The Delphi method is how medicine, policy, and technical standards bodies reach agreement when hard data is missing. A panel of experts scores statements anonymously over several rounds. Between rounds, a coordinator shows everyone the group's score distribution. Experts whose scores fall far from the middle are asked to explain themselves. The whole method rests on one person or team, the coordinator, who sees every raw score, decides which dissenting arguments get circulated, and computes whether a statement crossed the consensus bar (for example, 67% agreement). This proposal argues that the arithmetic half of Delphi is easy to protect with existing cryptography. The real power the coordinator holds sits in the feedback channel: who is allowed to argue, whose words are shown, and whether an argument presented as coming from a dissenting panelist really did. We propose a small composed protocol that removes that power, and a set of experiments asking whether removing it changes what panels conclude.
Concretely, an expert whose committed score really lies outside the group's middle range can post an argument that is provably from a registered panelist and provably from a genuine outlier. The post cannot be traced to its author by the coordinator, and only one such post per item per round is allowed. Everyone can check the group statistics against the committed scores. We then test three things directly. We measure what the protocol costs at real panel sizes. We test whether the dissenter's own writing style gives them away anyway. We test whether verified dissent makes panels more accurate and less steerable than a standard trusted-coordinator Delphi, using forecasting questions that later resolve so accuracy can be scored.
Core Claim
For Likert-scale Delphi rounds, the published feedback (the histogram, median, interquartile range, and percentage agreement) is a function of the vote counts. Computing those counts privately is a solved kind of problem: two non-colluding servers each hold a random-looking share of every ballot and add them up. That is the "secret sharing" approach, and each ballot carries a short proof that it encodes exactly one valid score. The novelty is not there. The new and testable claim is narrower. Binding each circulated argument to a zero-knowledge proof of outlier status, plus a per-round anonymous rate limit, closes the two channels that the file identifies as the coordinator's real levers: de-anonymizing unpopular positions and fabricating or suppressing dissent. We further claim this produces measurable behavioral differences. Panels should retain more correct dissent and resist injected manipulation better. Whether that behavioral claim holds is the central empirical question, and a null result would be informative too.
The statement an outlier proves, given a hiding commitment $C$ to their score $s$ and the publicly verified quartiles $Q_1, Q_3$ of the round, is
$$ C = \mathrm{Com}(s; r) \;\wedge\; \big(s < Q_1 \;\vee\; s > Q_3\big) \;\wedge\; \nu = \mathrm{PRF}_k(\text{round}, \text{item}), $$
where $r$ is the commitment randomness, $k$ is the panelist's registered secret key, and $\nu$ is a tag that is deterministic for a given panelist, round, and item. A second post on the same item and round repeats $\nu$ and is rejected. Tags from different rounds or items cannot be linked to each other.
Relation to Prior Work
Nothing in the retrieved literature merges the two directions. The closest name match, Delphi: A Cryptographic Inference Service for Neural Networks, is a two-party private neural-network inference system that combines secret sharing and garbled circuits. It shares only the name and does not touch expert elicitation. A Consensus Privacy Metrics Framework for Synthetic Data runs a 13-expert Delphi panel about privacy, but the panel itself was conventional and coordinator-run. Its finding that differential privacy budgets are uninterpretable unless $\epsilon$ is near zero constrains our design and rules out a noise-based answer. The clinical panels The Vascular Impairment of Cognition Classification Consensus Study and Conservative Management and Biological Treatment Strategies motivate the work. They show consensus decided by hard percentage thresholds (≥67% in one; 51/75/100% tiers in the other), where a few votes or a few circulated arguments move a statement across a line, and where the evidence base is mostly expert opinion ("Grade E").
