Owned Edits to a Shared Decision Rule: Learning Private Exceptions as Typed, Gated, Withdrawable Weight Edits
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
| Application | Public rule | Private holders; why no pooling | A declaration reveals | Possible acceptance body | Most likely failure |
|---|---|---|---|---|---|
| Component derating and qualification (worked example) | Per-family load ratios with a test-derived safe-operating limit [ 9 ] | Test houses, integrators’ parts engineers, suppliers; test cost, export control, procurement intent | Part, slot values, signer; hence which part the signer is qualifying | The standard’s specification authority for qualified parts; open for commercial parts | Slot too small for real records (energy, lot); no body willing to carry liability |
| Earth-observation calibration and data products | Published cross-sensor adjustment coefficients; published cloud-screening thresholds | Operators of commercial sensors, each owning one sensor’s calibration evidence | Gain and offset for the signer’s own sensor | A third-party mission-quality assessment programme | Corrections are continuous; for masks, truth is public and labels plentiful |
| In-situ sensor-network quality control | Published plausibility flags for soil-moisture records | Network operators holding site calibration and maintenance logs | A site-specific limit, e.g. where a site saturates | The network data centre or its quality service | Operators already share data, so the case for not pooling is weak |
| Space-weather anomaly screening | Public alert thresholds on the particle environment | Satellite operators holding anomaly logs; customer confidence, security | That a bus family is susceptible above some level: a vulnerability | A national space-weather centre, or a trusted operators’ association | What is declared is itself sensitive; anonymity preferred to attribution |
| Other sectors (unassessed) | Varies | Varies | Varies | Varies | Not examined |
| Tier | Components relevant to the protocol |
|---|---|
| Running today | Self-sovereign identities as public keys, with standard Schnorr signatures from established, audited libraries; verification of signed requests at the inbound boundary, failing closed; a per-identity data store with deny-by-default access control; an append-only, bi-temporal history of assertions that can be queried at any past time (its writes are not yet transactional, so a partial record is possible). |
| Built and tested, not switched on | A signing service that offers a closed list of named operations and refuses generic signing; a library for salted-hash commitments with signed disclosure and revocation receipts; a durable, crash-safe store for review cases from proposal to publication; an independent re-check of a proposed change by a pinned external checker that is not a build dependency of the system it checks; a timeline view that shows what was added and retracted between two times. |
| Designed only | Separate keys per party and for the acceptance body; a declaration event type and a hash-chained ledger of declarations, optionally anchored externally; a sealed reference set with rationed probing; a recorder, beyond the reach of the learner, of the model and prompt behind a declaration; a confined configuration in which a party’s records provably stay on its premises; erasure of a disclosed value. |
| Plan | Commit | What it fixed | Deviations, as recorded |
| v1: fixed-truth replication (exploratory) | hashed 0edf77ae ; first committed in 86bb5fb | Three fixed truths; 30 draws; seven record counts; pre-stated tie-break primary; judgement gate. | Hashed but not committed before its run, and draws 0–9 had been seen by an earlier pilot, so all of it is exploratory. Four departures: (1) an ungated no-prior arm was added, because the gate’s null degenerates when the reference rule is worse than a constant rule; (2) the confined-gradient patch was checked under a 16-edit limit; (3) federated values counted changed floats rather than transmitted floats, which undercounts; (4) the pooled large net’s step size, which the plan left open, was fixed before the run. |
