KCR1: Keep/Continue/Restart Causal Revision

Status: closed negative at the source-disjoint transaction canary on 2026-08-10. CPU admission and the complete 512-update fit passed, but all five frozen treatment conditions failed. No trained controls or broad capability evaluation opened, NDR1 remains closed, and holdout remains sealed.

Closed result

The aligned model scores 1,144/1,566 = 73.0524% canonical action accuracy, 1,294/1,566 = 82.6309% executed semantic accuracy, 681/1,566 = 43.4866% exact execution, 673/692 = 97.2543% KEEP byte preservation, and 170/522 = 32.5670% three-state counterfactual consistency. Branch action accuracy is KEEP 673/692, CONTINUE 444/522, and RESTART only 27/352. There are 14 malformed transactions and 21 decode-limit exhaustions. Thus action overall/per-branch, semantic execution, KEEP preservation, and counterfactual consistency all miss their frozen floors. Exact KCR1 closes without controls, broad scoring, holdout, or syntax/rank/layer/update/seed/ decoding/parser/threshold rescue.

The single allowed read-only attribution identifies label nonidentifiability, not a silent executor failure. On canonical RESTART rows the model emits CONTINUE 269 times, KEEP 52 times, RESTART 27 times, and malformed output four times. Despite the wrong canonical action, 210 of those rows are semantically correct. Across all rows, 244 wrong-action transactions still execute to a correct answer, while 94 correct-action transactions are semantically wrong. This does not rescue KCR1, but it motivates a structurally different set-valued successor in which every verifier-correct coherent transaction is a valid training target instead of forcing one arbitrary branch label.

canary result SHA-256: 198bc6985f7108553190afed9b3e17ce34e2f7a3d2b1424ec65cdcda7a4b81c7
attribution result SHA-256: 39ded1e7a841e43d525d50acb0326fd5d7031c045044b343f3a4cd2c9dc78220
local files: SHOHIN_KCR1_CANARY_RESULT.json and SHOHIN_KCR1_CANARY_ATTRIBUTION_RESULT.json

CPU admission result

CPU job 750637 admitted 11,218/11,220 = 99.982% source identities and rejected only two sources lacking a safe semantic continuation boundary. It created 33,654 rows with all three presentations retained per source and 33,654/33,654 exact transaction round-trips. The action distribution is 16,396 KEEP, 11,218 CONTINUE, and 6,040 RESTART. The actual pinned Qwen tokenizer reports maximum 2,104 prompt tokens and 1,490 target tokens, zero 4,096-token truncation, and 3,416,424 charged target tokens. Holdout was not used and the only runtime-visible field is question.

train SHA-256: 3e63c4e248a32c9f4008a1c245ea6e18bcb2f8497058b6c0c237ca5a719863b9
report SHA-256: f6236490f2e435c841267c849460183db6fbf90fa30f6d1a61b3bd4a64b79b8e
local result: SHOHIN_KCR1_DATA_RESULT.json

The implemented training path preserves the action field through reservoir selection, proves that each tokenized response begins with its declared action, and applies the frozen loss exactly: KEEP averages action plus terminal EOS; payload transactions assign one half of presentation loss to action-marker tokens and one half to newline, payload, and terminal EOS. This path is opt-in; ordinary historical language CE is unchanged. One hash-bound update must pass before any 512-update fit.

Evidence that selects the changed mechanism

NDR1 trained ordinary full-trajectory CE on aligned natural drafts and a same-domain nearest-length shuffled control. Aligned scored 306/1,289, shuffled 343, and unchanged 340; aligned also exhausted 879 generations versus shuffled 768. Read-only identity-level attribution binds all 1,289 immutable rows and finds:

only 1,405/11,220 = 12.52% of training drafts were token-exhausted;
910/1,289 = 70.60% of development drafts were token-exhausted;
aligned loses 37 pairwise answers overall;
33 of those 37 answers are lost on wrong, token-exhausted drafts; and
aligned emits 75,804 more tokens than shuffled.

This does not reopen NDR1 or justify an exhaustion-reweighted retry. It selects a different output mechanism: draft termination state must choose a distinct, executed computation branch rather than remain optional prompt text for a full-answer generator.

Architecture

The first owner emits a complete draft d and one deterministic controller bit z recording whether generation stopped normally or reached its token ceiling. The later owner receives source x, exact draft d, and z, then emits one transaction:

<KEEP>

<CONTINUE>
payload

<RESTART>
payload

A generic deterministic transducer executes exactly one branch:

[ T(d,a,p)= \begin{cases} d & a=\mathrm{KEEP}\ d\Vert p & a=\mathrm{CONTINUE}\ p & a=\mathrm{RESTART}. \end{cases} ]

Malformed actions, a payload after KEEP, or an empty payload after CONTINUE/RESTART fail closed. No verifier, answer label, solver, task router, or host repair exists at inference. The cutoff bit is generation-controller state owned by the first pass, not semantic supervision.

