# WP31 research mandate (verbatim summary of the author's brief, 2026-10-29)

Working title: *PRISME: Intervention-Supervised Sparse Workspace Editing for Small Language Models*.
Author: Simon-Pierre Boucher (UQO). Status at initiation: research proposal + untrained reference prototype;
no demonstrated language-model advantage. Target: honest, reproducible, submission-quality paper + implementation.

## Immutable rules
1. Scientific truth first (no fabricated training, metrics, runtimes, citations, hardware, screenshots, proofs).
2. No novelty/superiority claims without evidence; up-to-date primary-source prior-art review; negative results acceptable.
3. Phases in order with acceptance gates; failures documented, never relabelled.
4. Minimal falsifiable version first (no persistent memory, RAG, tools, MoE, verifiers, RL).
5. Respect resource/permission boundaries (existing approved access only).
6. Never hide costs (params, executed ops, training FLOPs incl. counterfactual branches, wall-clock, RAM, caches).
7. Independent train / controller-calibration / dev / final-test splits; no contamination across paired branches.
8. Pre-register primary outcome and decision criteria before unsealing final test evaluations.
9. Provenance for every number (run id, config hash, code revision, data manifest, log, result file).
10. Runnable by another researcher (setup, scripts, seeds, model/data cards, licences, compiled PDF).

## Central question and hypotheses
Can counterfactual supervision of *where* to edit a small latent workspace improve the quality–compute frontier
of autoregressive LM beyond uniform workspace computation and adaptive global recurrence?
H1 editor usefulness; H2 spatial selection (learned 1–2 slot selection beats uniform/random at fixed cost);
H3 intervention supervision beats uncertainty/entropy/norm/fixed-budget routing (controller overhead included);
H4 compute adaptivity moves the Pareto frontier vs the best eligible matched baseline; H5 differential identity at
budget 0; H6 generalisation (unseen templates, deeper compositions, distractors, paraphrases, seeds, independent
distribution); H7 scaling robustness at 459M (conditional).
Null possibilities: oracle finds edits but controller cannot predict them; loss improves but answers worsen;
larger ordinary Transformer / shared-depth recurrence wins at equal compute; fixed two-slot editing matches
learned routing; dispatch overhead dominates sparse savings; source bank hurts long-context latency/memory.

## Architecture contract (Section 3 of the brief)
Configurations 103M / 459M / 1.18B (L 12/16/24, d 768/1536/2048, FFN 2048/4096/5632, heads 12/3, 24/6, 32/8,
w 256/512/768, FFN_w 768/1536/2048; vocab 32768, M = 16, L_w = 2, R_max = 8, tied head, RoPE, GQA, single final
norm). Reference counts: backbone 99,502,848 / 446,744,064 / 1,149,339,648; workspace+controller 3,088,260 /
12,336,900 / 25,779,844. Source bank = backbone final representations projected to width w, strictly causal.
S0[j] = e[j] + z_t, reset per token. 24 nonterminal actions (16 singles + 8 fixed pairs) + STOP. Two shared
editor blocks (bank cross-attention, slot attention over a coherent snapshot, SwiGLU, RMSNorm/residuals,
step/budget context); clipped gated writes (η_max = 0.25, c = 1); exact locality. Controller predicts ĝ, ĉ, σ̂,
p_harm; Q = ĝ − λĉ − κσ̂, Q(STOP) = 0; never executes candidate edits at deployment. Differential readout
Δ = U(r(S_final) − r(S0)), U = 0 at init, explicit budget-0 bypass (bitwise). Separate backbone KV and bank caches.
Reference ops/token at 459M, T = 4096: 1.296 / 1.645 / 1.767 / 2.195 / 5.608 billion (backbone; 4×k1; 4×k2;
8×k2; 8×dense).

## Training (Section 4)
Paired target g(s,a;y) = loss(D(s),y) − loss(D(T_a(s)),y). Label sampling: 4 candidates/state (policy preference,
competitive alternative, 2 random; balanced singles/pairs; propensities logged); exhaustive 24-action audit subset.
Stages A backbone → B editor bootstrap (frozen backbone, forced budgets {0,1,2,4,8}) → C counterfactual collection
→ D controller fit (Huber gain, BCE harm, calibrated σ; λ, κ tuned on calibration data) → E alternation/lock.
Delayed benefits: compare 1-step oracle, 2-action lookahead oracle, learned controller. Token loss ≠ task success:
evaluate free generation, track correct↔incorrect transitions vs budget.

## Phases 0–15, baselines (Section 7), ablations (Section 8), data/evaluation (Section 9), hardware policy (10),
tracking (11), paper structure (12), failure modes (13), decision rules (14), final report (15): as in the brief.
Suggested (non-binding) decision criteria: advance if PRISME exceeds the strongest eligible baseline by ≥ 3 pp at
comparable latency, or cuts median latency ≥ 20 % at non-inferior quality, with the uncertainty interval excluding
a null practical effect; zero-budget perplexity preservation within 1 %.