🧩 PRISME: Intervention-Supervised Sparse Workspace Editing for Small Language Models
UQO Working Paper No. 31 — Design, controlled evaluation, and negative and positive evidence at the 100M-parameter scale
TL;DR — PRISME adds to a frozen causal decoder a read-only source bank, a 16-slot ephemeral workspace, a shared sparse editor (1–2 slots per action, 24 actions + STOP), a value controller trained on executed counterfactual gains, and a differential readout with an exact zero-budget bypass. We implement, test, train and evaluate it against matched baselines. Result: the editor is learnable and helps a frozen backbone by a few percent of loss, but a global-recurrence adapter trained identically helps as much; where to write does not matter for fixed schedules; a target-reading oracle shows large opportunity that the learned controller cannot capture (it is worse than no routing on held-out data); and dispatch overhead erases sparse savings at small scale. A careful negative result with a mechanistic explanation.
📌 Status
| Item | Value |
|---|---|
| Version | 1.0 (October 2026) |
| Phases executed | 0–10, 13–15 (see PROGRESS.md); 11 (MLX port) not executed; 12 limited to meta-instantiation |
| Decisive experiments | 5 adapter seeds × {sparse editor, global recurrence, dense writes} on a frozen synthetic backbone; 103M FineWeb-Edu pilot (one backbone seed) |
| Code & data | REPRODUCE.md; synthetic suite regenerated from seeds; FineWeb-Edu subset documented, not redistributed |
🗂 Repository layout
prisme/ reference implementation · training/ stages A–D · baselines/ global recurrence + reference routers ·
evaluation/ routing, latency, perplexity, locked driver, statistics · data/ generators, processors, splits ·
tests/ invariants + cache equivalence · scripts/ audits, counts, FLOPs, launchers, tables, figures, claim audit ·
protocol/ pre-registration, locked test manifest, deviations · literature/ verified bibliography + novelty audit ·
experiments/runs/ manifests + metrics (checkpoints on the training nodes) · results/ raw/analyzed/tables/figures ·
paper/ LaTeX + PDF · reproducibility/ environment, cards, run index, audit · PROGRESS.md, DECISIONS.md, FAILURES.md.
🚀 Quick start
uv venv --python 3.12 .venv && uv pip install -p .venv/bin/python -r requirements-lock.txt
.venv/bin/python -m pytest tests -q
.venv/bin/python scripts/count_params.py --all-configs --device meta
scripts/reproduce_pilot.sh # miniature end-to-end pipeline on CPU (~1 min)
.venv/bin/python scripts/make_tables.py && .venv/bin/python scripts/reproduce_figures.py