UQO Working Paper No. 31· v1.1· November 2026· 72 pages

PRISME: Intervention-Supervised Sparse Workspace Editing for Small Language Models

Design, controlled evaluation, and a bounded negative result on learned spatial routing (five-seed synthetic suite; 103M pilot with two backbone seeds)

AI Systems

Quality controlled65.8/100

Learned spatial routing in a sparse latent workspace does not pay: the primary effect is +0.9 points against a three-point threshold, and the learned controllers are indistinguishable from no routing (+0.3 ± 1.0 % over five seeds).

On a frozen synthetic backbone, one or two sparse edits lower held-out loss by 3.6 %, but every global-recurrence variant does at least as well and a per-token MLP with no workspace does better (5.2 %), so nothing attributes the gain to the workspace. At 103M parameters on 98.3M tokens of web text, no workspace adapter or decoder variant moves the loss by more than the backbone's own seed-to-seed gap. The result is bounded: it concerns an ephemeral workspace on frozen backbones at the scales and token budgets tested.

Simon-Pierre Boucher — contact@spboucher.ai

Paper PDF (revised)Read online Code & data repository

+0.9 points

Primary effect on the locked test regimes (pooled exact match, sparse editor at budget 2 against budget 0, five seeds) against a three-point threshold; pooled item interval [+0.31, +1.57]. The primary-comparison definition was fixed after the development results, which favours the architecture.

−3.6 % vs −5.2 %

Held-out loss change on the frozen synthetic backbone at budget 2: sparse workspace edits against a per-token MLP adapter with no workspace. Every global-recurrence variant does at least as well as the sparse editor, so the gain is not attributable to the workspace.

+0.3 ± 1.0 %

Held-out loss of the learned value controllers relative to no routing, over five seeds: indistinguishable from not routing and far from the fixed schedules (−4.2 %). A stop-only decider with a random action does as well.

−15.4 %

Loss change found by an exhaustive oracle that reads the target: a ceiling, not a deployable policy. Only 25 % of the variance of intervention gains is action-level (12.1–13.4 % across five label sets); the rest is state-level.

0.375 → −0.011

Share of the one-step opportunity the controllers capture on audited states versus on the states they reach when deployed; the gap is distribution shift (−0.386), not the cost penalty (+0.004).

0.00523 nats

Seed-to-seed gap of the 103M backbone on validation (two seeds, 98.3M tokens of web text): no workspace adapter or decoder variant moves the held-out loss by more than this; the only effect outside it is the per-token MLP (−0.00663 nats). One run per adapter.

3.0× vs 1.92×

Measured decode latency of eight sparse edits at 103M on Apple Silicon against the analytic operation count; dense writes are only 18 % slower than sparse ones, so per-action dispatch overhead erases most of the arithmetic savings of sparsity.

Abstract

PRISME augments a frozen causal decoder with a read-only source bank, an ephemeral sixteen-slot latent workspace re-created at every position, a shared sparse editor that rewrites one or two slots per action, a value controller trained on counterfactual intervention gains measured by executing paired edit/no-edit branches, and a differential readout with an exact zero-budget bypass. We implement the architecture, prove its structural properties, reproduce the proposal’s parameter and operation accounting, and test seven hypotheses under a pre-registered decision rule and thresholds, with seeded comparisons against a parameter-matched plain decoder, a dense-slot editor, global-recurrence adapters (at the proposal’s size, at the editor’s own parameter count, and with cross-position reads restricted to the frozen latents), two workspace-free controls, five reference routers and exhaustive oracles. The result is a bounded negative one. On a frozen synthetic backbone, one or two sparse edits lower the held-out loss by 3.6 % and raise exact match by about two points, but every global-recurrence variant does at least as well, and a per-token MLP with no workspace does better (5.2 %): the workspace adds nothing these comparisons can identify. Where to write does not matter for fixed schedules. An oracle that reads the target finds large opportunity (−15.4 % of loss), but only 25 % of the variance of intervention gains is action-level (12.1–13.4 % across five label sets), the rest being state-level; the learned controllers are indistinguishable from no routing on held-out data (+0.3 ± 1.0 % over five seeds) and far from the fixed schedules (−4.2 %), a stop-only decider with a random action does as well, and the gap between what the controllers capture on audited states (0.375) and on the states they reach when deployed (−0.011) is distribution shift, not the cost penalty. On the locked test regimes the pre-registered primary effect is +0.9 points against a three-point threshold. At 103M parameters on 98.3M tokens of web text with two backbone seeds, no workspace adapter or decoder variant moves the held-out loss by more than the backbone’s own seed-to-seed gap (0.00523 nats); the only effect outside it is the per-token MLP (−0.00663 nats). The zero-initialised readout prescribed by the proposal learns slowly because of the saddle at U = 0, from which a fixed scale, a non-zero initialisation or a larger learning rate on U each recover most or all of the gap, and on Apple Silicon per-action dispatch overhead erases most of the arithmetic savings of sparsity.

