← Complete research archive
Evaluation & auditsResult80 lines

CSDC Lexical Backbone Transfer Result

Decision: pass; advance model-owned SmolLM2 role-copy predictions into the frozen CSDC candidate constructor

docs/research/CSDC_LEXICAL_BACKBONE_TRANSFER_RESULT.mdOpen original Markdown ↗

CSDC Lexical Backbone Transfer Result

Decision: pass; advance model-owned SmolLM2 role-copy predictions into the frozen CSDC candidate constructor

Result

The capability-floor hypothesis passes every frozen condition. On exactly the same semantic questions and labels:

ArmKnown programLexical-OOD programLexical-OOD answerLexical all-four
protected Shohin step-300k100.000%77.344%85.352%221/512
exact SmolLM2-135M-Instruct100.000%95.947%96.582%443/512
shuffled-label SmolLM20.342%0.293%0.391%0/512

SmolLM2 improves exact unseen-lexical programs by 18.604 points over Shohin, above the fixed 15-point margin. Its 95.947% program score, 96.582% answer score, 443 exact quartets, and 0.293% shuffled control all clear the remaining gates.

What changed

The UTF-8 questions, character spans, semantic programs, answers, split identities, and matched twins are identical. Semantic identity SHA-256 is 3ec090657cd0ee90d509d7313147a156ec65b617b13b9c7edb4f32fd47d18cab. Only the frozen backbone and its tokenizer representation differ. Both real arms use the same 8,607,886-parameter ordinary role-copy compiler at layer 19, the same 96,000 examples once, seed, optimizer, labels, and exact executor.

Shohin tokenizes the 96,000 training questions into 9,624,734 tokens; SmolLM2 uses 9,493,338. Length-bucket packing produces 1,517 versus 1,515 optimizer updates. This 0.13% schedule difference is disclosed rather than described as exact update matching. The result isolates the combined practical effect of pretrained backbone plus tokenizer, not backbone weights alone.

Jobs and cost

JobPurposeStateElapsed
739553paired corpus build and retokenizationcompleted3:58 CPU
739556real imported-parent load probecompleted0:08 CPU
739590Shohin role-copy armcompleted9:36 H100
739591SmolLM2 role-copy armcompleted9:41 H100
739592shuffled SmolLM2 armcompleted9:49 H100
739615automatic assessmentcompleted<1s CPU

Total neural allocation is 29:06, or approximately 0.485 H100-hours.

Artifact receipts

  • runtime archive SHA-256: 3946ab3c7bef4303bc876f291f28ad9e99221c3b1f8cd9ddf0703db866c2470f;
  • Shohin/Smol corpus report SHA-256: f102629f2e6c5cec1770ec2ca585b948d1f2052bf8279bd5f1eb194c3261c6a6 / 59b886fd2268cff25b488b5dca98aac01450e4b2532b2a142da6b27886d78098;
  • assessment SHA-256: 6ac70d3d0a1938366342411fab380182ea3c3c9ab5ab6269a9971232ccfa3fa7;
  • Shohin/Smol/shuffled compiler SHA-256: 1439b5499f612cef1cc4ce52eeea27ca931fdfbd27d6c9e3f7bd03856b470567 / abd22528da0d8dc4718c7a89d9c94520540a2f38b7f0b1d9a9e623d0af23cf4d / 16a7b12b5ec499b6c83912b18aa9918197612abc5e23f836e8f10b0544c32c40.

All Newton outputs are read-only. The compact assessment is preserved in the private repository.

Interpretation and next gate

This is a meaningful semantic-interface result: the copied source positions are already exact, while the stronger pretrained representation supplies the previously missing unseen-word polarity. It is not yet CSDC natural-language reasoning, because this board ends in a host list-machine executor.

The authorized successor freezes the trained Smol role-copy compiler and the existing CSDC reasoning core. Model-decoded source roles must construct the complete CSDC hypotheses, source challenges must falsify them, and one lineage must answer the late query. The end-to-end gate must preserve the shuffled- challenge and lineage-swap failures. No confirmation or public claim follows from the compiler result alone.