DIVERGE-v0 Neural Component Pilot

Status: the tiny and pooled source compilers are closed. The token-role/source- copy compiler qualifies. The resource-corrected A--G gate is a five-seed broad- promotion negative, while retaining a narrower delayed-recovery result.

Date: 2026-08-05

Purpose

The exact CPU packet proved that DIVERGE can preserve and refine a compact version space without losing coherent worlds. It did not prove that a learned source compiler can identify the fault lines and construct that packet. This pilot isolates four learned operations while retaining syntactic candidate record and option-span scaffolds:

1. classify candidate records versus distractors;
2. compile each option into one of four finite ordered programs;
3. compile the deliberately wrong support prior; and
4. bind delayed evidence to one source-sealed option key.

The exact packet, verifier, conflict refinement, execution, and late reader are unchanged. The result is deliberately narrower than unrestricted language compilation or general reasoning.

Tiny compiler result

The from-scratch compiler has 243,319 trainable parameters. It uses local lexical convolutions for record classification, a bidirectional GRU for option program/prior compilation, and a character encoder for exact delayed-evidence binding. Counterfactual wrapper pairs prevent a record position from serving as the candidate label.

Five matched seeds completed 600 updates and 9,600 charged episodes each. Confirmation is perfect for all five seeds, but the development renderer is not stable:

Only two of five seeds satisfy strict joint development and confirmation. The frozen four-of-five-seed requirement therefore fails. The tiny compiler is closed; its perfect confirmation split is not a promotion result.

Diagnosis and one successor

The exact program and delayed-alias fields transfer reliably, while candidate record detection and one seed's prior semantics do not. This pattern is consistent with inadequate lexical semantics in a 243k-parameter compiler, not an exact-packet failure. The one allowed successor changes only the source representation: frozen layer-17 SmolLM2-135M residuals feed a roughly million-parameter projection, two-layer local Transformer, and the same finite heads. The packet runtime, data budget, split, causal release, evaluator, and strict joint metric remain unchanged.

The real-backbone interface smoke completed on Stokes in 28 seconds with a finite update, gradient, report, and checkpoint. The first 600-update H100 seed then completed on evc30 in 2m54s, training 1,068,775 parameters over 9,600 charged episodes at 91.07 episodes/s. It fails decisively:

Development also has 12.694% distractor false positives; confirmation fault-line recall is only 93.164%. Report/checkpoint SHA-256 values are 6b9ecebcf94cb5bbdab62b7d29dddd00bdd6b85f5867628b767da48cd87ff6ca / 246a4600f1ed9c3d8269582be10c56baadb30ed2943c4a11c8cba0726a9fd42f. The remaining four seeds are not launched because seed one misses the strict joint gate by orders of magnitude.

The pooled-residual compiler is closed. Its renderer-dependent inversion (development program order versus confirmation prior/candidate support) calls for an interface redesign, not another tiny width, threshold, wrapper, loss, duration, or seed. The next implementation must preserve token-level semantic roles and source order through constrained copy/closure rather than compressing an entire option into one pooled vector. A read-only fieldwise error audit is permitted before freezing that replacement.

Token-role/source-copy qualification

The replacement preserves the semantic boundary at token resolution. Separate learned roles identify candidate/background cues, favored/reserve support, four finite action identities, and their source order. A constrained decoder copies those roles into one complete option before packet closure. It never averages program fields or consumes the late query. The adapter has 1,013,962 trainable parameters, fewer than the failed pooled compiler.

Five matched seeds each trained for 600 updates / 9,600 charged episodes. Every seed reaches 100% on both the 256-episode development split and the 256-episode held-renderer/held-ontology confirmation split for gold-support recall, exact record set, program, prior, evidence binding, exact packet, DIVERGE answer, and strict joint packet-plus-answer. Distractor false-positive rate is zero. The immediate top-1 and no-conflict diagnostics remain zero as predicted.

Runtime r2 is pinned to private commit e1bdf8b at tar SHA-256 21c9179727d2a721b281b336d6abd18e28ec1ff8bcb22e0fa13170f45c4d8c12. All reports/checkpoints are read-only and hash-verified locally and on Newton. This clears only the learned source-boundary prerequisite. The next gate must implement the complete matched A--G arms, three late-query types, resource receipts, and causal interventions already frozen in the promotion contract.

That matched CPU gate is now complete. Five seeds / 720 episodes / 2,160 late queries give G DIVERGE 100% versus 66.667% for the strongest soft and no- conflict controls. Exact details and the two invalid pre-promotion attempts are recorded in DIVERGE_V0_SOURCE_SEALED_FACTORIZED_CSDC.md. Delayed evidence no longer reuses source aliases: it binds only to option commitments in the sealed packet. The aggregate report SHA-256 is 0d78b271cd4bf4761dde5aa80b929e24ad50fe5057ae5f9cd2c1ea763a8b918d.

Superseding resource correction: that V3 aggregate undercharged DIVERGE's active state groups and therefore underfunded complete-particle B. Exact V4 raises B from 33.333% to 72.222% while G remains 100%; C averages 65.278% and E/F remain 66.667%. V4 fails the frozen >=2x storage advantage at four worlds (1.893x) on all five seeds, although corrected sharing grows to 3.412x at eight worlds and 27.365x at 64. Aggregate SHA-256 is 8e4405920379b7c0a2f4a0c9acc463839a3816b4a87e84f78fcb9d656e17aaab. The broad architecture gate is negative; only the learned compiler qualification and a narrow high-ambiguity delayed-recovery mechanism survive.

Claim boundary

The scores named A--G in these pilots are mechanism diagnostics, not complete parameter/FLOP-matched neural controls. In particular, A--D currently share the compiler's immediate top-1 decision. No result from this component pilot can promote DIVERGE. Promotion still requires the frozen complete controls and resource receipts in DIVERGE_V0_NEURAL_PROMOTION_GATE.json.