← Complete research archive
Functor compilerTheory227 lines

CSDC SmolLM2 Whole-Mention Span-Quotient Gate

Purpose: resolve exactly one remaining interface question before DIVERGE: can a frozen pretrained language backbone identify and preserve an entire episode-local alias mention, independent of its BPE segmentation, well enough to drive the protected CSDC hypothesis selector?

docs/research/CSDC_SMOLLM2_SPAN_QUOTIENT_GATE.mdOpen original Markdown ↗

CSDC SmolLM2 Whole-Mention Span-Quotient Gate

Status: frozen before implementation and neural evaluation

Decision date: 2026-08-05

Purpose: resolve exactly one remaining interface question before DIVERGE: can a frozen pretrained language backbone identify and preserve an entire episode-local alias mention, independent of its BPE segmentation, well enough to drive the protected CSDC hypothesis selector?

This is one pass/kill experiment. No seed, width, update-count, loss-weight, threshold, renderer, or candidate-width repair follows its result.

1. Evidence and narrow hypothesis

The frozen first-subtoken lexical bridge reaches 99.691% development and 95.915% lexical-shift answers. Shuffled outcomes and lineage swaps cause the predicted large failures. It therefore carries causal evidence into CSDC, but it recovers all eight shifted challenge tuples in only 17.920% of episodes. The interface labels and copies only the first BPE token overlapping each alias.

The frozen hypothesis is:

Pooling and classifying each complete model-selected alias span, while sharing evidence across exact-surface occurrences, will preserve lexical identity across unseen aliases and tokenizations and raise shifted all-eight tuple exactness from 17.920% to at least 90% without changing CSDC.

This gate tests occurrence identity, not general coreference, unrestricted language understanding, DIVERGE, or public reasoning.

2. Immutable inputs and budget

The treatment reuses the exact inputs from the first-subtoken bridge:

InputFrozen SHA-256
SmolLM2 parent8196f810a31e0abe7f3bf0eae0a37b103195f109b7a8e962c7b74b5710c98a02
tokenizer9ca9acddb6525a194ec8ac7a87f24fbba7232a9a15ffa1af0c1224fcd888e47c
warm lexical adapterabd22528da0d8dc4718c7a89d9c94520540a2f38b7f0b1d9a9e623d0af23cf4d
frozen CSDC reasonerc374e3b566808cb317ffcd2725653c9073d2e7aebeb75e93ed7ea2a7e2e27044

Training remains seed 2026080507, 1,500 updates, batch 128, 192,000 episodes, AdamW at learning rate 3e-4 and weight decay 0.01. Data, renderer, evaluation, and evaluation-renderer seeds remain 202608051000, 202608052000, 202608053000, and 202608054000. Templates 0--2 and the training alias pools are visible in training. Template 3 and the disjoint shift alias pools remain evaluation-only. Evaluation uses 1,024 episodes in each family-by-depth cohort at depths 8 and 12.

The frozen first-subtoken report b3ae0526e9e28ef21e93f2b32bacd7845f40cdb663d8d4b48d74ed9b7cfc05c5 is the matched control. It is not retrained.

3. Only architectural change

The frozen SmolLM2 layer-19 residual extractor and warm 384-wide, five-layer context encoder remain unchanged. Record-kind prediction remains unchanged. The token role head is replaced by a bounded whole-span quotient head:

  1. Enumerate every nonempty contiguous token span of width one through four wholly inside one rendered record. Enumeration has no alias dictionary, field label, target value, query, or answer.
  2. Represent a span by its first residual, last residual, and mean residual, followed by one learned span projection.
  3. Quotient candidates by exact normalized source bytes within an episode. Mean-pool candidate representations inside each exact-surface class and send that class message back to every occurrence.
  4. Classify each candidate as OTHER, START, OUTCOME, or WORD from its local whole-span representation and its occurrence-class message.
  5. Decode nonoverlapping whole mentions inside each model-selected challenge record. Select one START, one OUTCOME, and up to twelve WORD mentions; order WORD mentions solely by source position.
  6. Convert only an exactly selected annotated source mention to the same episode-local state or generator identity used by the frozen typed CSDC compiler. Partial spans, supersets, overlapping selections, missing fields, duplicate START/OUTCOME, and excess words fail closed.

The maximum span width of four is frozen before the board. The executable must audit that every labeled alias occurrence in both alias pools is representable under that cap before training. Any miss is an infrastructure failure and no neural result may be reported.

The trainable span projection is LayerNorm(3*384) -> Linear(3*384,384) -> GELU. The class projection is LayerNorm(384) -> Linear(384,384) -> GELU. The role head is LayerNorm(768) -> Linear(768,384) -> GELU -> Linear(384,4). The loss is record-kind cross entropy plus span-role cross entropy over candidates inside true challenge records. Span-role class weights are [0.05, 1, 1, 1] for OTHER/START/OUTCOME/WORD. No auxiliary downstream, contrastive, table, answer, or consistency loss is allowed.

The quotient groups exact surface mentions only. It may not merge aliases, infer synonyms, or repair a malformed prediction. Complete particles, version-space state, conflict clauses, and query-time ambiguity are absent.

4. Ownership boundary

Model-owned

  • challenge-record selection;
  • every selected mention start and end;
  • every selected mention role;
  • the number and source order of generator mentions;
  • all challenge tuples consumed by CSDC.

