# R12 Dual-Rail Lexical COMMAND Preregistration

## Status

Closed negative. Matched H100 job `725519` completed all 1,000 updates in
11:33. Report SHA-256 is
`e693588b0a680aa123f835f8e34fe18968713cb1e51e2926960871ec550c717d`;
SHA256SUMS SHA-256 is
`55120e01aa11ffa2e03a2f9dca16a18e73578c140d35f81a537126c656724202`.

## Diagnosis

The atomic state algebra reconstructs every real-batch target exactly when
given canonical edits. Training the edit selector from exact WORLD state plus
Shohin's frozen layer-19 COMMAND residual nevertheless leaves held-out exact
terminal packets at zero and degrades value transport. The ETTR surfaces use
opaque lexical identities among distractors; forcing every binding through an
undertrained frozen residual is an avoidable information bottleneck.

## Treatment

Retain the exact atomic node, relation, root, and disposition algebra. For each
COMMAND token position, add a trainable projection of Shohin's raw tied token
embedding to the existing projected layer-19 residual before the compiler's
normalization and cross-attention. The base embedding and transformer remain
frozen. The editor therefore receives two synchronized rails:

1. the frozen contextual residual for semantic context; and
2. the frozen raw embedding for exact lexical identity.

No QUERY tokens, answer labels, terminal packet, oracle program, transaction
trace, host solver, verifier, candidate search, or assessor state reaches the
runtime forward. Oracle initial WORLD state is used only for this isolation
training arm. Raw COMMAND tokens are destroyed after edit compilation exactly
as before.

## Matched Gate

Use architecture seed 31, data seed 11, H100 BF16, 1,000 updates, LR `3e-4`,
clip `1.0`, causal-delta weight `1.0`, atomic-action weight `1.0`, 32 held-out
batches, and the same protected checkpoint, release, reader, and joint parent
as job `725517`. The only architectural change is the direct lexical COMMAND
rail. The complete system must remain below 200M parameters.

The treatment is useful only if held-out oracle-initial terminal value
accuracy materially exceeds the residual-only 71.506550% result and movement
is not purchased by structural regression. Nonzero exact terminal packets are
the decisive positive gate. Fully autonomous reasoning cannot be promoted
from this oracle-initial diagnostic. If the treatment fails, reject generic
slot cross-attention and move to explicit token-occurrence/value pointers. If
it crosses, retain it and add the same stage-separated lexical preservation to
a jointly trained WORLD compiler before the fully autonomous gate.

Decision:
`reject_unfiltered_lexical_fusion_and_delete_token_native_transport_cover_before_explicit_occurrence_pointers`.

## Sealed Result

The direct lexical rail does not repair edit binding. On the exact held-out
interface, oracle-initial value accuracy changes from `74.344978%` before fit
to `70.196507%` after fit; autonomous-initial value accuracy changes from
`58.542576%` to `55.376638%`. Exact terminal packets remain `0/512` in both
arms. The fully autonomous source-deleted evaluator remains WORLD strict
`0/80` and COMMAND strict `0/144`, with all margin-1 rates zero. Autonomous
factual top-1 reaches `60.9375%`, which is not causal reasoning because the
counterfactual pairs remain invariant.

Inspection found that fixed-width token-native packing places the semantic
COMMAND AST first and fills the remainder of all 96 model-visible positions
with deterministic public-codebook cover. The ordinary transport attention
mask marks the cover as valid. Direct lexical fusion therefore gives the
editor more exact identity while also forcing it to attend dozens of
nonsemantic codewords. The next treatment removes only that cover using the
public grammar; it does not parse ontology semantics or execute a program.