On the cryptographic side, several papers supply method rather than motivation. PINE shows how to check range and norm conditions on secret-shared vectors with small overhead. We reuse its style of validity check for one-hot ballots, though our vectors are tiny and fixed costs will dominate. Efficient Non-Interactive Zero-Knowledge Proofs in Cross-Domains without Trusted Setup combines hash-based and discrete-log-based proofs without a trusted setup. That is exactly the mix we need: hash-based tags, algebraic commitments, and an "or" range statement. We inherit its known weakness that the algebraic part is not quantum-safe. EVA-S3PC assembles multi-party regression from small verified sub-protocols, but it assumes semi-honest parties and an uncolluding helper server. We adopt the same honest assumption explicitly, two non-colluding aggregators, rather than claiming more. The Honey Badger of BFT Protocols builds asynchronous agreement from threshold encryption and reliable broadcast. For panels of 13–80 we need only its reliable-broadcast idea for the argument board, so that suppression is detectable. Its dependence on a trusted dealer for threshold keys also warns us to keep threshold keys out of the design. Finally, AI Agents and the Future of Deliberation proposes AI moderators for deliberation without any guarantees. We use a language model only on the expert's own device, as described below.
Research Program
The build has four pieces, each small.
The first is ballot aggregation. Each expert submits a one-hot score vector per item, secret-shared to two aggregators, with a validity proof. Each expert also posts a commitment to the score on a public board. The aggregators publish the summed histogram together with a proof that it is consistent with the posted commitments, so any panelist can verify the median, quartiles, and percentage agreement.
The second is the outlier credential. After quartiles are published, an outlier produces the proof above and attaches it to their argument.
The third is the argument board. It is append-only and replicated across the two aggregators plus any willing panelists. Each submission returns a signed receipt, so an argument that was accepted and later missing is provable.
The fourth is the style-normalization step. The ballot is protected cryptographically, but the justification text is not: writing style can unmask authors in a pool of 13 people who know each other's papers. Before submission, a local language model rewrites the argument into a neutral register, and the expert approves the rewrite. This is the one place we expect leakage to survive, so we measure it rather than assume it away.
The analysis question is which guarantees survive when the threat model is honest. The aggregators must not collude. The board needs a majority of honest replicas. Text anonymity is only empirical. We will state each assumption next to the guarantee it supports.
Baselines
The naive reference point is a standard trusted-coordinator Delphi. Participants enter scores into ordinary survey software, and a coordinator computes the histogram and circulates hand-selected anonymized arguments, as in the clinical panels cited above. It is run with the same questions, round count (three), panel sizes, and wording as every other arm. It counts as beaten if the full protocol yields lower final Brier score or smaller manipulation-induced shift at equal participation.
A simpler cryptographic baseline is commit-and-publish. The coordinator still sees raw scores, but posts per-ballot hash commitments and the histogram so tampering with counts is detectable. This is what a skeptic would build in an afternoon. It shares our integrity property for numbers and lacks confidentiality from the coordinator and any authenticity guarantee on arguments.
The key ablation is the full protocol with outlier credentials removed. Ballots stay private and verified, and the board stays anonymous, but anyone can post any argument. Comparing this arm with the full protocol isolates whether proof of genuine dissent, rather than anonymity alone, changes how panelists weigh arguments.
For forecasting accuracy, the strongest reported reference points are external. ForecastBench reports a Brier score of 0.096 for superforecasters and 0.121 for the general public. Agentic Forecasting using Sequential Bayesian Updating of Linguistic Beliefs reports a difficulty-adjusted Brier Index of 71.0 against a superforecaster median of 70.9, and +3.4 over a crowd-plus-empirical-prior baseline. We also run a single-round crowd median on the same questions as the no-iteration reference. Our panels are recruited forecasters, so "beating" means outperforming the coordinator-run Delphi and the single-round median on identical questions. Matching superforecasters is not the target.
For protocol cost, no reported result was found on Lacuna for cryptographic execution of a Delphi round. PINE's 0.13% proof overhead at $d=10^7$ does not transfer to our setting of roughly 80 ballots of 9 entries. We will report absolute latency and bytes instead.
Why not just…?
Why not just trust the coordinator under IRB oversight? For most clinical panels this is adequate, and we do not claim otherwise. The case for the protocol rests on panels where trust is the problem: competing firms setting shared benchmarks (the setting in A Consensus Privacy Metrics Framework for Synthetic Data), or contested policy questions. It also rests on the empirical question of whether participants behave differently when anonymity is verifiable rather than promised. If the behavioral study finds no difference in dissent retention or accuracy between the trusted-coordinator and full-protocol arms, the honest conclusion is that the protocol matters only for adversarial-coordinator threat models. We would say so.
