Solver-Checked Delphi Consensus for Health Data-Sharing Norms

Privacy law tells hospitals to share patient data only in "appropriate" ways, but software needs exact rules: who may send which attribute to whom, under what conditions, and with what technical protection. Today those rules are written either by small expert committees or by averaging crowd votes, and neither includes the patients most affected by digital health's failures. We propose building the rules with the Delphi method, a structured survey in which an anonymous expert panel rates items over several rounds and sees summarized group feedback between rounds. We add one ingredient that no Delphi study has used: between rounds, an automated logic solver checks whether the panel's emerging rules contradict each other. It then returns the smallest set of conflicting positions to the panel as part of the feedback. The panel deliberately includes safety-net patients, caregivers, and community health workers alongside lawyers, engineers, and clinicians.

The practical payoff comes from what the process yields besides a consistent rule set. It also yields an explicit map of where different stakeholder groups genuinely disagree. That map lets a personal health data system apply community consensus as the default. It then asks an individual patient only about the flows that remain contested. A patient with limited digital literacy might face a handful of meaningful choices instead of a wall of ontology terms. The core claim is testable. Solver-in-the-loop consensus with a stratified panel should yield health data-flow rules that contain fewer logical contradictions than crowd voting or expert committees. The rules should also better predict the judgments of held-out safety-net patients, and they should shrink the number of consent decisions a patient must make without lowering how well those decisions match the patient's own preferences.

Background and the gap

Contextual integrity is the idea that privacy means information flowing according to the norms of its context rather than staying secret. It describes each flow by sender, recipient, subject, attribute, and transmission principle, the condition under which sharing is allowed, such as consent. Integrating Differential Privacy and Contextual Integrity adds a sixth element, the transmission property. This is the technical protection on the flow, such as aggregation, local computation, or a differential privacy noise level. The authors argue that social context should set these parameters, but they leave open who decides which properties matter and how. Their work supplies our representation; we supply the elicitation procedure it lacks. The closest method is Learning Privacy Expectations by Crowdsourcing Contextual Informational Norms. It collected 1,411 education-privacy vignettes from 451 Mechanical Turk workers, encoded majority judgments in a decidable logic, and used the Z3 solver after the fact. The solver found 138 semantic inconsistencies and 59 transitivity violations, and only 29% of norms reached a two-thirds majority. That paper motivates our design, since one-shot voting produces contradictory rules. It also supplies our method, since solver encoding is feasible. We differ by moving the solver inside an iterative process where humans resolve the conflicts, instead of repairing them afterwards, and by moving from education to health with a stratified panel instead of anonymous crowd workers.

On the consensus side, BRIDGE pilot study uses structured bilateral expert consensus to reconcile EU and US health-data regulations. Health Delphi studies such as an international Delphi on behavioral-trial methodology rely on anonymized rounds with median and interquartile feedback. Both stay at the policy level, draw panels from elites, and never check their outputs for logical coherence. On the systems side, TIDAL compiles machine-readable consent into queries over patient-held Solid pods, which are personal data stores the patient controls. However, its dashboard assumes users who can parse biomedical vocabulary. The digital-divide literature explains why that assumption fails. Eliminating the AI digital divide by building local capacity documents that safety-net organizations lack the legal and technical capacity to govern data systems. Usability work on privacy labels shows that informational disclosures are ignored or misread. We treat TIDAL as the deployment target and the divide as the design constraint, and we reduce the decisions asked of each patient rather than adding explanation.

Core Claim

We make three claims. First, feeding solver-detected conflicts back to a Delphi panel produces a final rule set that is logically consistent by construction, and it does so without the panel simply capitulating to whichever position the coordinator presents. Second, a panel that includes patients and community health workers produces rules that differ measurably from an expert-only panel on specific flows, and those differences improve prediction of held-out safety-net patients' judgments. Third, the explicit disagreement map supports a consent interface that asks only about contested flows. For low-literacy users, that interface should beat both granular consent and a single blanket default on decision agreement and completion.

Research Program

We first build a health vignette space. Each vignette is a six-part flow in the augmented contextual integrity form, drawn from health settings typical of safety-net care: a patient's glucose readings sent to a community health worker, mental-health notes to an insurer, or pod data to a research consortium under local computation. We seed the attributes and actors from the health-related regulations and scenarios underlying PrivacyLens and the health domain of CIMemories. We then prune the combinatorial space to roughly 300 to 500 vignettes, following the pruning strategy Shvartzshnaider et al. used to reduce 20,000 combinations to 1,411. Transmission properties are phrased concretely, for example "only a monthly average leaves your phone," rather than as raw epsilon values.

The panel has about 60 members in four strata. These are legal and privacy scholars, health-IT engineers, clinicians at safety-net sites, and patients, caregivers, and community health workers recruited through safety-net clinics. Delphi panels lose members over rounds, and low-literacy participants are the likeliest to drop out. The patient stratum therefore completes rounds by phone or with a community health worker reading items aloud. Each round covers at most about 40 items, and participants are compensated. Panelists rate each vignette's acceptability on a 1–9 scale and give short reasons. An item reaches consensus when roughly 70% of ratings fall in the top or bottom third and the interquartile range is at most 2. We treat these thresholds as design choices and will report results under alternatives.

Between rounds, consensus items are encoded as allow or deny facts in the same decidable logic fragment Shvartzshnaider et al. used. We add domain axioms: a rule granting a flow under aggregation is not violated by a flow under stronger aggregation, and a chain of individually allowed flows can reproduce a denied flow. Z3 then returns minimal unsatisfiable cores, meaning the smallest sets of the panel's own positions that cannot all hold. Each core is rewritten in plain language as a short dilemma and added to the next round's feedback, together with the standard medians and anonymized reasons. Where agreement holds within each stratum but not across strata, we do not force convergence. Instead, the item is recorded as contested, along with each stratum's position. The output is a consistent rule set plus a list of contested flows, compiled into the Data Privacy Vocabulary consent format that TIDAL consumes.

