Shohin Reasoning Phase Handoff
Status: phase closed on 2026-08-11 after the one authorized VTE1 evaluation. No successor experiment is frozen, queued, or authorized by this handoff.
Surviving qualified architecture
The surviving practical architecture is the immutable dense 9B same-family temporal-revision release:
source -> model-owned draft -> trained same-family revision
-> learned whole-trajectory commit -> one coherent answer
On the protected 538-row product board, unchanged continuation solves 316,
trained revision 374, and learned commitment 383; the coherent oracle ceiling
is 399. The package pins Qwen3.5-9B revision
c202236235762e1c871ad0ccb60c8ee5ba337b9a, release manifest SHA-256
554e841f71edd3a19063411348340e337532db2db05dd5e1e2adc25a3d347e7b,
release SHA256SUMS SHA-256
0dad031312dec0859e35bb7e9daea8aef688ef350b9053f587fba5acdc9c58c5,
and product-result SHA-256
3e86751bb234ee29465885206da5316890060ad8b0b88ea752c4fb012bbf7187.
This is a qualified multi-pass dense system, not a claim about the original
125M checkpoint or universal reasoning.
LAM1 remains the strongest controlled source-to-terminal mechanism: a learned byte compiler, grammar-constrained program selection, and learned recurrent digit microcode solve 3,917/3,917 source-disjoint development programs. It is not holdout-qualified and retains explicit grammar/stack/rational scaffolding.
Final bounded result
VTE1 trained a Qwen3.5-9B final-four rank-8 reviser with a set-valued loss over independently verified KEEP/CONTINUE/RESTART transactions. Its immutable source-disjoint result is:
- semantic correctness:
1285/1566 = 82.0562%, versus KCR1 parent1294/1566 = 82.6309%(-9answers); - per presentation: natural
379/522, continue414/522, keep492/522; - all three states correct:
324/522; - valid transactions:
1566/1566, malformed: zero; - emitted actions: RESTART
1566, KEEP0, CONTINUE0; - exact execution:
69/1566; KEEP byte preservation:0/692; - decode-limit exhaustions:
197, generated tokens:359,269.
VTE1 is a conjunctive failure. Equivalence supervision removed arbitrary label penalties but induced universal regeneration and did not improve semantics. Draft-hidden control, broad development, and holdout were not opened.
Closed lanes and prohibited retries
The exact NDR1, KCR1, and VTE1 natural-draft/transaction family is closed. Do not retry source mix, parser, prompt, delimiter, action ontology, equivalence temperature, candidate family, rank, layer, duration, seed, decoding, or threshold variants. DTMC1, DTC1, CTE1, LTR1, CTF1, and ECTR0 close the current natural-language-to-microcode bridge. DSEO1, DSET1, PSET1, and later edit variants close their exact objectives/interfaces. Small-OLMoE MTR/RCR/ECR/MPR/OBR/DPR routes do not authorize large-MoE scaling. Historical synthetic ETTR lanes remain evidence, not qualified public capability.
Exact VTE1 custody
- KCR1 parent checkpoint:
07e08abe2480782afc77e35031d23bea71a737d019f307066af2bde786dd2ebd - VTE1 training data/report:
cc312363f880e9048622b57cb0cb609acaf92a1ab9ed0552ec8383ea20da1c33/d1c828f4645a15175535d7712805b489019948bf11e0ab2c3899881c4c2705fb - mechanics checkpoint/report:
e74e929cb17ffc8825e538996aba9b06b36ce5dee633d4384e3c9b65e308b370/b91f218f0e9dd4f75522cf9c705cad3aa5f1cbf1b23ea3c1c0891b61996d12e0 - fit checkpoint/report:
8d8f813341aa8eeab3b78cd010908e80fac75ec8d8ad690a18818deb5543381a/efc78413c319160cdbd942c69a07b65243504a1340772bdbdfa0273a9c03b846 - canary data/report:
b9c97c28438b6a87e89c816c177f1a23818564bb16fed0c9551dd07c05d7e139/78fef3045c75a8aa1c61afb172b6e1474f2a4b2888c7e8c2cec1f5fd4b8638a5 - merged candidates/report:
45efcefded3d6a06633c478bf6b265ebd8fd2a44d313cf9db780ca33374d03d3/e47a5f0e8cfcbdad920e02fab1aeabda49e5e12b8ac4103a278b524132070189
Full paths, shard hashes, compute receipts, and frozen gate booleans are in
docs/research/SHOHIN_VTE1_RESULT.json.
Unresolved bottleneck
The system can generate and consume useful complete trajectories on capable dense hosts, and it can execute typed arithmetic programs causally. It still lacks a reliable model-owned transaction planner that preserves correct drafts, repairs semantic faults, and commits efficiently. KCR forced an arbitrary canonical branch; VTE quotiented that ambiguity but learned the easiest objective solution, universal RESTART. The remaining issue is not transaction syntax or verifier quality. It is semantic planning and capability preservation under a causally draft-dependent action.
Compute state and one recommendation
VTE1 evaluation used 10,652.98 aggregate GPU-seconds (2.9592 H100-hours); its
256-update fit used 1,017 Slurm seconds (0.2825 H100-hours). No holdout or
successor compute was consumed. A final squeue -u [redacted user] check returned no
queued or running jobs; the VTE1 data, mechanics, fit, dispatcher, four
evaluation shards, and merge all report COMPLETED in Slurm accounting.
The single recommended next publication gate, if a later phase is authorized, is an unopened cross-family confirmation of the already-qualified immutable 9B temporal-revision release contract on a capable dense host, with the same model-owned draft, trained revision, unchanged-pass control, whole-trajectory commit accounting, and a sealed source-disjoint benchmark. This tests the surviving transferable claim without reopening the failed transaction family.