PSET1: Two-Stream Pointer Edit Transducer

Status: frozen prospective development canary, 2026-08-09. DSET1-v0 is closed. No PSET1 model output exists at freeze time. Holdout remains unopened.

Hypothesis

DSET1 proves causal draft use but loses reliability while autoregressively serializing an old surface and a replacement surface. PSET1 instead treats revision as sequence transduction. The source and draft have separate causal streams. A model-owned policy emits an operation, draft-character pointers, and at most sixteen UTF-8 replacement bytes. A generic deterministic executor copies every untouched draft byte and applies the predicted splice. It has no semantic knowledge, verifier, solver, answer label, task route, or repair rule.

For a Unicode draft D=(c_0,...,c_n), PSET1 predicts

[ a\in{KEEP,REPLACE},\quad i,j\in[0,n),\quad r_{1:m},\ m\le 16. ]

KEEP returns the exact draft bytes. REPLACE requires valid character boundaries and returns draft[:i] + utf8_decode(r) + draft[j+1:]. Malformed pointers, absent offsets, overlong replacement, or decode mismatch fail closed. The final trajectory is therefore causally impossible without the draft: all but the bounded edit are copied through model-owned pointers.

Encoder and edit head

• pinned OLMoE host and frozen DSET1 aligned update-512 stage owner;
• all host weights, routers, experts, and the DSET residual frozen;
• source problem and draft encoded in separate frozen forward passes;
• shared 256-wide projections followed by one draft-query/source-key cross-attention block and feed-forward block;
• action and start/end pointer heads over character states obtained by broadcasting each contextual draft-token state to its tokenizer-reported character span and adding a learned hashed-character embedding;
• a two-layer 256-wide autoregressive replacement decoder cross-attending to source states and the selected draft state;
• replacement input/output is a compact 257-symbol byte alphabet (256 bytes plus EOS), avoiding BPE-boundary ambiguity;
• no full-response autoregressive decoder.

The aligned and label-permuted heads have identical initialization, parameter count, source/pair order, updates, optimizer, and token geometry. The label-permuted arm trains clean drafts to restore the known fault and fault drafts to KEEP. Draft-hidden and same-family near-length shuffled-draft interventions are evaluation-only causal falsifiers of the aligned head.

Data and custody

Stage 0 reuses only the already-opened DSET1 development lineage:

• 4,096 DSET1 train identities and 256 DSET1 diagnostic identities, selected deterministically at the frozen 7:1 numeric/choice ratio;
• exact train/diagnostic source disjointness inherited and rechecked;
• tokenizer offsets must cover every draft character exactly once;
• replacement UTF-8 bytes must decode byte-for-byte to the registered new surface and contain at most sixteen bytes;
• source and complete draft each fit 4,096 tokens independently;
• all selection, drop, span, offset, and hash receipts are stored.

No public benchmark or sealed holdout is part of Stage 0.

Optimization

• seed 2026080917;
• 512 updates, one same-source clean/fault pair per update;
• AdamW, LR 3e-4, cosine decay, fused optimizer;
• equal aggregate action, start-pointer, end-pointer, and replacement-token losses on fault examples; clean examples contribute action loss only;
• gradient norm clipping at 1.0;
• maximum replacement generation sixteen bytes plus EOS.

Frozen Stage-0 gate

All conditions are conjunctive on the 256-identity source-disjoint diagnostic:

1. aligned exact edit-program accuracy >=95% overall and per family;
2. aligned executed complete-trajectory accuracy >=95%;
3. aligned clean copy >=99% and fault repair >=90%;
4. aligned same-source counterfactual consistency >=95%;
5. aligned exceeds label-permuted, draft-hidden, and shuffled-draft executed accuracy by at least 13 answers each;
6. hidden and label-permuted exact edit-program accuracy are each <=60%;
7. forcing the gold program versus a wrong program changes downstream executed trajectories in the expected direction on at least 95% of fault rows;
8. zero accepted malformed pointers, offset violations, replacement decode mismatches, or generation exhaustion;
9. complete parameter, update, token, FLOP, memory, latency, and hash receipts.

A one-update mechanics receipt precedes the two fits. A miss closes exact PSET1-v0 without width, layer, seed, duration, pointer, or replacement-budget variants. A pass permits one broader natural-draft gate, not a public reasoning claim.

Controls and claim boundary

DSET1 full-script generation, source-only/draft-hidden, shuffled draft, and label-permuted pointer learning are explicit controls. A pass would establish a reliable model-owned copy/edit mechanism on deterministic final-span faults. It would not establish arbitrary multi-edit planning, natural fault discovery, general reasoning, or large-MoE transfer.