Why not just add differential privacy noise to the feedback? The published feedback is the histogram itself. Noising counts in a panel of 13 to 80 people at the near-zero $\epsilon$ that the synthetic-data panel deemed interpretable would routinely flip whether an item crosses a 67% or 75% threshold. Noise also does nothing about a coordinator who sees raw ballots or plants arguments. We will report how often Laplace-noised histograms at several $\epsilon$ values change the consensus classification on the real panel data. That gives a direct, citable measure of the damage.
Why not just let an LLM moderate and summarize the arguments? An LLM moderator is still a coordinator whose framing can steer the panel. Elicitation Format Drives Divergent LLM Geopolitical Forecasts reports a 44.3% contradiction rate and a governance gap that swings from +0.228 to −0.158 depending only on question format. We do use a language model, but only locally and with author approval, for style normalization. Because the same framing sensitivity could distort an argument's meaning during rewriting, semantic fidelity is measured explicitly.
Why not just use commit-and-publish hashes? That baseline is in the experiments. It catches miscounted numbers but leaves the argument channel fully under coordinator control. The manipulation-injection experiment below is designed to show whether that gap matters.
Proposed Experiments
Cost comes first. We implement the protocol and measure end-to-end time per round, bytes per expert, and proof generation time on a laptop. The range is 13 to 80 panelists and 12 to 100 items, matching the reported panel sizes of 13, 75, and the VICCCS domains. The bar for practicality is that a round completes in minutes, not hours, on ordinary hardware.
The text-leakage experiment asks whether style normalization actually hides authors. We collect justifications from the behavioral study, plus prior writing from consenting participants. We train an authorship classifier per pool and report top-1 attribution accuracy against chance (1/N), before and after the local rewrite. Blind raters judge whether each rewrite preserves the argument. If attribution stays far above chance after rewriting, the text-anonymity guarantee fails. We would then report credentials as protecting authenticity but not anonymity.
The behavioral study is the decisive test. Panels of 15–25 recruited forecasters run three-round Delphi on ForecastBench questions that resolve within the study window. They are randomized across four arms: trusted coordinator, commit-and-publish, the protocol without credentials, and the full protocol. A single-round crowd median serves as the no-iteration reference.
We measure three outcomes. The first is final-round Brier score. The second is dissent retention: the fraction of round-one outliers who remain outside the interquartile range, split by whether they turned out right. The third is manipulation susceptibility. In every arm, the experimenters inject one fabricated, persuasive argument per selected item in round two, presented as coming from a dissenting panelist. In the full protocol that argument cannot carry a valid credential, and its unbadged status is visible to panelists. We then measure the round-three median shift toward the planted position.
The proposal's central bet is twofold. First, the full protocol should show a smaller manipulation shift than both the trusted-coordinator arm and the no-credential ablation. Second, correct dissenters should hold their position more often in the full protocol, lowering Brier score. If the full protocol and the no-credential ablation are indistinguishable, then outlier proofs add nothing beyond anonymity, and the credential machinery should be dropped.
References
- Delphi: A Cryptographic Inference Service for Neural Networks
- A Consensus Privacy Metrics Framework for Synthetic Data
- The Vascular Impairment of Cognition Classification Consensus Study
- Conservative Management and Biological Treatment Strategies: Proceedings of the International Consensus Meeting on Cartilage Repair of the Ankle
- PINE: Efficient Verification of a Euclidean Norm Bound of a Secret-Shared Vector
- Efficient Non-Interactive Zero-Knowledge Proofs in Cross-Domains Without Trusted Setup
- EVA-S3PC: Efficient, verifiable, accurate secure matrix multiplication protocol assembly and its application in regression
- The Honey Badger of BFT Protocols
- AI Agents and the Future of Deliberation: Designing Human-AI Collaboration for Democratic Dialogue
- ForecastBench: A Dynamic Benchmark of AI Forecasting Capabilities
- Agentic Forecasting using Sequential Bayesian Updating of Linguistic Beliefs
- Elicitation Format Drives Divergent LLM Geopolitical Forecasts