Generate one complete internal trajectory
A pinned Qwen3.5-9B B1 owner receives only the source problem and produces a full first attempt.
Architecture / current systemPublic record · 15 August 2026
Same-family temporal reasoning
Draft. Revise.
Commit.
Shohin's current product architecture gives one model family multiple temporally distinct roles: produce a complete internal draft, revise it with a trained later owner, compare the coherent trajectories, and commit one final answer. The mechanism now has a confirmed broad-board result.
Qwen3.5-9B · same-family owners383 / 538 protected problems+67 over matched second pass
01The architectural move
Reasoning can unfold across learned roles
Do not force every solution
into one irreversible pass.
A standard generation commits tokens while it is still deciding what the answer should be. Shohin preserves that first trajectory, gives a later owner the explicit job of revision, and delays the final commitment until both coherent attempts exist.
02The current computation
Model-owned temporal workspace
Draft, revise, compare, commit.
Both generative owners use the same pinned model family with different small adapter states. The commit owner sees complete candidate trajectories, and all inference-time supervision channels used during training are absent.
Train a later owner to transform the attempt
A distinct small LoRA state on the same pinned backbone sees source plus internal draft and emits a complete replacement solution.
Assess coherent trajectories as wholes
A learned commit owner evaluates the original and revised trajectories without candidate provenance, task routing, or correctness labels.
Keep exactly one complete answer
Selection never averages tokens or answer fields. The output is one intact model-generated trajectory.
CAUSAL PATH
source → internal draft → trained revision → coherent comparison → one final trajectory03The strongest confirmed result
Protected product board · seven task families
Learned temporal revision changes what the model can solve.
The matched original-B1 control gets the same internal draft and a second generation pass. Trained revision still adds 58 product answers by itself; commitment raises the final result to 383/538.
- Final macro
- 75.815% +8.552 points
- Final solved
- 383 of 538 problems
- Matched control
- 316 same draft + pass two
- Trained IDR1
- 374 revision alone
- Oracle ceiling
- 399 not deployable
This is the project's first protected broad product pass. It is a post-trained architecture on a Qwen3.5-9B host—not yet a result from the 125M scratch checkpoint and not a claim of frontier-model parity.
04Controlled mechanism foundation
DIVERGE · independently falsified owners
The mechanism laboratory remains intact.
Before the product pass, the project isolated source-deleted state, learned execution, anonymous identity, raw command and value reading, and unseen-law synthesis in controlled worlds. These are mechanism results, not the source of the 9B public benchmark claim.
01CONFIRMED
NPL2 · verified plasticity
85.6104%64 mutable session scalarsNatural verifier transactions drive branch-local learning. After source deletion, transfer exactly matches the typed oracle on all five seeds; the strongest non-oracle reaches 3.9185%.
02CONFIRMED
MZE1 · recurrent executor
75,272 / 75,272400 learned parametersOutcome-only training recovers every transition and composes the learned laws through depths 4, 8, 16, and 32 while preserving NPL2 exactly.
03CONFIRMED
EAL2 · episode laws
40,960 / 40,960397,250-parameter byte GRUNatural before/after evidence identifies eight fresh laws per episode, then source-deleted execution answers every late confirmation query.
04CONFIRMED
NCP1 · command compiler
20,480 / 20,480273,794-parameter CTC pointerRaw variable-length commands compile into ordered episode-local operation pointers without typed symbols, target length, or inference-time alignment.
05CONFIRMED
OQB1 · identity bus
40,960 / 40,960Occurrence quotient + neural attachmentUnseen names and coherent table reindexing remain exact. Breaking occurrence identity or crossing owner bases collapses the result.
06CONFIRMED
SVE1 · value-event reader
30,720 / 30,720472,136-parameter byte GRURaw bytes become complete value-bearing evidence with no numeric-span scanner or integer parser; all 40,960 late answers remain exact.
07CONFIRMED
SNL1 · neural episode laws
40,960 / 40,960Frozen no-retraining compositionThe qualified readers feed a neural law synthesizer without exact support intersection: 1,280/1,280 laws and 20,480/20,480 terminal states across five boards.
08CONFIRMED
OPB1 · operation binding
30,720 / 30,720Source-deleted evidence-to-operation ownerFive source-disjoint seeds preserve exact operation binding, terminal state, and late answers under renaming and reindexing; scrub and decoy controls collapse to zero.
05Two evidence layers
Turn a model's first attempt into editable internal state
Same-family owners separate proposal, correction, and commitment while preserving complete trajectories and inference-time autonomy.
383/538 · 75.815% macroProve causal state, binding, and execution under deletion
Successive gates remove typed commands, lexical identity, numeric spans, host parsing, exact support search, and explicit operation binding while maintaining destructive controls.
Five source-disjoint confirmation boards per ownerWhat is established
- Training the same-family revision owner adds 130 source-disjoint holdout answers.
- The complete architecture adds 67 solved product problems and +8.552 macro points.
- One intact trajectory is selected; candidate order is exactly invariant.
- No tool, task router, answer label, correctness bit, or verifier is used at inference.
- Executable code is retained at 35/40 on the protected board.
- A 32,784-parameter temporal-causal gate improves Qwen3.6-35B-A3B from 111/256 to 143/256 with native router and expert weights frozen.
- DIVERGE separately confirms causal state and compiler mechanisms under deletion.
What remains engineered
- The product system currently uses a Qwen3.5-9B host, not the scratch 125M model.
- The distinct benefit of antisymmetric scoring was not established over its matched control.
- The commit policy still leaves a 16-answer gap to the coherent oracle.
- The current result is not a claim of unrestricted or frontier-level reasoning.
- The controlled DIVERGE system still uses bounded worlds and exact verification.
Evidence before adjectives
The next target is larger sparse-MoE hosts.
The first sparse-MoE source-disjoint screen is positive: 143/256 versus 111/256 unchanged. The next falsifiable step is to test the same owner-to-revision temporal intervention on larger sparse hosts with matched unchanged and self-refinement controls.