CSDC SmolLM2 Lexical Integration Result
Date: 2026-08-05
Commit: 96217cb
Newton job: 739765 on evc26
Decision: exact bridge closed; preserve the end-to-end near-pass and replace
first-subtoken copying with tokenization-invariant model-owned alias grounding.
Result
The frozen SmolLM2 lexical adapter successfully controls CSDC's complete hypothesis selection and late execution. It does not pass the stricter exact interface gate because the combined unseen-template/unseen-alias split does not recover all eight complete challenge tuples per episode.
| Metric | Development | Lexical shift |
|---|---|---|
| Learned exact answers | 99.691% | 95.915% |
| Typed-oracle exact answers | 99.691% | 99.463% |
| Exact selected tables | 99.202% | 93.197% |
| All-eight challenge tuples exact | 100.000% | 17.920% |
| All-eight decoded fields source-valid | 100.000% | 75.472% |
| Shuffled-outcome answers | 52.604% | 53.304% |
| Whole-lineage-swap answers | 13.867% | 13.037% |
| Minimum family/depth answer cohort | 99.414% | 94.824% |
The learned path therefore remains strongly causal. Shuffling decoded
outcomes costs 47.087 / 42.611 points and swapping the selected whole world
costs 85.824 / 82.878 points. Every held family/depth cohort remains above
94.8% exact answers.
The gate still fails exactly as frozen. Twelve of thirteen conditions pass;
shift_tuple_ge_90 fails (17.920%). No threshold, seed, width, duration, or
loss-weight repair is permitted for this exact bridge.
Work and compute
- 1,500 updates
- batch 128
- 192,000 generated source-only episodes
- 8,604,806 trainable adapter/head parameters
- 134,515,008 frozen SmolLM2 backbone parameters
- 817.266 training seconds at 234.930 episodes/s
- 14:54 total one-H100 allocation including both 6,144-row evaluations
- 6,144 examples per split: three families by lengths 8 and 12
The frozen reasoner, CSDC candidate constructor, selector, commit, and late executor were unchanged. The parser received no query, answer, selected-table, terminal-state, or CSDC loss.
Interpretation
The main capability question receives a positive answer: natural-language challenge fields decoded by a real pretrained backbone can select a coherent CSDC world and answer late queries under lexical/syntax shift. This is much stronger than the prior list-machine lexical gate.
The exact interface question receives a negative answer. The decoder assigns roles to one tokenizer position per alias and treats that first subtoken as the copied identity. Under the combined unseen alias/template split, at least one of eight challenge records is often not completely source-valid. CSDC's redundant counterexamples tolerate those missing constraints, so table and answer accuracy stay high while all-eight tuple exactness collapses. This localization is consistent with a tokenization-boundary grounding failure; it does not prove which individual field is responsible because the frozen report records row-level complete tuples rather than post-hoc per-field diagnostics.
Successor
Do not rerun this parser. The next mechanism is a tokenization-invariant alias grounder:
- predict mention boundaries/equivalence independently of semantic role;
- quotient all subtokens of one source mention into one identity-bearing mention state;
- classify
START,OUTCOME, and orderedWORDroles over mention states; - copy one whole mention identity, never a privileged first subtoken;
- keep the same frozen CSDC candidates, falsifier, commit, executor, cohorts, and causal controls;
- compare against this preserved first-subtoken bridge at the same examples and update budget.
This is explicit grounding, not a cosmetic parser variant. Only a pass on exact shifted tuples authorizes lexicalizing observation records and the late query.
Artifacts
- Report SHA-256:
b3ae0526e9e28ef21e93f2b32bacd7845f40cdb663d8d4b48d74ed9b7cfc05c5 - Checkpoint SHA-256:
12a95731e5be263dce96a3bf13c21d3e28b55167fecae70e714228eb3a5bdcec - Immutable runtime archive SHA-256:
4328e07470e474d2eed02da14123d9d07c708ee413bd9e3b8a1012cc876e8d02