← Complete research archive
Pretraining & dataTheory111 lines

DIVERGE-PQI1: Pretrained Antisymmetric Query Interpreter

STI1's fresh byte encoder is exact on five query renderers and exactly inverted on the sixth. Its mention-marker deletion has no effect. The architecture has therefore learned renderer polarity without a transferable language semantic.

docs/research/DIVERGE_PQI1_PRETRAINED_QUERY_INTERPRETER.mdOpen original Markdown ↗

DIVERGE-PQI1: Pretrained Antisymmetric Query Interpreter

Status: frozen after STI1 closure and before data materialization, training, or neural evaluation. This is one structurally different semantic-interface gate, not an RRG1/STI1 variant.

Hypothesis

STI1's fresh byte encoder is exact on five query renderers and exactly inverted on the sixth. Its mention-marker deletion has no effect. The architecture has therefore learned renderer polarity without a transferable language semantic.

PQI1 tests whether an already pretrained language residual plus a role- symmetric pointer is the missing substrate:

A frozen pretrained 135M language backbone can represent unseen lexical target/distractor polarity, while one shared antisymmetric candidate scorer prevents independent stage labels or free role slots from selecting an incoherent assignment.

Every source entity is replaced by alpha or beta according only to its model-supplied occurrence group. Frozen layer-19 residuals are pooled over the two mentions and the whole query. One shared function emits candidate scores s0,s1; the only legal assignment logits are (s0-s1, s1-s0). Swapping role columns therefore inverts the assignment exactly by construction. No evidence, execution, answer, renderer ID, or hidden target label enters inference.

Matched arms

Three independent one-H100 arms receive the exact 100,000 counterfactually complete RRG1 QUERY rows, one epoch, example order, batch 128, AdamW schedule, seed, adapter geometry, and direct evaluator:

  1. protected Shohin step-300k backbone;
  2. exact imported SmolLM2-135M-Instruct backbone; and
  3. SmolLM2 with role labels shuffled once before training.

Shohin and SmolLM2 have the same 576-wide, 30-layer, 9-head geometry. Their tokenizers differ and token/compute receipts must be reported. The intervention is the pretrained backbone plus tokenizer substrate, not a claim about weights alone. All backbone weights remain frozen.

Split custody

The opened CCR1 board is development-only. Before any arm trains, a new 256- episode confirmation board is materialized at seed 2026080630 with 32 new entities and six new query renderers. Exact source, query, identity, and symbol overlap with RRG1 training, SRP1, and CCR1 must be zero. Selection is generator- only and records model_score_used=false.

Confirmation remains unread by candidate evaluators until all three development arms and the matched assessor pass. No confirmation template, threshold, seed, or name may change after training.

Development gate

The Smol arm must satisfy every condition:

  • at least 765/768 exact query assignments;
  • every mode at least 254/256 and renderer at least 127/128;
  • at least 64 more exact assignments than the matched Shohin arm;
  • shuffled-label Smol at most 430/768;
  • forced role swap loses at least 500 assignments;
  • context scrub loses at least 250 assignments; and
  • arbitrary entity renaming changes zero assignments and zero logit bits.

Only that conjunction opens confirmation. A development miss closes PQI1 without backbone layer, adapter width, duration, seed, learning-rate, label, renderer, tokenizer, threshold, or normalization variants.

Confirmation and natural-plasticity boundary

Direct confirmation repeats the absolute semantic and causal floors. A pass then permits one zero-retraining composition with protected TOL3 WORLD and NVE1 EVIDENCE. That composite must recover the STI1 end-to-end floors: WORLD and sealing at least 255/256, EVIDENCE at least 3,070/3,072, QUERY at least 765/768, and sensitive/invariant/abstention metrics at least 254/256, with protected hashes exact.

Only a full composite pass may activate the already frozen natural PL1 contract on development data. PQI1 does not authorize continuation pretraining or establish open-domain reasoning by itself.

Result: closed negative

Matched H100 jobs 744506/744507/744508 and assessor 744512 completed the single development gate. Protected Shohin and exact imported SmolLM2 both score 640/768; shuffled SmolLM2 scores 128/768. The two real arms emit the same assignment vector: renderers 0--4 are each 128/128, while renderer 5 is 0/128. Forced role-slot swap makes renderer 5 exact and all others zero. Context scrub falls to 384/768; arbitrary entity renaming is bit-exact. The stronger pretrained residual therefore does not change the failure through this pooled mention-scoring interface.

The matched assessment SHA-256 is c8b482de5384e8b75796c34436948e26927bde3ef3289deb696322582ab0db48. The three development-result SHA-256 values are e7bdfbfc7c5e077f390de8a4dd870cfc303ae55236d4aa23664632594f09b3bb, 588ad26fa2ee88c3f03c9e06b433c91712f16e963fbe190854191f875fb62e2d, and 7ebb61273220bcf4b67e18cf49985c65fab73748fb81be68898a84a0d9a6f7a7. Confirmation stayed sealed.

Read-only autoregressive attribution jobs 744542/744543 score both raw backbones at 384/768, with renderers 0/2/4 exact and 1/3/5 zero. Their result hashes are fc53ad879ecd6a42c7906509879ebc47fca7105dd34b9b24d2d702a080d87f5e and b92c481ad0efe12d4895f6146d317ad97fb2978264bc2791fbdd947bfc332f1e. Independent Stokes job 767010 regenerated the sealed board byte-for-byte with SHA-256 27f198680cc7bcd7e0203949fe4dee1658fc057fe75302b7e2d43d74321201b8. PQI1 is closed without variants.