← Complete research archive
Architecture researchClosed / no-go203 lines

DIVERGE-HSC1 Hierarchical Structured Compiler

The autonomous source boundary has one sharp asymmetry:

docs/research/DIVERGE_HSC1_HIERARCHICAL_STRUCTURED_COMPILER.mdOpen original Markdown ↗

DIVERGE-HSC1 Hierarchical Structured Compiler

Status: neural seed-one negative; closed without variants

Decision date: 2026-08-05

1. Evidence and hypothesis

The autonomous source boundary has one sharp asymmetry:

  • the option-local token-role compiler is exact on five of five seeds when a complete option is supplied as one source object;
  • SC1 loses lineage by multiplying independent token and pair decisions; and
  • WRA1 has exact record boundaries and bounded cardinality, but permutation- matched parallel slots still fail to bind all fields into one object.

HSC1 tests a distinct interface:

First infer one globally normalized monotonic hierarchy over the raw source. Then infer each complete option through one globally normalized finite-state semantic template. No field is normalized independently, and the runtime never constructs a pair graph, Cartesian proposal set, exchangeable slot, Hungarian assignment, retry set, or beam.

This is not a WRA1 decoder, matching, width, duration, seed, threshold, or loss variant. WRA1 remains closed. HSC1 uses source contiguity and ordered semantic templates rather than unordered object slots.

2. Frozen architecture

One raw source pass

Load the immutable failed-SC1 checkpoint only as a frozen SmolLM2 layer-17 source encoder and frozen record-boundary detector. Encode the raw WORLD once. Pair positive record boundaries in source order and fail closed under the existing nine-record / 108-word limits. No gold record, option, renderer, ontology, query, answer, execution state, or delayed evidence enters runtime.

Globally normalized record parse

For each predicted record segment, three learned gap-score channels identify:

  1. the start of source-ordered option A;
  2. the start of source-ordered option B; and
  3. the start of the nonoperative trailer.

Only triples 0 < a < b < t < record_width are legal. Training uses the exact log-partition over every legal triple. Inference uses exact Viterbi decoding, not three independent argmaxes. The resulting phases are:

HEADER | OPTION_A | OPTION_B | TRAILER

A globally normalized cue-role head chooses exactly one source position and one of fault-line or background inside HEADER.

Globally normalized option object

The two predicted option spans are source-ordered, not exchangeable. A shared two-layer local encoder reads each span's already-computed raw-source states. It emits token-role scores for OTHER, alias begin/inside, favored/reserve, and four action identities.

The candidate runtime enumerates only a fixed semantic grammar:

  • four possible component orders;
  • prior class in two values;
  • program class in four values; and
  • alias length from one to four words.

For each finite template, an exact dynamic program sums or maximizes over all legal monotonic source alignments. Alias continuation positions must be physically consecutive. The training loss is one global log-partition over all templates and alignments minus the gold complete-object path. Viterbi returns one coherent alias/prior/program/witness object. Unselected words contribute a shared OTHER baseline. There is no independently normalized alias pointer, prior pointer, action pointer, or class head.

The record and two options are validated, canonicalized by source position, committed into the existing DIVERGE packet, and then all raw source bytes, residuals, and KV state are deleted.

3. Frozen training charge

The parent, tokenizer, source layer, source encoder, record-boundary detector, batch size, total update count, and optimizer family remain unchanged.

  • stage A: 200 updates / 1,600 episodes train only record cut and cue heads;
  • freeze stage-A weights permanently;
  • stage B: 1,000 updates / 8,000 new episodes train only the shared local option encoder and globally normalized semantic-template scores;
  • stage B receives only stage-A Viterbi option spans; records whose predicted cuts do not equal the supervised record are charged and reported but provide no hidden gold-span fallback;
  • total: 1,200 updates / exactly 9,600 charged episodes;
  • 256 fresh episodes each for train, lexical, renderer, and composition cohorts; and
  • one seed first, four more only if seed one clears every floor.

4. CPU mechanics gate

Before neural execution:

  1. cut log-partition and Viterbi must match exhaustive triple enumeration;
  2. template log-partition and Viterbi must match exhaustive path enumeration;
  3. calibrated raw-source scores must reconstruct 1,000/1,000 complete packets;
  4. semantic-role and cut-channel shuffles must each lose at least 20 points;
  5. malformed/nonmonotonic phases and overlapping witnesses must fail closed;
  6. post-seal source poisoning must change no packet semantics; and
  7. accounting must remain linear in source words plus the fixed 128-template grammar, with no pair matrix, complete-particle materialization, retry, or answer-guided path.

