Shohin IDR4 Scale Transfer

Status: exact matched source-disjoint attribution passes on both splits. The protected product confirmation is now permitted and live.

Capability Hypothesis

The 9B IDR1 result established that training a model to revise its own earlier trajectory contributes 130 source-disjoint answers over an untrained second pass and 58 protected product answers. IDR4 asks whether that learned temporal reasoning survives when both owners are the same pinned Qwen3.5-4B family. This is a scale intervention, not a new architecture search.

At inference, one 4B B1 owner generates a complete source-only draft. A later 4B revision owner receives only the original source and that internal draft, then emits one complete final solution. There is no 9B/4B cross-host proposal, external verifier, task router, answer label, or tool.

Frozen Inputs And Compute

• Backbone: Qwen/Qwen3.5-4B@851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a.
• Draft owner: existing 4B B1 checkpoint SHA-256 f7354e6a0c4311ad792b73358b4e62d9dbe0ae1bd2d41896cf55482d9ce81feb.
• Bank: exact IDR1 source-disjoint 4,096 math, 4,096 logic/science, and 200 execution-verified code identities.
• Draft decoding: greedy, one sample, no thinking mode, batch 4, 768 new tokens, seed 2026080818.
• Revision: exact IDR1 verified targets, split, 256-update LoRA geometry, optimizer, context, order, and seeds, warm-started from the 4B B1 adapter.
• Estimated cost: 10--16 H100-hours for drafts and 2--4 for fit plus source-disjoint evaluation. This estimate was declared before launch.

Jobs 745767--745783 are isolated single-H100 draft shards. No downstream fit may start until their merger proves all 8,392 identities exactly once and binds the pinned model revision, B1 checkpoint, bank hashes, decoding fields, and every candidate/report hash.

Draft And Training Result

All 17 draft shards completed cleanly for 16.952 H100-hours. The 4B B1 first pass solves 2,276/8,392 = 27.12%:

Merged draft SHA-256 is 7d7e833de5b646ff18e08780c8d6760fb0d4a5ea8d6b3fd43c61c6d6e4a7c5e4; receipt SHA-256 is d91138a34a294f5edbc472272946e43f46ff11c296fc21028e04d9171f19f7ce. Train/development/holdout data SHA-256 values are a40d209a...ee7b, f7444eef...afa1, and ec99bc35...7a70.

Revision fit 745819 completed all 256 updates in 16m32s, charging 365,028 target tokens at 382.4 target tok/s. Checkpoint SHA-256 is ae3847fe0728b1debcc13049822ea7499f744836b62d6d1c5bcb7c1000d8560b. The trained revision scores 529/1,289 = 41.04% on development and 554/1,279 = 43.32% on holdout. The corrected eight-shard unchanged-B1 development control scores 371/1,289 = 28.78%, establishing a +158-answer / +12.26-point development gain. Domain deltas are all positive: math 140->208, logic/science 223->305, and executable code 8->16.

The corrected holdout control scores 380/1,279 = 29.71%, versus trained revision 554/1,279 = 43.32%: +174 answers and +13.61 absolute points. Holdout domain deltas are math 151->217, logic/science 223->319, and executable code 6->18; every development and holdout domain is positive. The control report SHA-256 is a3f4c824e644b13aa412535c553edb5706305430780d604fa114b074a0a03248.

Two original holdout shards on evc46 failed before model execution because CUDA was unavailable; exact shard replacements 745877/745878 excluded that node and completed cleanly. No examples, weights, prompts, thresholds, decoding fields, or evaluator code changed. Corrected merge 745879 proves all eight shard receipts and complete identity coverage.

Matched Gate

The protected baseline is the original 4B B1 adapter performing the same source-plus-draft second pass with identical prompts, decoding, evaluator, and token budget. The treatment changes only the trained revision LoRA state.

Pass requires all of:

1. trained revision improves overall source-disjoint accuracy by at least five absolute percentage points over the matched untrained second pass;
2. math, logic/science, and executable-code deltas are each nonnegative;
3. the direction reproduces independently on development and holdout;
4. complete custody and generation receipts pass without exclusions.

The pass opens the unchanged protected product board exactly once. Product confirmation uses a new 4B B1 internal draft for every protected identity, then compares the trained 4B revision owner with the unchanged 4B B1 second pass on those identical drafts. It does not reuse 9B product trajectories. Before any product completion was opened, confirmation success was frozen as at least 27 additional correct main-board answers, at least +0.05 five-domain macro accuracy, and nonnegative correct-answer deltas in every one of the five main domains. AIME is reported separately rather than used as a 30-item gate.

Transfer Role

IDR4 is the strongest clean within-Qwen source-disjoint transfer point and a control for the transferable architecture campaign. Together with 0.8B and 9B it defines a scale curve, but it does not select a scratch Shohin trunk. The next missing intervention is model-family transfer under TTR1; the protected 125M checkpoint and former 390M/920M candidates are optional efficiency baselines rather than the primary deployment target.