This differs materially from NDR1: the action changes output semantics and correct drafts can be preserved byte-for-byte. It also differs from DSET1's last-span synthetic editor: KCR1 handles nonlocal natural failures by choosing between exact preservation, append-only continuation, and complete restart.

Prospective data contract

Use the already-admitted, evaluation-disjoint 11,220-source NDR1 bank and its immutable natural B1 drafts. Each source contributes a balanced three-state episode around one verified response. A conservatively verified natural KEEP may preserve a different but independently correct model-owned trajectory:

KEEP: the complete verified response is the visible draft; transaction target is exactly <KEEP>.
CONTINUE: a deterministic semantic-boundary prefix of the verified response is the visible cutoff draft; transaction target is <CONTINUE> plus the exact remaining suffix.
NATURAL: the exact immutable B1 draft is visible. A conservative deterministic assessor may label a complete verified non-code draft KEEP; all unverified, code, or exhausted natural drafts receive RESTART plus the full verified response. No natural draft may be relabeled CONTINUE without an independently verified prefix certificate.

The builder must preserve all natural drafts, report every action by domain and draft termination state, and prove exact transaction execution. The complete prompt plus transaction must fit 4,096 Qwen tokens without truncation. A source is admitted only if all three presentations fit; no presentation-level filtering is allowed. Development and holdout sources are forbidden by the existing exact and protected word-13-gram boundary.

Model and optimization

pinned Qwen3.5-9B revision c202236235762e1c871ad0ccb60c8ee5ba337b9a;
immutable B1 update-256 checkpoint SHA-256 854a7cc44fbc2b54418f4e5bd09b7efeed0da44fc9ce217b0bb6b1997b722971;
final-four rank-8 revision LoRA, exactly 2,704,896 trainables;
pair-balanced source episodes, 512 updates, batch 1, accumulation 8, learning rate 2e-5, and the NDR1 seeds;
action-token CE and payload-token CE each contribute one half of the per-presentation objective when a payload exists; KEEP uses action CE only;
greedy transaction generation with a separately frozen bounded payload budget.

No GPU fit opens until CPU data/token/execution admission and a one-update gradient mechanics check pass.

Matched arms

KCR aligned: correct source-local draft state and transaction.
Within-source action permutation diagnostic: the three visible drafts remain source-local, but KEEP/CONTINUE/RESTART transactions are cyclically permuted. This is a causal canary, not a capability baseline; its executed targets intentionally differ.
Constant RESTART: identical model, source/drafts, parameters, and target responses, but every presentation executes a complete restart. Updates are adjusted prospectively to match charged target tokens and training FLOPs rather than presentation count.
Draft-hidden: exact prompt/token geometry with draft content causally hidden through every model layer; the cutoff bit remains visible and is reported separately.

The already-closed NDR1 aligned/shuffled arms and the qualified IDR1 reviser remain immutable references, not tunable controls.

Gates

CPU admission

100% exact execution round-trip for all generated transactions;
zero prompt/target truncation at 4,096;
all three presentations retained per admitted source;
exact source and natural-draft coverage receipts;
zero evaluation overlap under the existing exact and protected-13-gram assessor; and
no verifier or correctness field in runtime-visible examples.

Source-disjoint transaction canary

Before broad capability scoring, aligned must achieve at least 95% action accuracy overall and per branch, at least 95% executed-trajectory correctness, at least 99% KEEP byte preservation, and at least 90% counterfactual consistency across each source's three states. Action-permuted and draft-hidden controls must each remain at or below 60% action accuracy. Forced-action intervention must change the executed trajectory exactly as specified.

Capability development

On the frozen 1,289-row IDR1 development board, all conditions are conjunctive:

at least 603/1,289 answers;
at least 13 answers over constant RESTART and draft-hidden controls;
math at least 223, logic/science at least 349, and code at least 17;
at least 13 net repairs on wrong exhausted drafts;
no more than 400 payload exhaustions; and
complete parameter, token, FLOP, memory, latency, action, branch, and hash receipts.

Only a conjunctive development pass opens one sealed holdout. Any miss closes exact KCR1 without action syntax, split rule, loss weight, rank, layer, update, seed, decoding, parser, or threshold variants.

Claim boundary

A pass would establish that explicit model-owned termination state and a causally executed branch transaction improve temporal revision. It would not prove unrestricted reasoning, optimal branch semantics, or a native 125M Shohin mechanism. A failure closes this exact three-branch interface.