5. Neural pass/kill contract

Seed one passes only if every cohort reaches:

  • record segmentation >=99%;
  • hierarchical option-phase parsing >=99%;
  • gold-support recall >=95%;
  • exact sealed packets >=95%;
  • zero overflow and zero accepted invalid/overlapping objects;
  • post-seal source-poison invariance exactly 100%;
  • at least 20 points exact collapse under semantic-template lineage shuffle; and
  • at least 15 points exact advantage over closed SC1 and WRA1 on every shift.

Five-seed promotion requires every seed to beat both closed compilers and at least four of five to clear every absolute floor. Failure of seed one closes HSC1 without a threshold, stage-length, width, source-layer, template, renderer, seed, optimizer, or loss variant.

6. Claim boundary

Linear-chain CRFs, semi-Markov parsing, finite-state templates, Viterbi, structured prediction, and source copying are established methods. HSC1 is not the architecture novelty claim. It is one bounded attempt to provide DIVERGE's source-sealed factorized epistemic state with an autonomous compiler that preserves complete-object lineage. A pass authorizes exactly one unchanged >=8-world DIVERGE recovery gate. It does not reopen the failed broad resource claim or authorize continuation pretraining.

7. Exact CPU result

The frozen seed-202608056800 board passes all gates on 1,000 episodes / 5,904 records across the four cohorts:

  • complete packet reconstruction: 1000/1000;
  • cut-channel shuffle exactness: 0/1000;
  • semantic-role shuffle exactness: 0/1000;
  • malformed option-width fail-closed: 1000/1000;
  • post-seal source-poison invariance: 1000/1000;
  • overflow: 0/1000; and
  • exact linear score/object accounting: 1000/1000.

The receipt charges 280,803 source words, 195,251 option words, 798,297 cut cells, 798,297 cue cells, 2,147,761 semantic-role cells, and exactly 1,511,424 fixed-template evaluations. It allocates no pair matrix. The cut and option dynamic programs also match independent exhaustive references; the differentiable implementations match the CPU partitions and have finite gradients.

Canonical report digest is 23fa3a02f4299a2ff2b29dde415e8fcefc7bc8ecd2a56877cdee68d446c81809; stored JSON SHA-256 is ef795c698a9ab4195eba43a6d8f84b46e55528aad66f708893de59dfd22dbad2. This authorizes only the frozen two-update real-checkpoint smoke and, if that is mechanically clean, one 1,200-update neural seed.

8. Neural result and decision

The portable immutable runtime is commit fce7efb, archive SHA-256 317e85f1182ba6bb63d7bcf2e1256a751b59d487a26e0118aa89d9f71b4e2c0e, and manifest SHA-256 b738e6f31b4bee91de3b033307027228b9801dfd83563ec70f52f9fb38f6afd0. Smoke job 742764 reaches 100% stage-A phase parsing and exercises one exact stage-B update over all 96 predicted option spans without fallback. Full job 742775 completes the frozen 1,200 updates / 9,600 episodes in 8m09s. The 893,393 trainable parameters peak at 759,383,040 allocated GPU bytes. Stage-B loss falls from 9.47937 to 4.849e-5.

The autonomous result fails the gate:

CohortSegmentationPhase parseGold supportExact packet
Train100%95.313%96.875%96.094%
Lexical shift100%97.656%47.266%8.594%
Renderer shift100%99.219%96.094%0%
Composition shift100%99.609%11.328%0%

Overflow and accepted overlap remain zero, and source-poison invariance stays 100%. Those mechanical passes cannot override first-pass support loss and zero shifted packet exactness. Report/checkpoint SHA-256 values are 62ac144d66818e32ba261fade1ac9103d5adbe962d3572004f3a249ba94c56ad / 34c7eaee885ba5201e6e07335add1737b7b7d26b2709861b7967e0b97be64a05.

HSC1 is closed after seed one. Do not run another seed, width, duration, source layer, stage allocation, optimizer, loss, cue repair, template repair, or threshold. The result localizes the remaining failure to shifted language grounding rather than exact structured decoding. It does not authorize a DIVERGE composition, continuation pretraining, or reasoning claim.