DIVERGE-ULC1 Uncertainty-Lifted Compiler

Status: exact CPU mechanics and frozen-HSC1 matched runtime gate passed; model-owned execution/evidence remains open

Decision date: 2026-08-05

1. Causal thesis

HSC1 does not mainly lack semantic knowledge. On 12,288 fresh complete options, the frozen failed checkpoint ranks the gold semantic template first in 99.316--100% of shifted options and the gold token alignment is Viterbi in 100%. It fails because a single globally normalized parse commits every record to one cue kind. Cue top-1 falls to 13.845% on composition shift, and one wrong record decision invalidates the complete packet.

ULC1 changes the computational object, not the classifier:

A language compiler must not turn local uncertainty into irreversible global deletion. It seals a packed structured support lattice, lifts unresolved source interpretations into guarded epistemic variables, executes each coherent interpretation through shared state groups, and commits only after evidence or query invariance removes the ambiguity.

The frozen rank audit shows that K=2 retains the valid interpretation in 100% of 768 shifted episodes. Across typical six-record episodes, even binary record uncertainty implies 64 coherent worlds. Resource-corrected DIVERGE-v4 already measured a 27.365x storage advantage over duplicated particles at 64 worlds. ULC1 joins those two independently measured facts.

2. Architecture

Score producer

The first matched gate freezes the exact failed HSC1 checkpoint. It may emit record-boundary, phase, cue, and semantic-role potentials, but it may not update weights, inspect delayed evidence, query, answer, renderer ID, ontology ID, or gold source objects. The hard-Viterbi control and ULC1 receive identical score tensors.

A later trainable version, only if this matched gate passes, replaces HSC1's span-conditioned option encoder with one record-wide encoder. This is required so all legal phase paths share one score grid and no gold/Viterbi span is needed to score an alternative.

Sealed parse lattice

For each exact learned record boundary, ULC1 stores:

1. a monotonic three-cut dynamic-program chart;
2. cue-position/kind alternatives with source-occurrence provenance;
3. two option finite-state charts over the fixed 128 semantic templates;
4. contiguous alias-span commitments and physical occurrence IDs;
5. exact learned log support for every retained factor; and
6. a source commitment covering the complete raw record.

No complete parse Cartesian product is materialized. Raw source bytes, residuals, and KV state are deleted after the lattice and occurrence table are sealed. Scores may order support but cannot delete a grammar-valid world in the CPU reference. The matched neural diagnostic may additionally report K=2 support, but the exact reference retains the full bounded grammar.

Uncertainty lift

Every semantic record interpretation maps to a guarded quotient:

• each record owns one categorical BACKGROUND / ACTIVE-left / ACTIVE-right variable, with the first record restricted to the two active alternatives;
• membership and active-option fields are derived from that one category rather than represented as independent axes;
• phase/template/alignment witnesses remain attached to that lineage;
• option programs become guarded state patches; and
• witness variants may merge only when an independent semantic certificate proves identical membership, occurrence identities, programs, support, and future state/query behavior.

Fields from incompatible parses are never averaged. A record's membership from one hypothesis cannot be paired with the program or alias from another.

Shared recurrent execution

Records execute in source order. The exact gate uses the existing typed DIVERGE transactions and groups worlds only by identical complete typed state. Every group carries a disjoint support mask and exact mass. Delayed observations produce verifier-checked nogoods; a nogood may remove support but never create or repair a world. Late queries return an answer only when all surviving worlds agree, otherwise ABSTAIN.

The typed exact executor is an assessor/reference, not the final model-owned reasoning claim. A neural shared reactor is authorized only after the unchanged score/runtime gate proves that uncertainty lifting itself is useful.

3. First discriminating gate

Use fresh raw-source episodes from all four existing cohorts. For each record, the independent assessor supplies the observable effect signature produced by the gold record program on a fixed typed probe state, then removes all source access before execution. The signature never contains alias identity, prior class, parse index, query, or answer. Every candidate independently predicts the same effect signature from its own complete program, and the verifier may remove only candidates whose predicted effect conflicts with the observation. Initial learned top-1 must be wrong in a measured subset; evidence must identify the valid behavioral lineage through state consequences rather than labels.

Matched arms receive the same frozen score tensors and typed transition budget:

Additional interventions are shuffled evidence guards, packet swap between matched episodes, forced premature top-1, state reset before query, and post-seal source poisoning. Full hypothesis fields are never averaged. The underdetermined query is selected late from the sealed typed state as the first register read with certified disagreement across represented worlds. If no such read exists, the episode fails the board contract; a safe invariant answer is never mislabeled as a false commitment.

4. CPU mechanics gate

Before using a neural score tensor, the independent reference must prove on at least 1,000 calibrated episodes:

1. 100% extensional parity with complete world enumeration;
2. 100% valid-world support immediately after compilation;
3. zero verifier-accepted nogoods that remove a valid world;
4. zero false query certificates or commitments;
5. exact monotone support and mass after every refinement;
6. exact source-order noncommuting execution;
7. 100% source-poison invariance after sealing;
8. fail-closed overflow and malformed provenance;
9. canonical byte/transaction accounting; and
10. at least 2x worlds/byte over whole particles at eight or more semantic worlds, reported rather than assumed.