JEL: C45, C63, C88

Key Results

A complete, tested reference implementation

PRISME is implemented with exact parameter and operation accounting for 103M, 459M and 1.18B configurations (nothing above 103M is trained), an incremental inference engine with separate backbone and bank caches, a 15-invariant test suite and a zero-budget bypass that holds bitwise.

Counterfactual labelling and oracle audits

A pipeline that executes paired edit/no-edit branches to label intervention gains, exhaustive one-step and two-step oracles, a variance decomposition of gains into state and action components, an on-policy audit that separates distribution shift from the cost penalty, a stop-only decider and target-free surrogate oracles.

Controlled, seeded comparisons that do not favour the workspace

The sparse editor is compared with a parameter-matched plain decoder, a dense-slot editor, global-recurrence adapters at three sizings (the proposal's, the editor's own parameter count, and with reads restricted to the frozen latents), two workspace-free controls and five reference routers, with item-level and seed-paired uncertainty.

An explanation of why the learned router fails

The opportunity the oracle finds is mostly a property of the state, not of the action; the controllers capture part of it only on the states their labels describe and none on the states they reach when deployed, and deciding when to edit does as well as also deciding where.

Measured cost of sparsity on Apple Silicon

Latency and memory profiles show where sparse computation does and does not pay: at 103M eight sparse edits cost 3.0 times the backbone's decode latency against an analytic 1.92 times, and at batch 8 the controller is slower than eight forced edits.

The zero-initialised readout starts slowly, it does not fail

The readout prescribed by the proposal learns slowly (−0.4 % after 3,000 steps against −1.6 % with the recipe used throughout) because of the saddle at U = 0; a fixed scale, a non-zero initialisation or a larger learning rate on U each recover most or all of the gap.

Quality control

This paper went through a multi-stage quality control: 7 quality-control reports and 105 points checked. The version presented on this page is the revised manuscript. Scores are on a 0–100 scale and are a diagnostic of the manuscript, not an acceptance probability.

Score after revision

65.8/100

57.0 before quality control · +8.8

Quality control dated November 9, 2026

Quality-control decision: major revision
Scientific integrity: PASS
Reproducibility: R1reproduced from documented analysis data
Maturity: Bone revision cycle from submission
Quality-control reports: 7
Points checked: 105

Scores by dimension

Before quality controlAfter revision

Originality & contribution53.1 → 56.0
Scientific rigor53.7 → 66.0
Identification & methodology51.6 → 64.0
Data & reproducibility69.3 → 75.0
Robustness & validation57.9 → 66.0
Writing58.9 → 66.0
Figures & tables52.9 → 70.0
Literature positioning71.4 → 78.0

Revised manuscript72 pages10 figures22 tables13 equations59 references30,304 words

What quality control changed

Quality control left the thesis unchanged — a negative result on learned spatial routing — but re-scoped five abstract-level sentences and bounded the result. New controls show that the small gain of the sparse editor cannot be attributed to the workspace, and the claim that a plain decoder beats both adapters at 103M was withdrawn once a second backbone seed showed every difference lies inside the seed-to-seed gap.

Zero-initialised readout: "does not learn because of a scale mismatch" → learns slowly because of the saddle at U = 0; a fixed scale, a non-zero initialisation or a larger learning rate on U each recover most or all of the gap.
"Parameter-matched" global recurrence → qualified: the adapter as first sized trains 26 % (tiny) and 37.5 % (103M) more parameters than the sparse editor; variants at the editor's exact count and with frozen-latent reads still do at least as well.
Editor usefulness: sparse edits help the frozen backbone → weakened: a per-token MLP with no workspace lowers the loss more (5.2 % against 3.6 %), so the workspace adds nothing these comparisons can identify.
103M pilot: "a parameter-matched plain decoder does better than both" → withdrawn; with two backbone seeds and full splits, every workspace and decoder difference lies inside the backbone seed gap (0.00523 nats).
Variance of intervention gains: 25 % described as state-level → 25 % is the action-level share (12.1–13.4 % across five label sets), the rest being state-level.
Controller: "worse than no routing", ranking "near chance" → indistinguishable from no routing (+0.3 ± 1.0 % over five seeds); the audit-to-deployment gap is distribution shift, and a stop-only decider does as well.

