DIVERGE-TFS1: Typed Semantic Fault-Line Composition

Status: passed every frozen condition on 2026-08-06; controlled typed mechanism promoted, natural-language/evidence transfer not yet established.

1. Hypothesis

TOL3 qualifies a model-owned local semantic compiler but always commits to one operation. DIVERGE-TFS1 tests the next distinct mechanism:

When source text presents several locally plausible operation anchors, keep each complete operation meaning as one discrete fault-line option, execute all coherent programs through shared typed state groups, and commit only after delayed state evidence makes the requested query invariant.

The source compiler is the immutable TOL3 checkpoint b8b9dfe54b7ab4a31a74739625b8650fa4ee93a41221ab5d82610ebc1c030328. It receives no updates. Fixed verb dictionaries may label its old training snippets but are forbidden from candidate compilation. Each retained option must originate from a distinct positive-margin model-selected source span. Arguments are decoded from the retained document symbol table and the selected span's relations. Incompatible operation fields are never averaged.

2. Frozen board

The one CPU board uses seed 2026080607 and exactly 256 episodes. Each episode contains:

• five fresh named rational registers, one of which is an untouched sentinel;
• twelve binary ambiguous direct-action clauses, interleaved with swaps and guarded updates;
• exactly 4,096 coherent complete programs;
• six gold option-zero and six gold option-one choices, shuffled independently of TOL3 scores;
• one delayed typed state observation after every ambiguous action;
• one sensitive late query, one invariant sentinel query, and one query that remains underdetermined when the final evidence item is withheld; and
• a full gold program whose answer differs from at least one represented world.

Ambiguous source clauses contain two learned anchors separated by /; no new operation word is introduced. Train, TOL1 OOD, TOL3 confirmation, and TFS1 register-name banks are disjoint. Gold option ordering is source order and may not be selected from model confidence. The independent assessor enumerates all 4,096 programs, verifies every trajectory and evidence observation, and proves that full evidence preserves the gold while partial evidence leaves the named underdetermined query non-invariant.

After compilation, raw source bytes and model hidden states are deleted. The sealed packet contains source commitments, the symbol table, unambiguous typed instructions, twelve binary operation fault lines with calibrated TOL3 support, their guarded typed instructions, and provenance. Delayed evidence contains only committed step/register/value observations. It may remove worlds through verified conflict masks but may not create, repair, or rescore one.

3. Matched arms

All arms receive the same compiled options, initial state, operations, query, and evidence budget.

Whole-particle bytes include duplicated source commitment, complete program, state, provenance, support, and evidence receipt. G includes both serialized packet bytes and peak state-group activation bytes. FLOP proxies count every logical and unique typed instruction application; wall time and peak groups are also reported.

Interventions are shuffled evidence between episodes, source commitment swap, initial-state reset, disabled conflicts, operation-option shift, and post-seal source poisoning. Packet/source swaps must reject rather than silently answer.

4. Frozen gates

All conditions are conjunctive:

1. exactly two valid model-derived operation options per fault line and 100% gold support immediately after compilation;
2. exact extensional parity between factorized execution and independent enumeration for every represented assignment;
3. zero verifier-accepted evidence items that remove the gold assignment;
4. at least 243/256 exact sensitive answers for G and on the initially wrong top-1 subset;
5. G beats A and equal-memory B by at least 20 percentage points each;
6. F abstains on at least 243/256 sensitive queries, while invariant queries answer 256/256 and partial-evidence underdetermined queries abstain 256/256;
7. shuffled evidence, state reset, and operation shift each reduce sensitive exactness by at least 50 points; packet swaps reject and post-seal source poisoning is bit-invariant;
8. conservative whole-particle bytes are at least 2x G's static-plus-peak bytes, and logical instruction applications are at least 1.25x unique shared applications; and
9. zero malformed packets, false query commitments, source/provenance mismatches, or overflow.

If equal-memory particles match G within five points, model-derived gold support is absent, or state groups explode enough to miss the resource gate, TFS1 is negative. It receives no seed, width, ambiguity-count, value-range, evidence, renderer, threshold, or checkpoint repair.

5. Claim boundary

A pass would establish that the learned TOL3 source interface can populate a source-sealed factorized epistemic program and that delayed evidence can drive one coherent typed commitment more efficiently than premature top-1 or duplicated particles on this controlled language. The rational executor and evidence verifier remain exact architecture components. A pass is therefore not unrestricted language understanding or general reasoning; it authorizes a learned natural evidence/interface successor, not continuation pretraining.

6. Frozen result

Private implementation commits 5489c6f and f71f7ba bind the exact board, compiler, runtime, controls, accounting, and evaluator. The board contains 256 unique episodes, 3,072 model-owned binary operation fault lines, and exactly 1,048,576 represented complete programs. Every operation-pair family occurs 512 times, and withholding the last evidence receipt leaves exactly two answer-disagreeing worlds in every episode.

The unchanged 28,109-parameter TOL3 checkpoint compiles all 256 programs and all 3,072 fault lines exactly. Both options at every fault line come from a distinct positive-margin learned source span, so gold support is 256/256 before evidence. Exact state-group execution has extensional parity with independent enumeration in all 256 episodes and never deletes the gold assignment.

All destructive controls behave as frozen: shuffled evidence, state reset, and operation shift score 0/256; all 256 packet swaps reject; post-seal source poisoning is bit-invariant. There are zero compiler failures, false query commitments, gold-support deletions, malformed accepted packets, or overflows.

Canonical accounting charges complete particles 8,671,220,064 bytes versus 4,143,311 bytes for the factorized packet plus peak state groups and evidence, a 2,092.824x ratio. The equal-memory control funds exactly one complete particle per episode. Factorized execution performs 7,639,040 logical instruction applications through 9,472 unique state-group applications, an 806.486x sharing ratio; peak state groups are two per episode because each delayed receipt immediately refines the newly split fault line.

Artifact SHA-256 values are:

• board: 818ec5581b4b93180d73939c31efbd39eee805a04b3c8b3628c68d15cf2fc511;
• board report: ddcfe2cfc1e5025e8c8a76a2be7fdd536682a91d507f432a92602d74eaa4e3da;
• evaluation: 6f740db9b89adee048ee1b054da33052e2d4f10448058106bbf26ef1d12eecdc;
• TOL3 checkpoint: b8b9dfe54b7ab4a31a74739625b8650fa4ee93a41221ab5d82610ebc1c030328.

Local and Newton copies match. TFS1 receives no width, seed, duration, evidence, or renderer variants.

7. Decision

TFS1 establishes the controlled conjunction sought by DIVERGE: a learned language interface preserves mutually incompatible complete operation meanings, exact shared execution retains their coherent lineages, and delayed evidence turns an initially wrong top-1 interpretation into a justified late answer. It also shows that soft answer aggregation can be exact once evidence already isolates one world, so accuracy alone is not a factorization claim; the measured advantage is shared storage/execution and correct abstention before commitment.

The claim remains narrow. Ambiguity syntax, state observations, grammar partitioning, exact rational execution, and evidence verification are engineered. The next and only authorized successor is a learned source-bound evidence compiler on natural verified traces, preserving all alternatives before the exact TFS1 runtime. No continuation pretraining is authorized by this result.