CPU result

The frozen seed-202608057400 reference passes all ten requirements on 1,024 episodes and 16,896 represented worlds. The calibrated support widths are 2, 6, 16, and 42 coherent worlds; a sparse cross-record guard circuit rules out a background-to-ACTIVE-left transition on held-out deeper compositions.

The factorized representation is deliberately not promoted at two or six worlds, where its fixed packet overhead can lose to complete particles. It earns the storage gate at 16 and 42 worlds and shares 64,768 transaction applications. Source-order perturbation changes the answer, overflow discards all partial support, and post-seal source poisoning does not change packet bytes.

The immutable report is artifacts/reasoning/diverge_ulc1/cpu_calibration_seed202608057400.json with SHA-256 4123def2e71041987a14eef28385e6491565c34427fd93cc7fc5247fe09a061b. Its board commitment is 6013ea82c2949ac33b9744964f1144520e8c158ad72fa287496e553072417478. This result authorizes the bounded frozen-HSC1 score/runtime gate only. It is not learned-language, neural-execution, or model-owned reasoning evidence.

5. Neural pass/kill rule

Freeze the score checkpoint, seeds, cohort sizes, K, transactions, evidence, queries, and all controls before reading result accuracy.

ULC1 passes only if:

• every episode represents at least eight coherent worlds;
• gold source support is at least 95% in every cohort;
• final exactness is at least 90% in every cohort;
• wrong-top-1 exact recovery is at least 90% in every cohort;
• final sensitive-query exactness beats hard Viterbi and equal-memory whole particles by at least 10 points in every shifted cohort;
• final exactness beats equal-transaction single-state recurrence and soft aggregation by at least 10 points in every shifted cohort;
• invariant queries are 100% and underdetermined queries never falsely commit;
• provenance and state-reset interventions each lose at least 20 points;
• packet swaps are rejected and post-seal source poisoning is bit-invariant;
• verifier-derived guards never remove a represented gold world;
• conservative whole-particle component bytes are at least 2x the factorized packet bytes in every shifted cohort; and
• duplicated transaction applications are at least 1.25x unique shared applications in every shifted cohort.

If K=2 full particles equal or beat ULC1 at matched memory/FLOPs, the sharing claim fails. If support is absent before evidence, the compiler fails. If evidence helps only through an exact host operation unavailable to the future neural reactor, retain the result as mechanics only and do not promote it as model-owned reasoning.

Frozen-HSC1 runtime result

The corrected immutable job 743195 completes on evc38 in 00:02:44 from runtime commit bf049df. It evaluates 64 fresh episodes in each of four cohorts with no weight updates. The accepted report SHA-256 is b538565639e89f34ec6aa969e22a0017d00a4b1f3ef70e005ca727d8f2c2faa2.

Gold semantic support, represented-gold preservation, invariant answers, certified underdetermined abstention, packet-swap rejection, source-poison invariance, and gold-state execution are all 100%. Shuffled evidence guards score 0--3.125%; state reset scores 9.375--28.125%. All frozen gates pass.

The packed support is substantially wider than the calibrated CPU board: per episode it ranges from 125 to 25,930,800 worlds. Exact sharing keeps the MDD to 61--4,692 expression nodes and at most 224 simultaneous typed-state groups. The matched particle arm stores only 6--71 complete paths under the same byte and transaction budgets and never contains a sufficient surviving lineage. Aggregate conservative whole-particle/factorized byte ratios range from 23,098.7x to 74,490.7x; duplicated/unique transaction ratios range from 676.1x to 2,272.7x. These large ratios are specific to this highly redundant synthetic grammar and are not a general language-efficiency claim.

Job 743159 used the same scores and runtime but mislabeled one invariant slot-zero query as underdetermined. Its report is preserved under SHA-256 f90b14f91a3f626d31aaba74c0dd92b54d6a9d8fdb830b569141e6cdbef08f71 and explicitly invalidated. Schema v2 selects an underdetermined query only after certifying cross-world disagreement in the sealed typed state; no model, score, evidence, sensitive answer, arm, budget, or threshold changed.

This is the first learned-source DIVERGE mechanism pass, but it is not native neural reasoning. HSC1 supplies frozen learned source scores; the typed host executes candidate programs; and an independent assessor supplies observable program-effect evidence. The next gate must learn those evidence and execution interfaces while retaining whole-hypothesis coherence and the same sealed support behavior.

6. Claim boundary

Parse forests, CRFs, packed charts, version spaces, belief propagation, particle filters, delayed evidence, and guarded execution are established ideas. The candidate contribution is narrower: source-sealed lifting of language-parse uncertainty into coherent DIVERGE state variables, shared noncommuting execution, verified contradiction refinement, and query-invariant commitment. The rank audit and any CPU pass do not by themselves establish native or general reasoning.