Baselines

The naive reference point is a rule set derived directly from regulation. We translate the HIPAA-grounded seeds in PrivacyLens into allow and deny rules on our vignettes, which represents what a compliance team would do without eliciting any norms. It shows how much of patients' actual judgment regulation already captures.

The strongest existing elicitation method is crowd majority vote followed by solver-based repair, as in Learning Privacy Expectations by Crowdsourcing Contextual Informational Norms. Its reported figures are 68% of norms at simple majority, 29% at two-thirds, and 138 inconsistencies plus 59 transitivity violations before repair. We run it on the same vignettes, with a crowd sample of equal total rating effort and the same solver and axioms. Conflicts are removed by dropping the lowest-margin rule in each unsatisfiable core, which gives it the same logical machinery without human resolution. A second strong competitor is a norm set produced by a language model trained for contextual integrity reasoning. We use the released CI-RL pipeline, trained on the contextual integrity synthetic dataset. Its authors report that a fine-tuned Qwen2.5-7B outperformed a 14B baseline on integrity scores and that CI-RL cut PrivacyLens leakage by up to 40%. We prompt it to label every vignette and apply the same solver repair to its labels.

Two ablations isolate our ingredients. An expert-only Delphi, with identical rounds and solver feedback but no patient stratum, isolates the contribution of panel composition. A stratified Delphi with standard feedback but no solver-generated dilemmas isolates the solver's role. Beating a baseline means higher balanced accuracy on held-out safety-net patients' judgments, with confidence intervals excluding zero difference, at zero residual inconsistencies. Baselines that reach zero inconsistencies only through automatic deletion must also retain comparable rule coverage.

Why not just…?

Why not just crowdsource with majority vote and let the solver repair contradictions afterwards? This is the most serious threat, because it is cheaper and already demonstrated. Automatic repair must pick which conflicting rule to drop using only vote margins. The humans whose judgments conflict never see the conflict, so the repair may delete exactly the rule a patient cares about. The comparison against solver-repaired crowd voting in held-out prediction settles this. If crowd-plus-repair matches the solver-checked Delphi, we will report that the human resolution loop is unnecessary and recommend the cheaper pipeline.

Why not just have a trained language model label the flows? Language models recite privacy rules well but apply them poorly. PrivacyLens reports 98.99% probing accuracy for GPT-4 alongside a 25.68% leakage rate in agent actions. Can LLMs Keep a Secret? reports that GPT-4's correlation with human judgment falls from 0.8 to 0.1 as contexts grow complex. More fundamentally, a model trained on 729 synthetic examples encodes no information about what safety-net patients in particular find acceptable. The CI-RL baseline measures this directly. If it predicts held-out patients as well as the panel does, human elicitation is poor value for this purpose.

Why not just convene an expert committee or rely on HIPAA? Committees are fast but suffer from status deference, which Delphi anonymity exists to remove, as the ankle cartilage consensus proceedings describe. They also exclude the people for whom consent interfaces fail. Our expert-only Delphi ablation and the regulation-derived baseline test whether adding patients changes anything. If patient and expert strata agree on nearly all items, the equity argument for stratified panels is weak and we will say so. The contested-flow interface would then reduce to a single default.

Why not just give patients a simpler consent form with plain-language explanations? Explanations do not remove the number of decisions, and privacy label research shows that disclosure alone rarely changes understanding or behavior. The user study below includes a plain-language granular condition to test this directly.

Proposed Experiments

The first experiment tests consistency and non-coercion. We count solver-detected contradictions per 100 consensus rules at each round and in the final set for every method. To check that the Delphi is not manufacturing agreement, we record whether panelists who change positions after seeing a dilemma cite the conflict in their reasons. We also compare how often positions move toward the median in items with solver dilemmas versus items without them. Movement concentrated in conflicted items, with no inflated convergence elsewhere, supports genuine resolution over conformity.

The second experiment, the decisive test, is held-out prediction. A separate sample of about 300 safety-net patients, recruited from partner clinics and never part of any panel, privately rates a random subset of vignettes. Each method's rule set predicts these ratings, dichotomized at the scale midpoint. We report balanced accuracy overall and on items where the strata disagreed. Our method should beat all baselines, and the gain should concentrate on the contested items. If it does not, the core claim fails.

The third experiment is a consent-interface study with the same held-out population, stratified by a standard health-literacy screen. Participants configure a TIDAL-style pod under one of four randomly assigned conditions. The conditions are granular ontology-level consent as in TIDAL, a plain-language version of that granular consent, a single blanket default taken from the regulation baseline, and consensus defaults with questions only on contested flows. The outcomes are agreement between the configured policy and each participant's own later vignette ratings, completion rate, and time. The prediction is that the contested-flow condition matches granular consent on agreement, with far fewer decisions and higher completion among low-literacy participants.

The fourth experiment is an executability check. We compile the final rules into a pre-action guard that blocks agent tool calls violating a rule, and we evaluate it on the health-domain subsets of PrivacyLens, CIMemories, and CONFAIDE. The reference points are the reported numbers from those papers: 25.68% action leakage for GPT-4 on PrivacyLens and over 25% leakage even with privacy prompts, 14.8% violations with 43.9% completeness for GPT-4o on CIMemories, and CI-RL's reported reduction of up to 40%. We rerun CI-RL and prompting baselines on the same subsets with the same base model. A guard counts as useful if it lowers leakage below CI-RL while keeping task completeness within a few points of the unguarded agent. This experiment is secondary. It shows whether the rules are precise enough to execute, not whether they are right. The held-out prediction test decides that.