A shared decision rule can now be corrected one owner at a time: 96.6% at 30 records, non-inferior to all seven baselines
Synopsis
Several organisations apply the same published decision rule while each holds private records showing it is wrong for its own items, and none will pool them. One approach compiles the published rule exactly into a small ReLU network, lets each party submit only typed declarations, and gates them on evidence before compiling each into a weight patch confined to that party's reserved units. In two pre-registered synthetic tests, the learner scored 96.6% at 30 records per party against 91.7% for the unchanged rule and was non-inferior within 2 points to seven tuned baselines; the same search without the compiled prior scored 84.0%.
Interpretation
A shared decision rule can now be corrected on one party's own subjects without changing any other party's decisions. Previously the records were either pooled into a common model, which discloses them by design, or exchanged as gradients or weights in federated learning, from which records can be reconstructed, and removing a party afterwards needs retraining or approximate unlearning. At the primary cell of the second confirmatory run (owner-specific truths, no label noise, 30 records per party), the typed learner with the evidence gate scored 96.6% against 91.7% for the unchanged rule, and was non-inferior within 2 points to all seven cross-validation-tuned baselines, with every paired difference's lower bound above zero.
The unit of change is a declaration, which a public compiler turns into a weight patch written only into that party's reserved units. A declaration is a few named values drawn from fixed categories, such as subject, scope, form and size; the learner enumerates this small space on its own records and submits only when the records clear a selection-aware evidence gate. An acceptance body re-derives the patch from the declaration and runs exhaustive checks; in every draw of the synthetic benches typed patches spilled zero decisions onto other parties, withdrawal restored the model hash bit for bit, and each party disclosed about 2.2 declaration values against 210 to about 992 uploaded values for the pooled and personalised federated comparators.
The mechanism has boundaries beyond the rule: it only tightens, and it depends on a reference set that does not yet exist. When the truth is more lenient than the published rule, or shared by all parties, per-party declarations cannot express it, and the learner can only decline or fall behind pooled methods. On relaxation truths the learner proposed in 0.0% of party-draws and matched the public rule at 93.5% against 93.5%; on shared truths it trailed the pooled large net and pooled gradient-boosted trees; and the labelled reference set the acceptance test needs, chosen by no party, does not yet exist for any application the material identifies.
Perspective
The demonstrated scope is an accuracy comparison on synthetic screening benches, plus exact compilation, composition, non-interference and exact withdrawal of one published derating rule over 33 million valid inputs; the real-record bench has no accuracy evaluation because no held-out truth exists for real parts. The material addresses engineers and standards users who need several parties to correct one rule without pooling records.
The typed space was built to contain the owner-specific truths, so realisability holds by design; the evidence gate proposed on shuffled outcomes far above the pre-stated 10%, and in an exploratory null check written after the second run the tuned setting's false-proposal rate was 6.2% at 100 records, above the 5% cap at which it was chosen. A language model given the records decided 347 of 360 probe items correctly on one draw, against 348 of 360 for the patched network.