| v2: confirmatory run 1 | abfc9c8 | 20 sampled truths per class; truth as the unit of inference; minimal-change tie-break primary; seven CV-tuned baselines; endpoint of disclosure ratio plus accuracy non-inferiority; judgement gate unchanged. | (1) The endpoint failed as first run, in one draw . The code counted the federated floats a client changed , not the floats it transmitted as the plan defines. In one draw every record already agreed with the public rule with margin, so the federated updates were exactly zero and Ditto pFL’s bound was counted as 0 bits. The disclosure clause therefore failed in that draw, and the endpoint as first run fails. With uploads counted as the plan defines them and the cell re-run, the endpoint passes; accuracy and spill are byte-identical, and the as-run rows are kept. (2) The grid was reduced from 70 cells to 9 before any non-primary cell was analysed. (3) The boosted-tree leaf regularisation, which the plan left open, was set to 1. |
| Gate tuning (development) | 3a52b7e | Sixteen gate settings on fresh data; feasibility bounds; selection rule fixed before the sweep. | None in the selection. The chosen setting was selected on development data, so its false-proposal rate there is biased low. A column for wrong-party proposals was added to the report after the sweep, computed from the committed rows; the criterion was unchanged. |
| v3: confirmatory run 2 | 78c9304 | Tuned gate ( , ); fresh truths and draws; accuracy-only endpoint; S11; the nine-cell grid. | None in the design, gate, data or grid. The disclosure clause was removed from the endpoint: once uploads are counted as transmitted it is close to automatic, and in v2 it had added brittleness without evidence. The analysis routines for v3 were written after the plan and before any v3 row existed; the reconstruction of the judgement gate was validated on v2’s rows first. |
| Trained base (hybrid) | 8a5f6a5 | 40 truths with shared and owner components; endpoint of two superiorities plus guarantees; secondary family S1–S6. | One confined-gradient row exceeded the edit limit and is marked partial (slot ablation only). Pooled training was expected to be an upper bound and was not. |
| Null check (exploratory) | 6a38b4f | The tuned gate on a fresh public-rule null, at three record counts. | Not pre-registered: written after run 2. Its pooled false-proposal rate’s interval reaches above the 5% bound, and the rate itself exceeds the bound at 100 records (Section 6.4 ). |
| Baseline | Difference | 95% interval |
|---|---|---|
| confined gradient | [ , ] | |
| Ditto pFL | [ , ] | |
| FedRep pFL | [ , ] | |
| FedAvg | [ , ] | |
| pooled 79-param net | [ , ] | |
| pooled large net | [ , ] |
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
| space | , | , | for | |||
|---|---|---|---|---|---|---|
| margins only | 73 | 6.19 | 4.29 | 0.243 | 0.073 | 73 |
| all three forms | 127 | 6.99 | 4.84 | 0.261 | 0.078 | 79 |
| linear | 1,113,751 | 20.09 | 13.92 | 0.564 | 0.169 | 170 |
| arm | owner, 0%, 10 | owner, 0%, 30 | owner, 0%, 100 | owner, 10%, 30 | shared, 0%, 30 | interaction, 0%, 30 | relaxation, 0%, 30 |
|---|---|---|---|---|---|---|---|
| public rule | 91.7% [90.2%, 93.1%] | 91.7% [90.2%, 93.1%] | 91.7% [90.2%, 93.1%] | 91.7% [90.2%, 93.1%] | 91.4% [89.9%, 92.8%] | 93.4% [92.2%, 94.6%] | 93.5% [92.5%, 94.5%] |
| typed, minimal change, gated (primary) | 93.0% [92.1%, 93.9%] | 96.6% [96.2%, 96.9%] | 99.6% [99.5%, 99.8%] | 94.5% [94.0%, 95.1%] | 96.3% [95.7%, 96.9%] | 95.3% [94.7%, 96.0%] | 93.5% [92.5%, 94.5%] |