Limitations stated in the paper

Scale: the decisive seeded comparisons use 1.6M- and 4.75M-parameter character-level backbones on a synthetic suite; the 103M pilot has two backbone seeds, one run per adapter and one token per parameter. The evidence does not refute the architecture class at larger scale, with joint training or with persistent slots.
Every adapter was trained on a frozen backbone, and the recipe (bootstrap device, forced-budget distribution, learning rate) was chosen on development data; other recipes could change the ranking of adapters.
The primary-comparison definition was fixed after the development results and first committed after unsealing, with self-recorded freeze and unsealing times; the secondary locked-test reads foreseen by the pre-registration were not executed.
Four analyses requested during quality control were not run and the corresponding questions stay open: a matched-token continuation of the plain decoder, a persistent-slot arm, an on-policy relabelling iteration and output-head-only fine-tuning.

Read online (interactive reader)Revised PDF Revised branch on spbgit

Data

Synthetic compositional suite

Six generator families with programmatic ground truth (arithmetic, variable binding, register updates, constraints, scheduling and a held-out composition family), five phrasing templates each; 23,005 training items after duplicate removal, with iid, held-out-template, deeper-composition and held-out-family regimes, a separate calibration stream and locked test streams.

FineWeb-Edu subset (103M pilot)

The head of FineWeb-Edu sample-10BT (ODC-By 1.0) in canonical shard order, exactly deduplicated and split by content hash (98/1/1 %): 290.0M train, 2.91M validation and 2.88M locked test tokens. The 103M backbone was pretrained on 98.3M tokens in 2.3 hours on one Apple M3 Ultra (validation perplexity 46.53). It is not a random sample of the corpus.

Counterfactual intervention labels

199,860 executed counterfactual labels on 12,272 states, collected from an editor-unseen stream on states reached by random prefixes, plus exhaustive 24-action audits; the value controllers are trained on these labels over five seeds.

Methodology

Architecture with tested structural properties

A frozen causal decoder is augmented with a read-only source bank, an ephemeral 16-slot workspace re-created at every position, a shared two-block sparse editor that rewrites one or two slots per action (24 nonterminal actions plus STOP), a value controller and a differential readout that returns the backbone output bitwise when no edit is made.

Counterfactual supervision of the controller

Intervention gains are measured by executing paired edit/no-edit branches at training time; the controller predicts these gains and is evaluated against random, round-robin, fixed-pair, entropy-threshold and norm-change routers and against oracles that read the target, reported as ceilings only.

Matched-recipe baselines and workspace-free controls

All systems in a comparison share the backbone checkpoint, tokenizer, data stream, step count and batch size. The global-recurrence adapter as first sized trains 26 % (tiny scale) and 37.5 % (103M) more parameters than the sparse editor, so variants at the editor's exact count and with frozen-latent reads were added, together with a per-token MLP and a budget-indexed constant shift.

Decision rule, locked test and its chronology

Metrics, thresholds, baseline list and seed plan were committed before any adapter result; the primary-comparison definition (forced budget 2) was fixed after the development results and first committed after the locked splits were unsealed. "Pre-registered" therefore covers the decision rule and thresholds, not the primary-comparison definition. Each system was read once on the locked test.

Uncertainty

Differences between systems on the same items use a paired bootstrap over items (10,000 resamples); differences across training seeds use the per-seed statistic with a t-interval. Tokens of one training run are never treated as independent replications of training.

Operation and latency accounting

Analytic operation counts are compared with latency measured on Apple Silicon (fp32, PyTorch MPS backend, Python dispatch included) with warm-up and synchronisation, at the tiny scale and at 103M.

Reproducibility

Code, configurations, run manifests, result files, figures and the manuscript live in one repository, released under the MIT licence.

Every run has an immutable identifier, a configuration hash, the device and the measured wall-clock and memory; every number in the manuscript is produced by scripts from those files and checked by a claim-verification script, and make all rebuilds every table, macro, figure, the claim audit and the PDF.

The synthetic suite is generated deterministically from seeds; the token files and tokenizer of the FineWeb-Edu subset are redistributed under ODC-By with an attribution notice, the raw text is not.

REPRODUCE.md gives the commands for the environment, the tests, a miniature end-to-end pipeline and the figures.

Of the 84 run manifests of the original campaign, 75 record no code commit; the revision lists the SHA-256 and node path of 134 checkpoints and re-executes every essential analysis from committed files, but the original runs cannot be tied to a commit after the fact.

No paid API or cloud service was used; computation ran on the author's own Apple Silicon machines.

Cite this paper

Boucher, S.-P. (2026). PRISME: Intervention-Supervised Sparse Workspace Editing for Small Language Models. UQO Working Paper No. 31, Université du Québec en Outaouais.

@techreport{boucher2026wp31,
  author      = {Simon-Pierre Boucher},
  title       = {PRISME: Intervention-Supervised Sparse Workspace Editing for Small Language Models},
  institution = {Université du Québec en Outaouais, Département des sciences administratives},
  number      = {31},
  type        = {UQO Working Paper},
  year        = {2026},
  month       = {nov},
  url         = {https://www.spboucher.ai/research/wp31},
}