cs.LGNov 11, 2026

Owned Edits to a Shared Decision Rule: Learning Private Exceptions as Typed, Gated, Withdrawable Weight Edits

Authors: John J. O'Hare

Organizations: DreamLab AI Consulting Ltd dreamlab-ai.com

Abstract

Organisations that apply the same published decision rule often hold private evidence that the rule is wrong for their own cases, and will not pool it. We describe a learning mechanism for that setting whose unit of change is an owned edit. The published rule is compiled exactly into a small ReLU network, a structured prior. Each party's learner searches a small typed space of declarations, a few named values such as subject, scope, form and size, and proposes one only when its own records pass a selection-aware evidence gate. The learning is that gated search around the prior. In two analyses registered before they ran, on synthetic data with 20 sampled hidden truths, the learner at 30 records per party was non-inferior, within 2 points, to all seven tuned baselines (a confined gradient learner, Ditto, FedRep, federated averaging, two pooled networks and pooled gradient-boosted trees), scoring 96.6% against 91.7% for the unchanged rule; the same search without the prior or the gate scored 84.0%. In a third registered test, typed edits over a network trained on shared records beat that network alone by 5.4 points and the same edits over the published rule by 3.1. An acceptance body recomputes each edit from its declaration and writes it into units reserved for that party, so by construction an edit changes no decision about any other subject, withdraws exactly, discloses at most log2 of the typed space's size in bits, and can hide nothing beyond its declared values. We report where it loses: truths shared by all parties, relaxations, ten records, a pre-stated gate criterion that failed in both runs, and a language model given the records that decided about as accurately on one draw. On a rule-only base, a table of owned exceptions makes the same decisions.

Figures & tables

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Sep 30, 2026

Is This Evidence Decision-Critical? Learning to Verify Rule-Governed Decisions

Rule-based reasoning, as in eligibility checks and contract reviews, requires language models to assess evidence against individual conditions and combine their judgments under explicit rules. Errors in evidence assessment can leave a decision unchanged, but misinterpreting or overlooking decision-critical evidence can reverse it. Identifying such evidence allows more capable models to focus on checking the corresponding condition judgments, supporting accurate and safe decisions. Recognizing the evidence's criticality requires understanding how evidence affects a condition judgment and how that judgment affects the decision. To achieve the goal, we propose a INTERvention-based imPACT learning framework (InterPact), which enables counterfactual verification of evidence criticality in rule-governed decisions. Specifically, its evidence intervention constructor generates training pairs for a propagation verifier by editing case facts with a frozen language model while holding rules and non-target conditions fixed. Human-reviewed labels record the resulting condition and decision changes, while complete state-to-decision mappings supervise consequences beyond the observed edit. During training, the verifier weights learned conditional decision predictions by evidence-based condition probabilities through a fixed composition operation, propagating decision-change supervision into the base model. At inference, the trained base model directly judges criticality from the original case and target evidence, without human or stronger-model supervision. On single-case evidence criticality verification over adapted rule-governed decision cases, InterPact achieves 68.28% accuracy, outperforming all six baselines. These results support learned decision sensitivity as a basis for prioritizing evidence checks.
Oct 24, 2026

Exact Unlearning After Data Curation

Training pipelines often remove similar examples before fitting a model. Deleting a raw record can change that selection and admit an example that was excluded before. What must a service retain to reproduce training after such requests? We answer this question for a fixed rule in which earlier similar records block later ones. A limit on the total number of future deletions identifies exactly which records can still enter training. We use this boundary to build a summary of training statistics that can be repaired exactly after repeated deletions. The repaired summary is the same as rebuilding it on the surviving data with the remaining budget. We also construct finite-precision features for ridge regression, a linear learner with a penalty on coefficient size. With exact rational outputs, allowing one more deletion changes the required private information. It rises from zero to every private feature sign in the excluded candidates. An encoding attains this conditional private-information minimum on the constructed family throughout a sequence of requests. On 1,001 AskUbuntu development records, deleting an initially excluded record admits another record at 13 of 51 fixed similarity cutoffs. Exact membership repair passes 13,056 deletion-step checks. Two model-based semantic calibrations fail their fixed criteria, so the downstream task study is not run. The central result is a retention contract: the future deletions a service promises to support determine which excluded data can still be needed.
Apr 18, 2026

One Mask to Rule Them All: On Hidden Facts after Editing and How to Find Them

Knowledge editing methods such as ROME and MEMIT update factual associations in transformer models by modifying MLP weights. While evaluated mainly by output behavior, their internal mechanism remains underexplored. We investigate whether edits rely on a common mechanism, regardless of which fact is modified. Despite fact-specific weight changes, we argue that ROME and MEMIT target the same subset of weights critical for maintaining edits. To isolate this subset, we train a compact binary mask over the edited weights. The mask reverses 80% of edits on the training set and over 70% on the test set, confirming that diverse edits share a common functional structure. Our analysis reveals that the mask reverses edits by eliminating overattention in later layers. Additionally, we show that injecting the mask during editing drops editing success from 98% to 38%, demonstrating that this mechanism is necessary for edits to succeed. Our finding that edits suppress rather than overwrite knowledge explains why ROME and MEMIT fail to propagate changes to related facts. The identified common functional subspace informs detection and defense against unwanted edits.