← Complete research archive
Architecture researchResearch record95 lines

DIVERGE-TOL3: Position-Free Local Semantic Anchors

TOL2 fixes document binding but misses its gate because the unchanged TOL1 clause-global heads make 277 top-level operation mistakes and 89 local guard semantic mistakes. TOL3 keeps the TOL2 symbol table, relation graph, guard partition, canonical bytecode, and exact executor. It…

docs/research/DIVERGE_TOL3_LOCAL_SEMANTIC_ANCHORS.mdOpen original Markdown ↗

DIVERGE-TOL3: Position-Free Local Semantic Anchors

Status: qualified controlled-language front end on 2026-08-06.

TOL2 fixes document binding but misses its gate because the unchanged TOL1 clause-global heads make 277 top-level operation mistakes and 89 local guard semantic mistakes. TOL3 keeps the TOL2 symbol table, relation graph, guard partition, canonical bytecode, and exact executor. It replaces global semantic pooling with one small learned byte-level encoder applied independently to:

  • each candidate source word, classified as NONE or one operation anchor; and
  • the source substring between predicate operands, classified as one of six comparisons.

Supervisor-only training extraction may identify the gold anchor in the TOL1 training board. Runtime receives only source strings and model logits. It may not use the supervisor anchor dictionary, answer labels, generator metadata, or execution search. Position-free scoring makes operation meaning invariant to clause order by construction.

The one frozen fit uses seed 2026080505, width 64, one bidirectional GRU layer, full-batch class-balanced cross entropy over the deduplicated local training snippets, AdamW at 3e-3, and exactly 750 cosine-decayed updates. At runtime, every source word is scored independently. The compiler selects the single word with the largest positive non-NONE logit margin and passes that model-owned span, rather than a fixed verb lookup, to the relational argument decoder. Comparator phrases are extracted structurally between the two typed predicate operands and scored by the same local encoder's separate head.

The opened TOL1 OOD board is development only. TOL3 must reach at least 90% exact semantic programs, answers, exact guard clauses, and rows with every guard exact, with zero accepted malformed packets. A miss kills this local-anchor design. A pass permits one fresh confirmation renderer under the already frozen TOL2 thresholds. No width, duration, or seed variants are authorized.

Development result and confirmation contract

The one fit completes in 6.844 seconds on CPU with 28,109 parameters and reaches 68/68 exact deduplicated snippets. Checkpoint SHA-256 is b8b9dfe54b7ab4a31a74739625b8650fa4ee93a41221ab5d82610ebc1c030328. On the opened 1,024-row TOL1 OOD board, TOL3 reaches 1,024 exact programs and answers, all 2,506 guard clauses exact, and zero invalid rows. Development report SHA-256 is b86187b2bad4f3f953acfc2aa9665ae69fa944404f488dc80b3c17269277b40e.

This pass authorizes one fresh 1,024-row confirmation board at seed 2026080506. It uses disjoint register names, body depth 15--20, recombined direct-action orders, and the previously unused otherwise FALSE; if PREDICATE, then TRUE guard order. The unchanged checkpoint must satisfy the same four 90% exactness conditions and accept no malformed packet. No result-dependent renderer or decoder changes are permitted after the board hash is materialized.

Fresh confirmation result

The board materializes exactly 1,024 programs / 23,063 clauses at depths 15--20. Every row contains guards, swaps, register operands, and rational values. Its register names are disjoint from both prior name banks and its program identity overlap with all 25,536 earlier rows is zero. Board/report SHA-256 values are 36a5fb51f5129294fac4a6ea30cef22c4637d4ae79f633bbee65cec2b5735ed3 / 714eac5590209cb4b71974bb6da80273fe4e44338987eff49a542e451b8772e5.

The unchanged checkpoint passes every condition:

MetricFresh confirmation
exact semantic programs1,024 / 1,024
exact answers1,024 / 1,024
exact guard clauses3,663 / 3,663
rows with every guard exact1,024 / 1,024
exact top-level operations23,063 / 23,063
invalid rows / accepted malformed packets0 / 0
operation-shift answers0 / 1,024
binding-derangement answers21 / 1,024
state-reset answers0 / 1,024

Confirmation evaluation SHA-256 is 2f5b1ca3c08f4b82f8e220713441014178da72a257b8fafb67b0887f28ad5700. All six artifacts are hash-identical in local disaster recovery and immutable Newton path artifacts/reasoning/diverge_tol3_0dfaf53_r1.

TOL3 is promoted only as a controlled typed-language compiler. It has learned local operation/comparator semantics and composes them with document-owned binding across new names, order, and depth. Its exact rational executor and grammar partition remain engineered, so this is not unrestricted language or general reasoning. No TOL3 seed, width, duration, or renderer variant follows. The next gate must combine this source compiler with factorized language- derived fault lines, delayed evidence, and one coherent DIVERGE commitment; single top-1 parsing and full particles are protected controls.