← Complete research archive
Theory & no-go resultsTheory90 lines

Shohin IDR08 Capacity Boundary

Qwen3.5-4B proved that a model can improve difficult reasoning by generating one complete internal draft and then using a separately trained state of the same backbone to revise it. IDR08 asks whether that exact mechanism remains causally useful at 0.8B parameters. It is a capaci…

docs/research/SHOHIN_IDR08_CAPACITY_BOUNDARY.mdOpen original Markdown ↗

Shohin IDR08 Capacity Boundary

Status: complete. The frozen 0.8B comparison closed on 2026-08-08 and is now the lower point in the transferable temporal-revision scale curve.

Question

Qwen3.5-4B proved that a model can improve difficult reasoning by generating one complete internal draft and then using a separately trained state of the same backbone to revise it. IDR08 asks whether that exact mechanism remains causally useful at 0.8B parameters. It is a capacity-boundary experiment that selects the scratch Shohin scale, not a new mechanism search.

Frozen Intervention

  • Host: pinned Qwen/Qwen3.5-0.8B revision 2fc06364715b967f1860aea9cf38778875588b17.
  • Warm start: the existing final-four-layer rank-8 B1 checkpoint produced by the matched QST1 campaign; its path and SHA-256 must be recorded before admission.
  • Bank: the exact 8,392 source-disjoint identities used by IDR1/IDR4: 4,096 MATH, 4,096 logic/science, and 200 execution-verified code cases.
  • Draft: one source-only B1 trajectory, greedy, no thinking mode, batch 4, maximum 768 new tokens, seed 2026080818.
  • Revision data: the same verified IDR targets, train/development/holdout split, and source-plus-internal-draft presentation contract as IDR4.
  • Training: warm-start the 0.8B B1 state; 256 AdamW updates; batch 1 with gradient accumulation 8; 4,096-token context; learning rate 2e-5; final four layers; rank 8; alpha 16; seeds 2026080815/2026080814.
  • Control: the unchanged 0.8B B1 state performs the identical second pass over the same question/draft inputs, evaluator, token budget, batches, and seeds.

No external proposal model, verifier, solver, tool, answer router, or task router is present at inference.

Decision Gate

Development and holdout are conjunctive. Each split must independently show:

  1. at least +0.05 absolute overall accuracy for trained revision versus the unchanged second pass;
  2. nonnegative exact-answer delta on MATH;
  3. nonnegative exact-answer delta on logic/science;
  4. nonnegative exact-answer delta on executable code; and
  5. complete identity, model, checkpoint, prompt, generation, and evaluator receipts.

pipeline/compare_idr_scale.py applies this rule once to the four complete merged reports. A pass selects shohin_390m; any miss selects shohin_920m. There is no seed, rank, duration, layer, prompt, decoding, or threshold rescue family.

Cost And Sequencing

Expected cost is approximately 4--8 H100-hours for 17 independent draft shards and 1--3 H100-hours for revision fit plus sharded treatment/control evaluation. Record actual aggregate and critical-path GPU time. Draft shards may run concurrently, but data construction requires complete unique identity coverage and all downstream jobs remain dependency-gated.

The result no longer selects an active scratch program. It constrains the capacity boundary of temporal revision and motivates the cross-family TTR1 gate. No scratch capability canary or large scratch run is authorized.

Result

Development improves 236 -> 323 / 1,289 (+87, +6.749 points), with MATH 44->64, logic/science 190->257, and code 2->2. Holdout improves 242 -> 328 / 1,279 (+86, +6.724 points), with MATH 39->74 and logic/science 194->246, but code moves 9->8. The frozen conjunctive gate therefore fails only holdout code retention. Comparison report SHA-256 is 6f42de42dfb78ef77042238308e11d82f1fb748f624ba5babae3216c5c53347f.

This is positive aggregate causal evidence at 0.8B and an honest retention failure. There is no nearby 0.8B retry.

Launch Receipt

  • Immutable runtime: idr08_28c1246_r1
  • Runtime SHA256SUMS SHA-256: 52e6a0bc4b853bd3db195897d9d811d566d261d19fbee935c8089d26987c0a65
  • B1 checkpoint: baseline_v10_verified_u1000_2461d6f/checkpoint_0001000.pt
  • B1 checkpoint SHA-256: 14d1a2e34fc7c452b4af507d5e9cd39c039e2526fb9c0393999d2a4fabef0e28
  • Draft jobs: 745961--745977
  • Draft-dependent build: 745978
  • Training/evaluation/decision dispatcher: 745979
  • Fresh output root: artifacts/product_reasoning/idr08_0p8b_28c1246_r1