SOUL research papers
Revised research drafts · 4 November 2026.
These papers are not externally peer reviewed and share experiments from the same research programme.
A fifth draft, on personal logit steering, is not posted at present.
- A Failed Adapter-Steering Composition and a Possibly Format-Driven Adapter Gain on a Synthetic Extraction TaskTwo pre-specified comparisons on a synthetic extraction task. A combined adapter-steering route failed its criterion, which was unreachable at ceiling. On 32 new fictional profiles the trained student scored 32/32 against 16/32 bare; on the split tallied, every bare failure was one unquoted UNKNOWN, so the gain may be largely output format.Open PDF ↗ · research draft
- Duty-Cycled LoRA Training on One Consumer Phone: Step-Count Feasibility and Thermal Stops during ReloadThermal stop-rule studies on one USB-connected Galaxy S25. A 20 s train / 200 s pause schedule reached 232, 236 and 225 of 400 planned updates in three same-seed repeats; two of five reload checks stopped. Useful personalization, energy and a sustained loop were not measured.Open PDF ↗ · research draft
- Valid Edits, Unmeasured Authority: A Validator Fixture and Four Small-Model Proposer DiagnosticsA typed-state validator and four small-model proposer diagnostics that measure different layers. The primary proposer makes 16/16 required edits and 5/16 correct abstentions on 32 synthetic cases; no condition meets the joint screen. Authority enforcement on model output remains unmeasured.Open PDF ↗ · research draft
- Compiled Style Blocks versus a Trained Text Reader: A Negative Result and a Prompt-Card Edit Study on Synthetic UsersCompiled style blocks lost to a trained text reader on strict style accuracy, 0.780 versus 0.967, with the loss concentrated on identities holding one style pair never seen in training. A second study found lower off-scope drift after a fact edit, with qualifications. Synthetic users only.Open PDF ↗ · research draft