Shohin Phase 2 Scale Decision

Status: capacity-boundary decision complete, 2026-08-08. The protected 4B product aggregate and frozen 0.8B transfer both closed under their original conjunctive gates. Under the former scratch contingency the selected trunk would be shohin_920m. The project has since redirected to transferable reasoning on pretrained hosts; this result does not authorize a scratch canary or large pretraining launch.

Evidence Boundary

The original 125.08M Shohin completed 300,000 updates and 157.286B nominal token presentations over a 57.826B-token mounted corpus. It is a useful protected baseline, but its public and direct-interaction results do not meet the project objective.

The strongest causal result is now same-family internal draft/revision:

• pinned Qwen3.5-4B draft owner;
• the same 4B base with a learned revision state;
• one complete internal draft followed by one complete revised answer;
• no external proposal model, tool, verifier, answer router, or host solver at inference.

On source-disjoint data, trained revision beats the unchanged second pass by +158/1,289 development answers and +174/1,279 holdout answers. Every math, logic/science, and executable-code delta is positive. On the protected 568-case product board, treatment scores 320/538 = 59.48% and 61.391% five-domain macro versus unchanged-B1 272/538 = 50.56% and 51.053%. Science improves 52->90 and code 6->21, but GSM8K, MATH-500, and BBH logic regress by 2/1/2 answers. The no-domain-regression product gate therefore fails despite the +48 aggregate lift. The 4B artifact is retained as causal evidence and is not packaged as a qualified product.

Candidate Trunks

Both scratch candidates use the existing pre-norm GQA/SwiGLU/QK-norm Shohin trunk, tied embeddings, a 4,096-token training context, and a 49,152-token vocabulary. Parameter counts are exact and exclude no trainable tensor.

The training CLI requires --vocab-size 49152. The selected vocabulary is the Apache-2.0 SmolLM2 49,152-token tokenizer, physical SHA-256 9ca9acddb6525a194ec8ac7a87f24fbba7232a9a15ffa1af0c1224fcd888e47c; it is not a Qwen tokenizer. A hash-bound 500-document comparison across FineWeb-Edu, peS2o, Essential-Web, bounded FinePDF, and code measured 759,779 tokens versus 789,331 for Shohin's historical 32K tokenizer, a 3.744% token reduction overall and 7.19% on code, with zero encode/decode/re-encode mismatches in either arm. The report SHA-256 is 777d0fd6a28153be73bf55370e95a784582edc37a6a7173021d4d2bc6e15ea2e. Tokenizer identity remains an explicit physical launch field rather than being hidden in the size preset.

The deployed reasoning architecture is a shared trunk with two model-owned temporal roles:

1. Draft role: generate one complete source-only trajectory.
2. Revision role: reread source plus the internal trajectory and generate one coherent corrected answer.

The initial implementation may store the two role states as adapters. A later joint checkpoint may package them together, but packaging cannot change the measured inference path.

Scale Selection

The decision is capability-driven rather than parameter-count-driven.

1. Preserve the closed 4B mixed failure without local threshold, prompt, decoding, rank, or duration rescue variants.
2. Run one matched 0.8B internal-draft/revision transfer using the already pinned Qwen3.5-0.8B B1 state, exact source-disjoint bank, exact 256-update revision schedule, and unchanged control.
3. Select shohin_390m only if 0.8B preserves at least a five-point revision gain with nonnegative math, logic/science, and code deltas on both splits.
4. Otherwise select shohin_920m. Do not spend the remaining compute on a scratch scale already contradicted by the capacity-boundary result.

All seventeen frozen 0.8B source-only draft shards completed. Complete runtime r2 replay 746027 reproduced the 8,392-row merged draft bank and built 9,655/1,289/1,279 train/development/holdout rows. The legacy B1 checkpoint encoded a frozen trunk as unfreeze_layers=null; metadata-only migration 746045 normalized this to numeric zero while proving all 62 LoRA tensors and optimizer/state bitwise equal. Training 746046 completed all 256 updates in 13m38s, charging 365,028 tokens at 455.0 tok/s; checkpoint/report SHA-256 values are 540771a3cd6c446c6fb90e225b1b1dc69050152669152761b129d58e1a41b357 and 8a756b50de52545f552754e169914c320df85119f52afc6e5896b9c7e79b466a. Eight trained and eight unchanged-B1 evaluation shards ran under the same evaluator, followed by deterministic split merges and one conjunctive scale comparison. All completed cleanly; the comparison report SHA-256 is 6f42de42dfb78ef77042238308e11d82f1fb748f624ba5babae3216c5c53347f.

Development passes all four frozen conditions. Holdout passes the five-point overall, math, and logic/science conditions but fails nonnegative code by one answer. The conjunctive result is therefore FAIL and selects shohin_920m. The 0.8B mechanism shows real, source-disjoint overall improvement, but its code retention is not robust enough to justify the 389M risk. There is no nearby 0.8B retry family.

This test determined the smallest plausible scratch trunk. It is not a nearby variant of the failed QST1 workspace: the changed factor is the already qualified complete-trajectory revision mechanism.

Data Requirement

Current optimizer-authorized general data are 62.426B historical tokens from FineMath, OpenWebMath, and Python code. This stream is too narrow and too small for the intended scratch run without heavy replay.

The most developed fresh candidates include:

• 8.764B selected FineWeb-Edu tokens;
• 4.350B balanced peS2o technical/science tokens;
• 0.308B Essential-Web reasoning-core tokens;
• bounded FinePDF core and verified math/science/code sources.

They remain candidates until their existing provenance, residualization, privacy, contamination, license, and utility gates close. The production corpus must be mostly fresh broad data rather than repeated historical math and code.

Minimum staged targets:

Repeated presentations are reported separately and never counted as unique corpus size.

Training Sequence

1. Base pretraining: broad next-token training with source-balanced token sampling and immutable evaluation monitors.
2. Reasoning mid-training: interleave verified explanatory math, code, science, logic, and long-form decomposition without answer-only oversampling.
3. General instruction alignment: teach concise direct answers, complete solutions, code boundaries, and refusal/uncertainty behavior.
4. Internal draft/revision training: train both temporal roles on complete trajectories, including initially wrong drafts that can be repaired.
5. Verified-reward training: begin only after the model generates a useful positive rate on independently checked math/code/science tasks; preserve an identical no-RLVR continuation control.
6. Milestone evaluation: greedy pass@1 plus separately reported test-time compute on GSM8K, MATH-500, AIME, executable code, GPQA, and logic, followed by direct interaction and untouched distribution-shifted boards.

Launch Gate

A large run starts only after all of the following are true:

• the 4B product comparison and 0.8B capacity-boundary decision are recorded;
• the selected configuration instantiates at its exact parameter count;
• 1/2/4/8/16-H100 scaling is measured and the efficient geometry selected;
• the physical data contract proves source weights, unique tokens, tokenizer, licenses, contamination scans, and shard hashes;
• exact checkpoint/resume and data-cursor continuation pass;
• a 5B-token matched canary shows stable optimization and no broad regression;
• projected H100-hours are recorded against current quota before launch.

The target is not merely a benchmark-specialized adapter. The target is one deployable Shohin model that retains broad language and code utility while its own draft/revision path materially improves difficult reasoning.