MPR2: Bootstrap Draft-Owner Revision

Status: closed negative on development, 2026-08-09. Holdout and larger-MoE transfer remain sealed.

Hypothesis

MPR1 showed that a simple all-layer post-MoE residual improves pinned OLMoE, but its original weak draft is causally harmful: aligned scored 233/1,289 while both shuffled and draft-hidden controls scored 247. MPR2 changes only the quality of the model-owned temporal input. The successful MPR1 source-only model becomes a first-pass draft owner; a fresh, identically sized residual then learns to revise that owner's exact output.

This is not a rank, layer, seed, duration, prompt, or threshold retry. It tests whether temporal revision becomes useful when the preceding model-owned state is stronger.

Frozen system

• host: allenai/OLMoE-1B-7B-0125-Instruct at b89a7c4bc24fb9e55ce2543c9458ce0ca5c4650e;
• draft owner: exact MPR1 hidden/source-only checkpoint at update 256, SHA-256 6cbe9fdcb309513f8ff17e4aa9bc18549db2ccf74681e608cffac0213584610d;
• owner geometry: all 16 sparse layers, rank-18 shared post-MoE residual, 1,179,648 trainables;
• owner generation: greedy, no thinking, at most 768 new tokens, source plus the original revision envelope with the old draft causally unavailable to the full model;
• reviser geometry: the same all-16-layer rank-18 residual, initialized fresh;
• fit: 256 AdamW updates, LR 2e-5, batch 1, accumulation 8, seed 2026080901, data seed 2026080814, and context 4,096.

One canonical owner draft is generated for every unique training source. The existing development owner candidates are reused only after exact checkpoint, data, runtime, and candidate hashes are verified; they scored 247/1,289 with zero 768-token exhaustion.

Matched arms

Every admitted training presentation has the same source, verified target, update budget, target-token charge, and model geometry.

1. Aligned: the exact trained-owner draft for that source.
2. Shuffled: another source's same-task nearest-token-length trained-owner draft.
3. Hidden/source-only: the aligned token IDs and positions, but every draft token key is causally unavailable throughout the full model.
4. Draft-owner direct: the immutable first-pass owner output itself; this is the practical unchanged temporal baseline.

All train and development rows retain complete source, complete draft, and complete target at 4,096. The holdout is neither read nor generated.

Frozen development gate

On the exact 1,289-row source-disjoint development board, every condition is required:

1. aligned revision >=286, exactly 39 answers (+3.0 points) over the trained draft owner at 247;
2. aligned revision at least 13 answers (+1.0 point) above shuffled;
3. aligned revision at least 13 answers above hidden/source-only;
4. domains do not regress below the draft owner: MATH 57, logic/science 182, executable code 8;
5. conservative possible-semantic repairs minus strict breaks relative to the draft owner are at least 13;
6. exact protected hashes, zero frozen router/expert trainables, exact parameter/update/token/FLOP receipts, and complete 4,096-token custody.

A conjunctive development pass opens exactly one sealed holdout with the same margins and nonnegative domain deltas. Only a development-plus-holdout pass authorizes staged larger-MoE transfer. Any development miss closes exact MPR2 without geometry, duration, seed, renderer, prompt, threshold, or decoding variants.

Interpretation

• Pass: meaningful causal temporal revision is operational on small OLMoE; proceed to one sealed holdout and then staged larger-MoE transfer.
• Aligned improves but does not beat controls: stronger generic supervised adaptation remains useful, but temporal draft use is not established.
• Failure: close this bootstrap-owner route and do not scale it.

Result

Eight draft shards generated 5,823 unique trained-owner drafts and 61,263 tokens. Median length was 7 tokens, p95 10, p99 76, with 8/5,823 exhaustion. The matched corpus retained all 9,651 training presentations and all 1,289 development rows, rejected zero, had a maximum complete train sequence of 2,417/4,096, and matched 1,575,873 target tokens per arm. Each fresh fit completed exactly 256 updates and 338,931 charged tokens.

Development scores are aligned 243, shuffled 235, hidden/source-only 248, and direct owner 247. Aligned therefore gains eight over the unrelated draft but loses four to the owner and five to source-only. Domains are 54/181/8 versus owner 57/182/8. There are 18 strict repairs and 22 breaks, for semantic net -4. Every material and causal gate fails except code retention.

The trained draft carries a small source-correlated signal, but a fresh reviser does not convert it into a net capability gain. Exact MPR2 is closed without nearby variants. Result: SHOHIN_MPR2_RESULT.json.