Architectural

  • bounded contiguous-span enumeration;
  • exact source-byte equality among model-selected candidates;
  • permutation-invariant class pooling;
  • nonoverlap, cardinality, source-order, and fail-closed validation;
  • unchanged complete-candidate construction, counterexample scoring, coherent CSDC commitment, tied execution, and late answer scoring.

Forbidden at inference

  • gold spans or role labels;
  • candidate-name or alias dictionaries for proposal or selection;
  • raw query, answer, true table, selected table, or terminal state in the compiler;
  • retries, beam search, host semantic repair, or answer-guided selection;
  • changing candidate width or thresholds after evaluation.

Span/value annotations exist only to supervise and score exact source copying. The deterministic conversion of an exactly copied source mention to its episode-local anonymous identity is the same disclosed copy mechanism used by the protected bridge. This experiment does not claim a learned nominal quotient between different surface strings.

5. Required reports and controls

The report must include, by split and by cohort:

  • answer, oracle-answer, selected-table, all-eight tuple, and all-valid exact;
  • per-field mention exactness for START, OUTCOME, and WORD;
  • exact span-boundary and exact role accuracy over gold mentions;
  • attempted and accepted partial-span, strict-superset, and overlap counts, plus duplicate and missing-field counts;
  • exact-surface class consistency and class-use counts;
  • shuffled-outcome and whole-lineage-swap answer accuracy;
  • parameters, charged episodes, elapsed time, throughput, peak memory, input hashes, source commit, checkpoint hash, and report hash.

Required controls in the one executable are:

  1. unchanged oracle typed challenges;
  2. frozen first-subtoken metrics loaded from the immutable report;
  3. shuffled challenge outcomes;
  4. whole committed-lineage swaps;
  5. class-message-zeroed inference; and
  6. exact-surface class-ID reindexing, which must be bit-identical.

The class-message-zeroed result is diagnostic; it does not need to lose for the treatment to pass. Class-ID reindexing is a hard correctness gate.

6. Frozen pass/kill gate

Every condition must pass:

  1. 100% representability of every labeled alias occurrence under width four;
  2. development answer, all-eight tuple, and selected-table exactness >=95%;
  3. lexical-shift answer, all-eight tuple, and selected-table exactness >=90%;
  4. lexical-shift exact gold-mention boundaries and roles >=90%;
  5. every lexical-shift family-by-depth cohort answer >=90%;
  6. development and lexical-shift oracle answer >=98%;
  7. shuffled-outcome and lineage-swap drops each >=20 points on both splits;
  8. zero accepted partial or strict-superset spans as exact identity copies;
  9. exact-surface class-ID reindexing changes no decoded tuple, selected table, or answer; and
  10. all hashes, seeds, budget, and source-access receipts match this document.

A pass preserves this whole-mention compiler as the language front end for DIVERGE-v0 board construction. A miss closes this exact span architecture and records whole-mention grounding as an unresolved prerequisite. In either case, the next architecture work is DIVERGE-v0 CPU mechanics; this gate is not extended.

7. Ordered successor

After this result is immutable, begin DIVERGE as a source-sealed factorized extension of protected CSDC. PCSD and FCPT remain closed. The first DIVERGE deliverable is CPU-only packet semantics, exact enumeration parity, verified nogoods, safe merging, query invariance/abstention, canonical accounting, and fail-closed overflow on the Delayed Disambiguation/Recovery board. No long pretraining or public benchmark run is authorized by this span result.

8. Immutable result

Status: closed negative. Newton job 741299 completed on evc30 with exit code zero in 00:30:02. It trained the frozen 9,495,302-parameter span adapter for exactly 1,500 updates / 192,000 charged episodes. Training took 1,572.116 seconds at 122.128 episodes/s and reported 10,029,806,592 peak GPU bytes. The one hard minibatch at update 1,200 recovered immediately; no OOM, nonfinite value, or runtime failure occurred.

SplitAnswerComplete tupleSelected tableGold mentionOracle
Development99.691%100.000%99.202%100.000%99.691%
Lexical shift84.294%17.920%74.447%83.021%99.463%

All six development cohorts pass. On lexical shift, the weakest answer cohort is 77.832%. Shuffled outcomes reduce shifted answers to 51.807%, and whole- lineage swaps reduce them to 12.500%; the source challenges and coherent CSDC lineage are therefore still causally used. Exact-surface class-ID reindexing is bit-identical. No partial, strict-superset, or overlapping mention is ever accepted as an exact identity copy.

The failure is not tokenizer representability: every gold alias fits under the frozen width-four cap. It is joint role/identity generalization. Across the six shifted cohorts the strict decoder records 16,625 duplicate outcomes, 5,050 duplicate starts, 1,984 missing outcomes, 2,496 missing starts, 116 nonexact identities, and 223 selected partial spans that are correctly rejected. Shifted start/outcome mentions remain about 93.2%/93.4% exact, but ordered WORD mentions fall to 78.1%, so complete challenge records survive in only 17.9% of episodes. Exact-surface occurrence pooling solves neither unseen nominal grounding nor globally coherent role assignment.

The report/checkpoint SHA-256 values are d81a1c9648b10f8afb409116463b3ca8b5084abc472a33cd4922d0e5d17ebcca / a2b16103dcc63d1a1b08ac9e24be23520066b5cd772feb8e679b21e9a315b19b. Both artifacts are read-only and hash-verified on Newton and in local disaster recovery storage. This exact span architecture receives no repair run. DIVERGE-v0 begins from protected typed/role-copy CSDC and treats learned raw-language fault-line compilation as an unresolved later gate.