| typed, pre-stated tie-break, gated | 92.1% [91.1%, 93.0%] | 96.5% [95.8%, 97.2%] | 99.6% [99.5%, 99.7%] | 94.3% [93.8%, 94.7%] | 96.4% [95.8%, 97.0%] | 95.1% [94.5%, 95.7%] | 93.5% [92.5%, 94.5%] |
| typed, minimal change, no gate | 93.6% [92.7%, 94.5%] | 96.6% [96.3%, 97.0%] | 99.6% [99.5%, 99.8%] | 94.8% [94.3%, 95.3%] | 96.5% [95.8%, 97.1%] | 95.3% [94.7%, 95.9%] | 93.5% [92.5%, 94.5%] |
| typed, pre-stated, no gate | – | 96.5% [95.8%, 97.2%] | – | – | – | – | – |
| free-form linear, gated | 91.6% [90.4%, 92.9%] | 95.7% [95.1%, 96.2%] | 99.3% [99.1%, 99.6%] | 92.9% [92.2%, 93.6%] | 95.8% [95.0%, 96.6%] | 94.7% [94.2%, 95.3%] | 93.5% [92.5%, 94.5%] |
| cell | confined gradient | Ditto pFL | FedRep pFL | FedAvg | pooled 79-param net | pooled large net | pooled gradient-boosted trees |
|---|---|---|---|---|---|---|---|
| owner, 0%, 10 | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] |
| owner, 0%, 30 | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] |
| owner, 0%, 100 | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] |
| owner, 10%, 30 | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] |
| shared, 0%, 30 | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] |
| interaction, 0%, 30 | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] |
| arm | owner, 0%, 10 | owner, 0%, 30 | owner, 0%, 100 | owner, 10%, 30 | shared, 0%, 30 | interaction, 0%, 30 | relaxation, 0%, 30 |
|---|---|---|---|---|---|---|---|
| public rule | 90.9% [89.7%, 92.1%] | 90.9% [89.7%, 92.1%] | 90.9% [89.7%, 92.1%] | 90.9% [89.7%, 92.1%] | 89.6% [87.9%, 91.3%] | 94.1% [93.2%, 95.0%] | 93.3% [92.5%, 94.2%] |
| typed, minimal change, gated (primary) | 91.1% [90.0%, 92.2%] | 95.2% [94.5%, 95.8%] | 99.2% [99.0%, 99.4%] | 93.4% [92.7%, 94.1%] | 95.3% [94.8%, 95.8%] | 95.3% [94.8%, 95.8%] | 93.3% [92.5%, 94.2%] |
| typed, pre-stated tie-break, gated | 90.7% [89.5%, 92.0%] | 95.5% [95.0%, 96.0%] | 99.4% [99.2%, 99.6%] | 93.2% [92.5%, 93.9%] | 95.2% [94.8%, 95.7%] | 95.2% [94.7%, 95.7%] | 93.3% [92.5%, 94.2%] |
| typed, minimal change, no gate | 93.2% [92.3%, 94.0%] | 96.7% [96.2%, 97.3%] | 99.5% [99.3%, 99.6%] | 94.9% [94.3%, 95.6%] | 96.7% [96.2%, 97.2%] | 95.7% [95.2%, 96.2%] | 93.3% [92.5%, 94.2%] |
| typed, pre-stated, no gate | – | 97.4% [97.1%, 97.8%] | – | – | – | – | – |
| free-form linear, gated | 90.8% [89.6%, 92.1%] | 94.8% [94.3%, 95.4%] | 99.3% [99.0%, 99.5%] | 91.7% [90.7%, 92.7%] | 94.2% [93.4%, 95.0%] | 94.9% [94.4%, 95.5%] | 93.3% [92.5%, 94.2%] |
| cell | confined gradient | Ditto pFL | FedRep pFL | FedAvg | pooled 79-param net | pooled large net | pooled gradient-boosted trees |
|---|---|---|---|---|---|---|---|
| owner, 0%, 10 | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] |
| owner, 0%, 30 | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] |
| owner, 0%, 100 | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] |
| owner, 10%, 30 | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] |
| shared, 0%, 30 | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] |
| interaction, 0%, 30 | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] |
| comparison | run 1 difference | run 1 Holm | run 2 difference | run 2 Holm |
|---|---|---|---|---|
| S1 typed confined gradient | [ , ] | [ , ] | ||
| S2 typed Ditto pFL | [ , ] | [ , ] | ||
| S3 typed FedRep pFL | [ , ] | [ , ] | ||
| S4 typed FedAvg | [ , ] | [ , ] | ||
| S5 typed pooled 79-param net | [ , ] | [ , ] | ||
| S6 typed pooled large net | [ , ] | [ , ] |
| arm | owner, 0%, 10 | owner, 0%, 30 | owner, 0%, 100 | owner, 10%, 30 |
|---|---|---|---|---|
| typed learner (primary), run 2 (plan v3) | 93.0% [92.1%, 93.9%] | 96.6% [96.2%, 96.9%] | 99.6% [99.5%, 99.8%] | 94.5% [94.0%, 95.1%] |
| confined gradient, oracle-tuned | 92.6% | 94.2% | 96.4% | 92.7% |
