← Complete research archive
Architecture researchTheory105 lines

CSDC Role-Gated Copy Bridge

The closed semantic bridge identifies records, states, outcomes, and lengths almost perfectly but loses ordered generator words when it compresses each record into one summary vector and regenerates up to twelve operators through independent output heads. CSDC itself remains near…

docs/research/CSDC_ROLE_GATED_COPY_BRIDGE.mdOpen original Markdown ↗

CSDC Role-Gated Copy Bridge

Status: gate passed; promoted controlled rendered-source CSDC baseline.

Hypothesis

The closed semantic bridge identifies records, states, outcomes, and lengths almost perfectly but loses ordered generator words when it compresses each record into one summary vector and regenerates up to twelve operators through independent output heads. CSDC itself remains near exact with typed fields.

The Role-Gated Copy Bridge changes the interface rather than tuning the closed parser. A token encoder predicts four source roles: OTHER, START, OUTCOME, and WORD. At inference it selects challenge records, copies the state tokens at the predicted start/outcome positions, and copies every predicted word token in original source order. It cannot invent or reorder operator identities through a decoder. The complete copied challenges feed the frozen CSDC candidate selector and late executor.

This is not claimed as a novel pointer-network primitive. The architectural test is whether preserving source-token identity at the semantic-to-symbolic boundary removes CSDC's only measured end-to-end failure while retaining the one-coherent-lineage reasoning path.

Fixed controls and information boundary

  • identical seed-47 frozen row-local compiler and CSDC reasoner;
  • identical generated episodes, renderer templates, development cohorts, held-out field order, candidate constructor, evaluator, and interventions;
  • supervision only for source record kind and token roles;
  • no query, answer, table, selected-presentation, terminal-state, or CSDC loss; and
  • copied values come only from model-selected input token positions.

Gate

One seed trains for 1,500 updates at batch 128. It advances only if:

  1. development end-to-end exact answer is at least 95% and within five points of typed oracle CSDC;
  2. development complete challenge tuples and selected presentations are each at least 95%;
  3. shuffled outcomes and lineage swaps each lose at least 20 points;
  4. held-out-renderer exact answer and complete tuples are each at least 90%;
  5. held-out answer stays within five points of typed oracle CSDC; and
  6. every family/depth cohort reaches at least 90% end-to-end exactness.

Failure closes this copy interface without width, duration, seed, template, threshold, or loss variants. Passing establishes controlled rendered-source CSDC, not unrestricted natural-language or public-benchmark reasoning.

Resource envelope

Focused CPU mechanics and a two-update end-to-end smoke precede one H100 job. Hard ceiling is one H100-hour; expected use is below 0.25 H100-hour. No long pretraining or Shohin integration follows from this gate alone.

Result

Immutable job 739448 ran commit 6359453 on one H100 for 1,500 updates, 192,000 examples, and 461.997 training seconds. The copy parser contains 71,622 trainable parameters, 4,290 fewer than the closed summary decoder. The seed-47 row-local compiler and every CSDC reasoning component stayed frozen.

MetricDevelopmentHeld-out renderer
End-to-end exact answer99.593%99.723%
Typed CSDC oracle99.593%99.723%
Complete challenge tuple100.000%100.000%
Complete ordered word100.000%100.000%
Selected presentation99.007%99.284%
Shuffled outcome answer53.630%53.630%
Swapped lineage answer13.623%13.346%

All six cohorts in both splits exceed 98.9% answers; every complete challenge tuple is exact. The learned rendered-source system exactly matches typed-oracle CSDC on both aggregate splits. Shuffling copied challenge outcomes costs 45.963/46.094 points, while swapping the selected whole lineage costs 85.970/86.377 points. The copied source fields and one coherent committed presentation are therefore both causally necessary.

This passes every frozen condition. It establishes a model-owned compiler from controlled rendered source records into CSDC's symbolic challenge interface. The winning principle is:

learn semantic source roles -> preserve token identity and order by copying -> enumerate residual complete hypotheses -> falsify with source evidence -> commit one lineage -> execute the late query.

The result does not establish unrestricted natural-language parsing, external knowledge, or public-benchmark reasoning. The next scaling step must test a broader lexical and compositional language interface while preserving this copy/commit boundary; it must not return to summary-vector sequence regeneration.

Runtime SHA256SUMS SHA-256 is 145c87d760e2c7ee3aee0433c2609ef5849b3d31feb2c7d0c763aeb42dc9afa6. Report SHA-256 is 808f50e6e3a1026761f7fa0e29aa022346bde6befd419e9051e719fd9448ea37; checkpoint SHA-256 is 55b5ef79110625f383f6800ac89a20dba9d0a1420bd554fd928ee70f42fdf956.

Decision: promote_role_gated_copy_csdc_as_the_controlled_rendered_source_baseline_then_test_broader_lexical_semantics_without_changing_the_reasoning_core.