# TinyStyler mechanism pilot: development result

The frozen 64-candidate sweep completed with 64 EOS completions, zero complete factual-fidelity passes, zero exact-literal passes and zero joint voice passes. This is primary-session model review, not independent human evaluation. Four voices, two passages, two checkpoints, two style strengths and two seeds were tested. Both source controls preserve their supplied facts unchanged; copying them does not demonstrate target cadence.

All four 16-candidate conditions failed. Outputs omitted dates, commands, reported quantities and qualifications, substituted invented quantities, and added personal knowledge, organizational actions, software behavior or causal relationships. Generic forum-like prose persisted across the historical voices. Richer source facts did not rescue these particular TinyStyler outputs; topic and source length are confounded, so this does not decide the separate sparse/rich prompt-based hypothesis.

Generation used 4,636 decoder tokens including start tokens, approximately 66.35 seconds of measured generation across four processes, and at most 3.37 GB of allocated CUDA memory. Service window was 243.18 seconds including CPU embedding, loading and restoration. These are direct PyTorch measurements, not Inspect/Ollama or equal-budget model comparisons. Dependencies reused read-only Titan PyTorch 2.13.0+cu132 with an isolated Transformers 4.57.6 overlay. Initial standalone package download hit a disk-quota error; recovery did not modify service environments. Fortemi was restored healthy.

The implementation audit found correct core conditioning and strict checkpoint loading, with two reproduction limits: the pilot used top_k=0 instead of upstream default50, and preserved leading whitespace on two Twain samples where SentenceTransformer would strip it. The latter does not affect the other three voices. All frozen source/example token lengths fit the input limits; no silent truncation occurred. One over-limit whole Woolf sample was replaced before generation and recorded. No outputs were repaired or replaced. Raw input rather than production marker transport was tested, so this is not a test of protected-channel integration.

Decision: do not integrate this adapter or expand its sweep on current evidence. Preserve its failures as a distinct learned-conditioning mechanism result. The prepared sparse/rich study for instruction models remains unrun. A next pipeline should separate factual drafting, local expression edits and proposition-level verification, with an identity fallback; research already shows that unconstrained full rewrites and aggregate similarity are insufficient. Existing assisted span-repair evidence is the baseline for such work, not a novel result to claim again.

[Aggregate receipts](tinystyler-development-2026-09-07.json). Raw outputs, hash-bound primary reviews, source receipts and the implementation audit are retained privately under `.aiwg/working/voice-2292/tinystyler-pilot-01/`. Private authentic source text stays in the authorized workspace; no generated text is enrolled as an author sample.

The tested architecture and preservation variant come from the [published TinyStyler paper](https://aclanthology.org/2024.findings-emnlp.781/) and [pinned released inference source](https://huggingface.co/tinystyler/tinystyler/blob/2a879107b2ec342e57170b82cdc344d5179fa32b/tinystyler.py). The trial is an application-specific negative result, not a replication of its published benchmark.