| Ditto pFL, oracle-tuned | 93.4% | 95.6% | 97.8% | 90.9% |
| FedRep pFL, oracle-tuned | 92.8% | 94.8% | 97.0% | 87.4% |
| FedAvg, oracle-tuned | 92.7% | 93.6% | 95.2% | 87.8% |
| pooled 79-param net, oracle-tuned | 92.6% | 94.0% | 97.0% | 87.0% |
| arm | owner, 0%, 10 | owner, 0%, 30 | owner, 0%, 100 | owner, 10%, 30 |
|---|---|---|---|---|
| typed learner (primary), run 1 (plan v2) | 91.1% [90.0%, 92.2%] | 95.2% [94.5%, 95.8%] | 99.2% [99.0%, 99.4%] | 93.4% [92.7%, 94.1%] |
| confined gradient, oracle-tuned | 92.4% | 94.4% | 96.2% | 92.3% |
| Ditto pFL, oracle-tuned | 93.4% | 95.9% | 98.5% | 91.0% |
| FedRep pFL, oracle-tuned | 92.6% | 95.3% | 97.8% | 87.2% |
| FedAvg, oracle-tuned | 91.9% | 93.5% | 95.5% | 87.6% |
| pooled 79-param net, oracle-tuned | 92.0% | 94.2% | 97.9% | 87.0% |
| cell | proposes | worse than public (gated) | worse than public (no gate) | exact recovery |
| owner, 0%, 10 | 22.3% [18.0%, 27.4%] | 7.5% [3.2%, 16.3%] | 17.0% [11.8%, 23.9%] | 3.7% [2.1%, 6.4%] |
| owner, 0%, 30 | 77.0% [71.9%, 81.4%] | 6.5% [4.0%, 10.4%] | 7.5% [4.8%, 11.5%] | 28.3% [23.5%, 33.7%] |
| owner, 0%, 100 | 99.0% [97.1%, 99.7%] | 1.0% [0.3%, 2.9%] | 1.0% [0.3%, 2.9%] | 77.7% [72.6%, 82.0%] |
| owner, 10%, 30 | 50.7% [45.0%, 56.3%] | 17.8% [12.5%, 24.6%] | 28.0% [22.9%, 33.8%] | – |
| shared, 0%, 30 | 84.7% [80.2%, 88.3%] | 2.8% [1.3%, 5.6%] | 3.7% [2.0%, 6.7%] | 24.3% [19.8%, 29.5%] |
| interaction, 0%, 30 | 65.3% [59.8%, 70.5%] | 19.9% [14.9%, 26.0%] | 26.1% [20.8%, 32.2%] | – |
| arm or comparison | all 40 truths | interaction | relaxation | values / Holm |
| public rule | 85.5% [84.2%, 86.7%] | 85.2% [83.2%, 87.2%] | 85.8% [84.1%, 87.4%] | 0.0 |
| declarations over the public rule | 91.3% [90.4%, 92.1%] | 92.5% [91.7%, 93.4%] | 90.0% [88.6%, 91.4%] | 2.3 |
| trained base alone | 88.9% [88.0%, 89.9%] | 88.9% [87.5%, 90.3%] | 89.0% [87.5%, 90.5%] | 0.0 |
| combination (primary) | 94.3% [93.8%, 94.8%] | 93.4% [92.7%, 94.0%] | 95.3% [94.8%, 95.8%] | 2.2 |
| combination, no gate | 95.6% [95.0%, 96.2%] | 94.3% [93.6%, 95.0%] | 97.0% [96.6%, 97.3%] | 3.7 |
| gradient in reserved unit | 91.6% [91.0%, 92.2%] | 91.1% [90.4%, 91.8%] | 92.1% [91.1%, 93.1%] | 3.9 |
| cell | combination | base | decl. on public | Ditto pFL | fine-tuning | reserved-unit grad. | pooled |
|---|---|---|---|---|---|---|---|
| 100 shared, 10 private | 87.7% [86.7%, 88.7%] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] |
| 100 shared, 30 private | 92.9% [92.2%, 93.5%] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] |
| 100 shared, 100 private | 96.9% [96.3%, 97.4%] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] |
| 300 shared, 10 private | 89.2% [88.3%, 90.1%] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] |
| 300 shared, 30 private | 94.3% [93.8%, 94.8%] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] |
| 300 shared, 100 private | 98.3% [97.8%, 98.8%] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] | [ , ] |
| comparator | 10 private | 30 private | 100 private |
|---|---|---|---|
| combination over better base | [ , ] | [ , ] | [ , ] |
| trained base (CV) | [ , ] | [ , ] | [ , ] |
| trained base, oracle | [ , ] | [ , ] | [ , ] |
| better-trained base | [ , ] | [ , ] | [ , ] |
| private fine-tuning, oracle | [ , ] | [ , ] | [ , ] |
| fine-tuning, better base, CV | [ , ] | [ , ] | [ , ] |
| arm or comparison | accuracy or difference |
|---|---|
| public rule | 88.3% |
| typed, pre-stated, gated (v1 primary) | 96.1% |
| typed, no gate | 97.1% |
| typed, minimal change, gated | 96.3% |
| typed, minimal change, no gate | 97.5% |
| free-form linear, gated | 94.4% |
AI-Assisted Decision Making
AI-assisted decision making studies how computational systems can inform human choices while accounting for context, uncertainty, and consequences.
See all 154 papers