DIVERGE-NPW1: Narrative Program Weaver

Status: successor contract pre-staged while the sealed SOT1 confirmation runs. It may launch only if SOT1 passes every frozen gate. A SOT1 failure voids this route without opening NPW1 data or training results.

1. Capability hypothesis

SOT1 isolates semantic ownership, but its WORLD owner still receives an engineered line-per-instruction grammar, exact clause partition, and exact relation decoder. NPW1 removes those three scaffolds while retaining every qualified SOT1 component:

A document-level owner can recover an ordered typed event graph from one free-form narrative when local operation and comparison meaning is supplied by the immutable TOL3 semantic anchor. The resulting graph can enter the unchanged SOT1 transaction bus and factorized executor without raw-source access or host repair.

The intervention is one new WORLD_INGRESS owner. TOL3 WORLD_SEMANTICS, NVE1 EVIDENCE, SOT1 QUERY, the epistemic microkernel, and the exact factorized runtime remain hash-identical. NPW1 does not retrain or replace a qualified owner.

2. Model-owned source path

Runtime sees one ASCII narrative and a generic lexical lattice consisting only of word and rational-number spans. It receives no line boundaries, clause records, instruction count, role labels, generator fields, answers, evidence, query, state, or execution feedback.

The ingress owner has:

1. a two-layer bidirectional byte GRU, width 256, over the complete source;
2. pooled contextual states for every generic lexical candidate;
3. one monotone recurrent event cursor with a hard EVENT/STOP decision;
4. complete-event pointer heads for operation anchor, target, operand, predicate operands, true/false branch anchors and operands, and swap roles;
5. a structural DIRECT/SWAP/GUARD/AMBIGUOUS event-form head; and
6. an event commitment head binding event order and every selected source span.

The owner emits one coherent event at a time. It cannot emit fields from different event hypotheses or revise a committed prior event. The immutable TOL3 owner classifies selected local operation anchors and comparator phrases; the ingress owner cannot override those semantic labels. Runtime copies symbol and rational values from committed spans, validates the complete event, and fails closed on missing, repeated, crossing, nonmonotone, or semantically invalid spans. STOP is accepted only after every committed source span lies before the stop cursor.

After all events seal, raw bytes, byte states, candidate states, and cursor states are deleted. The ordinary SOT1 path then consumes natural evidence, executes the factorized support, and reads a late natural query.

3. Frozen data

The verified generator supplies semantic programs and labels only to training and the independent assessor. Surface realization changes presentation, never program semantics.

• training seed: 2026080618;
• training programs: 20,000;
• confirmation seed: 2026080619;
• confirmation programs: 256;
• five symbols, twelve binary fault lines, and 4,096 coherent worlds per confirmation episode;
• body includes direct updates, noncommuting swaps, guarded updates, symbol operands, rational values, and ambiguous operation pairs;
• training and confirmation use disjoint symbol banks, discourse templates, wrappers, connective patterns, grouping patterns, and semantic identities;
• each source is a paragraph with one to three events per sentence and no instruction lines or typed-program delimiters;
• confirmation source, identity, and exact narrative overlap with SOT1 and all NPW1 training data must be zero; and
• no model score may select or reject a confirmation row.

The renderer must preserve exact character spans for assessor supervision but those spans are removed from candidate runtime input. An independent audit reconstructs every semantic event from the stored span ledger and verifies that rendering does not change order or meaning.

4. Frozen optimization

Only WORLD_INGRESS parameters are plastic. The one fit uses seed 2026080620, AdamW, batch 32, learning rate 2e-3 with cosine decay, betas (0.9, 0.95), weight decay 0.01, gradient clipping at 1.0, and exactly 2,000 updates. Loss is the sum of class-balanced event-form, hard monotone cursor, whole-event pointer, commitment, and stop losses. Teacher forcing is permitted during optimization; every scored program uses autonomous recurrent decoding.

All four inherited owner hashes are checked before and after training and after composite reload. Any inherited-owner mutation is fatal.

5. Controls

• exact semantic event ledger and typed SOT1 ceiling;
• unchanged SOT1 controlled-language WORLD owner on the narrative source;
• ingress owner with event order shuffled;
• ingress owner with complete event packets swapped;
• fieldwise role aggregation in place of coherent event selection;
• forced premature STOP;
• source-symbol derangement;
• evidence shuffle, state reset, query-owner swap, and packet-provenance swap;
• post-seal source poison; and
• complete source deletion before evidence, execution, and query.

6. Frozen pass/kill gate

All conditions are conjunctive on the unopened confirmation board:

1. at least 245/256 exact complete source programs and 6,000/6,144 exact ambiguous options;
2. at least 98% exact event count, event order, event form, and every required role family, with zero accepted malformed event;
3. 100% gold fault-line support recall immediately after compilation;
4. protected SOT1 source/evidence/query retention remains above its frozen floors, and inherited owner hashes remain bit-identical;
5. at least 245/256 exact sensitive end-to-end answers, at least 245/256 invariant answers, and at least 245/256 correct underdetermined abstentions;
6. exact extensional parity on every answered episode, zero gold deletion, false commitment, overflow, invalid transaction, or valid-execution rejection;
7. event-order shuffle, packet swap, symbol derangement, evidence shuffle, and state reset each lose at least 50 points; every provenance swap rejects;
8. post-seal source poison changes no packet or answer bit; and
9. parameters, examples, source bytes, updates, FLOPs, peak memory, wall time, transaction counts, storage, and logical/unique state applications are reported.

A miss closes this exact narrative-ingress design without a width, duration, seed, renderer, threshold, optimizer, loss, or decoder retry. A pass qualifies one architecture-aware broad-language training campaign; it remains a verified generated-program result, not open-domain general reasoning.