CSDC Lexical Backbone Transfer Gate

Status: frozen fail-fast development gate; no CSDC language-integration claim until this gate passes

Capability hypothesis

The promoted CSDC mechanism already performs complete-hypothesis construction, counterexample falsification, one-lineage commitment, and late execution at 99.723% under a held renderer. Its remaining boundary is lexical: the successful compiler receives a small controlled vocabulary whose state and operator tokens preserve identity.

The next question is whether a pretrained language residual can assign unseen directional phrases to the correct copied operator role while preserving the same binding and ordered-copy mechanism. This is tested before building a larger natural-language CSDC front end.

Paired board

One factorized semantic corpus is generated once. A derivation pass retokenizes the exact same UTF-8 questions and character spans for the second tokenizer. It must prove byte-bound source artifacts, exact semantic-payload identity, zero cross-split prompt/name/13-gram overlap, exact dual-executor agreement, and unchanged quartet labels. The model sees only token IDs and the source length mask.

The development splits are:

known-atom compositional transfer;
lexical OOD, where every direction phrase is absent from training but is built from meaningful directional words;
order and binding twins with matched token bags.

Confirmation is not evaluated by this gate.

Matched arms

protected Shohin step-300k, frozen through layer 19;
exact imported HuggingFaceTB/SmolLM2-135M-Instruct backbone, frozen through layer 19;
the SmolLM2 arm with pointer and kind labels shuffled within equal-length buckets.

All arms use the existing ordinary token-role compiler: width 384, eight heads, five source encoder layers, identical optimizer, one pass, example order, batch size, seed, labels, and exact host executor. The backbones share the same 576-wide, 30-layer geometry. Token counts and elapsed compute are reported because the tokenizers differ.

Fixed decision

The lexical substrate passes only if:

both real-label arms reach at least 98% exact programs on known-atom compositional development;
SmolLM2 reaches at least 90% exact lexical-OOD programs and 95% lexical-OOD answers;
SmolLM2 exceeds matched Shohin lexical-OOD exact programs by at least 15 percentage points;
SmolLM2 reaches at least 410/512 all-four lexical-OOD exact groups; and
shuffled SmolLM2 remains at or below 5% exact programs.

A pass authorizes wiring the same copy predictions into the frozen CSDC candidate constructor and testing end-to-end lexical challenge falsification. A miss means pretrained residuals alone do not ground unseen operators. The next interface must provide explicit in-prompt definitions or contrastive operator grounding; it does not authorize more epochs, width, seeds, or a relaxed score on this board.

Resource envelope

Corpus generation and paired retokenization request eight CPU cores, 64 GiB, and at most two hours. The three independent neural arms each request one H100 for at most one hour and are expected to finish training plus both evaluations in roughly 12--20 minutes. The maximum gate is therefore three H100-hours; the expected use is below one H100-hour. No additional seed or duration follows a miss.

Claim boundary

This is a language-substrate and exact-program test. The host still dereferences copied tokens and executes a frozen list machine. Passing does not establish unrestricted natural-language reasoning, public-benchmark ability, or a new backbone result.