DIVERGE-TOL1: Typed Operation Language Gate

Status: completed and failed the one frozen OOD gate on 2026-08-05.

Capability hypothesis

The working FTA1/NTA3 subsystem recognizes closed arithmetic transactions but does not compile variable binding, exact rational values, predicates, or stateful programs. TOL1 tests whether one length-equivariant finite-state source encoder plus a parameter-free typed assignment layer can compile a broader operation language and execute held-out programs exactly.

The target machine stores exact rational registers and supports:

• SET, ADD, SUBTRACT, and MULTIPLY with literal or register operands;
• SWAP between two registers;
• six comparisons and one guarded true/false update; and
• a late QUERY of one register.

The compiler predicts a clause opcode, predicate and branch opcodes, and role scores over source spans. A maximum-weight assignment may enforce only the declared type and arity schema. It may not inspect answers, execute candidate programs, or use generator metadata. Register identity is the quotient of exact repeated source names. All execution uses fractions.Fraction.

Frozen splits

• Training: 24,000 generated programs, depths 4--8, training renderer and register-name banks.
• Development: 512 fresh programs from the training support.
• OOD: 1,024 programs, depths 9--14, disjoint register names, held-out syntax recombinations, and reserved operation adjacencies.
• Every split is generated independently from a fixed seed and written with a content hash. The assessor re-executes serialized typed programs and rejects malformed or non-finite records.

Frozen comparison

The treatment uses maximum-weight typed role assignment. The protected control uses the same checkpoint and logits but independent role argmax with no legal assignment repair. This isolates the structured decoder without changing parameters or training FLOPs. Operation-shift, binding-derangement, state-reset, and query-only interventions test causal dependence on the compiled program.

Pass / kill gate

On the OOD split, all of the following are required:

• at least 97% exact clause opcodes and 95% exact typed instructions;
• at least 80% exact complete programs and 85% exact terminal answers;
• at least 80% answer accuracy on programs containing each of rational operands, register operands, swaps, and guarded branches;
• treatment exceeds unconstrained decoding by at least 20 answer points;
• operation shift, binding derangement, and state reset each lose at least 50 answer points, and query-only remains below 20%; and
• zero executor acceptance of malformed packets.

A failure closes this finite-state compiler at the failed boundary. It does not authorize width, seed, duration, or renderer tuning. A pass authorizes one full-document integration gate and later replacement of exact-name quotienting with learned alias binding. Neither outcome is a claim of general reasoning.

Result

The board contains 24,000 / 512 / 1,024 train, development, and OOD programs, or 263,972 / 5,633 / 16,888 clauses. Train, development, and OOD SHA-256 values are d8b4af0744d3c4232c1de91989a7b6fd4dd3168f45e35c98f9495add1b52b8ba, 790e40984e042f3983d00724089370b4caa2063cab0d6811ab215d80b854a108, and f8f94746c4f5e1cef06204fc39bd1cfe4630ab81b1a15f0f6c6624a309189f70.

H100 job 743668 completed 2,000 updates over 30,896,564 sampled source bytes in 96.879 seconds. The compiler has 515,362 trainable parameters and used 621,024,768 peak allocated GPU bytes. It reaches 512/512 exact development programs and answers. Checkpoint SHA-256 is 6de732edf7158bfacc4a1627b7f2b8fbcb4274482b91fef8c524ee114af7e1b0.

The unopened OOD evaluation 743681 fails:

• clause opcode: 16,609/16,888 = 98.348%;
• exact typed instruction: 11,565/16,888 = 68.481%;
• exact complete program: 0/1,024;
• exact answer: 172/1,024 = 16.797%; and
• raw independent-role answer: 0/1,024.

Evaluation and gate SHA-256 values are d6ab92ca5400d0bc26d867e1a6d95f24b3d6a4985e88254f053567953e7d6c98 and a3f5e4cb868e6bb709d9e1350fdb70d12a5be5c63793c77424fff5ff567b3517.

Failure localization

The operation concept largely transfers; binding does not. Exact instruction rates by opcode are SET 5,502/5,502, ADD 1,247/1,402, SUBTRACT 1,367/1,478, MULTIPLY 2,349/2,515, QUERY 702/1,024, GUARD 403/2,506, and ordered SWAP 0/2,461. Nearly every correctly classified OOD swap reverses its two operands, which is execution-equivalent but exposes an over-strict ordered packet hash. It does not explain the low answer rate. The decisive errors are guards: clause-global pooling selects ordinary words as registers, confuses the moved true branch with the predicate, and reaches only 909/2,506 exact comparator plus true/false action tuples.

TOL1 is closed. No width, seed, duration, or renderer repair is authorized. The justified successor is a document-level anchor-relational compiler that builds one declared-register quotient, restricts all register pointers to that table, predicts action/comparison anchors locally, decodes guard regions independently, and canonicalizes symmetric bytecode. It requires a fresh confirmation renderer after the opened TOL1 OOD board is used for interface development.