DIVERGE-SRP1: Semantic Referent Primitive

Status: frozen after the SOT1 result and its one read-only attribution, before SRP1 data materialization, training, or neural result.

1. Capability hypothesis

SOT1 assigned TARGET/DISTRACTOR semantics to lifecycle stages. Its isolated QUERY owner fit all 50,000 training examples but inverted one fresh renderer on every counterfactual row. The already-qualified EVIDENCE symbol head also fails as a direct query reader. The defect is therefore not solved by choosing one existing stage owner.

SRP1 changes the architectural ownership axis:

A semantic role is owned once and reused across stages. EVIDENCE and QUERY may have different transaction schemas, but both must obtain referent roles from one exchange-equivariant REFERENT primitive.

This is not an SOT1 renderer, width, duration, seed, or loss retry. The SOT1 QUERY owner is a frozen baseline only.

2. Architecture

The composite contains three owners:

1. immutable TOL3 WORLD owner;
2. immutable NVE1 numeric-evidence owner for STEP/VALUE only; and
3. one plastic REFERENT owner shared by natural EVIDENCE and natural QUERY.

The REFERENT owner encodes source bytes and pools each complete symbol group. For candidate group i and the other group j, one shared scorer computes

s_i = f(global, h_i, h_j, h_i - h_j, h_i * h_j)
delta = s_0 - s_1

and emits the only two legal complete permutations from delta. Swapping the two candidate groups therefore swaps the prediction by construction. No stage-specific target head, independent field classification, host repair, or fieldwise averaging exists.

The REFERENT byte encoder is initialized from the qualified NVE1 encoder, then trained jointly on the immutable 50,000-row natural-evidence corpus and the immutable 50,000-row natural-query corpus. The NVE1 numeric owner remains a separate frozen copy; REFERENT plasticity cannot change numeric assignments.

3. Frozen fit

• seed: 2026080621;
• updates: 1,000;
• batch: 128 evidence plus 128 query examples per update;
• optimizer: AdamW, LR 3e-3 cosine to zero, betas (0.9, 0.95), weight decay 0.01, gradient clip 1.0;
• loss: equal-weight symbol-role cross entropy for the two stages;
• no answer, state, operation, comparator, or confirmation label enters fit.

Qualified WORLD and numeric-evidence hashes must remain bit-identical. The checkpoint serializes one model and one owner manifest.

4. Fresh board and controls

One 256-episode board at seed 2026080620 uses a disjoint 32-name entity bank, new program identities, and six source-disjoint query compositions built from training-supported semantic primitives. Every query mode crosses every renderer; mode and renderer cannot be confounded. Exact source, query, identity, and entity overlap with SOT1 confirmation is zero.

Matched controls are:

• frozen SOT1 isolated QUERY owner on the same fresh query transactions;
• frozen NVE1 EVIDENCE symbol head on the same transactions;
• typed-query ceiling;
• top-1 and equal-memory complete particles;
• referent-role swap, shuffled evidence, state reset, operation shift, packet/query swap, and post-seal poison.

5. Frozen pass/kill gate

All conditions are conjunctive:

1. WORLD programs at least 250/256 and natural EVIDENCE at least 3,041/3,072;
2. natural QUERY at least 752/768, every mode at least 245/256, and every renderer at least 122/128;
3. sensitive answers at least 245/256, with exact extensional parity and at least 95% exact conditional on initially wrong top-1;
4. no-evidence sensitive abstention, invariant answers, and partial-evidence underdetermined abstention each at least 245/256;
5. QUERY role swap, shuffled evidence, and state reset each lose at least 50 points; packet/query swaps all reject and post-seal poison is invariant;
6. zero invalid transaction, false commitment, malformed packet, gold deletion, overflow, or valid-execution rejection;
7. immutable owner hashes remain unchanged; and
8. SRP1 improves exact query transactions over frozen SOT1 by at least 77/768 (10 points). If frozen SOT1 already passes or trails by less than ten points, semantic sharing has not earned inclusion.

A miss closes this exact semantic-primitive design. Do not run SRP1 width, duration, seed, renderer, optimizer, warm-start, or loss variants. A pass qualifies this referent interface for one broader end-to-end successor; it is not open-domain reasoning or authorization for continuation pretraining.