code oracle misses its required 15/29, so the conjunctive gate fails and no panel-conditioned fit is authorized. The exact conclusion is that sampled OLMoE computation contains abundant diverse math/logic answers but does not solve the code-capability floor, while selecting a random/fixed trajectory is substantially worse than greedy. Close exact DPR1 without K/sampling variants. Preserve the pre-staged train-panel generator as inactive code. Result: SHOHIN_DPR1_RESULT.json.