Quarantined Feedback: Keeping Sybils from Steering Open Delphi Panels

The Delphi method gathers opinions from a panel of anonymous experts over several rounds. After each round, the panel is shown summary statistics (the median rating, the spread, anonymized rationales), and each expert may revise. This feedback is what makes Delphi converge, and it is also a weak point once panels move online or include AI agents. Suppose an attacker adds fake accounts ("Sybils", meaning one actor posing as many) or organizes a small group of colluding members. They can shift the median that everyone else sees in round one. Honest panelists then drift toward that contaminated median in later rounds. By the final round, removing the fake votes no longer recovers the honest answer, because the honest votes themselves have moved. Existing defenses against fake identities in voting only guarantee protection in a single, one-shot vote.

We propose to measure this amplification effect in multi-agent and human–AI Delphi panels, and to test a defense taken from how critical infrastructure gets upgraded. When an organization opens a vetted panel to a wider pool, it would not merge the two. It would run them as parallel tracks. A small verified core panel receives feedback computed only from itself, plus open-track rationales that pass a filtering gate. The open track can overturn the core's position only by a supermajority. The question is whether this dual-track design keeps fake accounts from steering the outcome ("safety") while still letting a genuinely better-informed open panel correct the core ("liveness"). We also want to know whether it beats the obvious fix of applying robust statistics at every round.

Core Claim

Our first claim is that output-time Sybil resilience does not carry over to iterative deliberation. Sybil-Resilient Reality-Aware Social Choice proves that trimming extreme votes before taking a median (its Suppress Outer-σ rule), anchored on a privileged status quo, preserves safety and liveness for a Sybil fraction σ < 1/3. That analysis treats the honest votes as fixed. In Delphi they are not fixed. They respond to the feedback.

A simple anchoring model makes the point. Let panelist $i$ hold rating $x_i^r$ in round $r$, and let panelists move a fraction $w$ toward the fed-back median $m^r$ each round:

$$x_i^{r+1} = (1-w)\,x_i^r + w\,m^r.$$

If Sybils shift the fed-back median by $\Delta$ per round, then after $R$ rounds the honest ratings themselves have shifted by about $(1-(1-w)^R)\,\Delta$. This follows from summing the geometric series of per-round pulls. Any trimming applied only to the final ballots therefore aggregates an already-contaminated honest population.

The update rule is a modelling assumption, not a law. Part of the work is to measure $w$ empirically for LLM panelists and for humans.

Our second claim is that the natural defenses trace out a Sybil–collusion frontier. Deduplicating near-identical rationales and capping each cluster's influence defeats cheap Sybils that repeat themselves. It does not stop a few real colluders who write diverse, mutually reinforcing arguments. Tightening the gate enough to catch them also suppresses legitimate open-track dissent, which is the liveness failure. This mirrors the impossibility results for referral networks, where no reward scheme is both proof against fake downstream identities and proof against coalitions merging (Threats and Trade-Offs in Resource Critical Crowdsourcing Tasks Over Networks; Collusion-proof And Sybil-proof Reward Mechanisms For Query Incentive Networks). We borrow it here as an empirical lens, not as a theorem we transfer.

Our third claim is that quarantining feedback moves this frontier outward. Separating the channel that forms honest beliefs (feedback) from the channel that aggregates votes (the final decision) should shift the frontier relative to single-track designs.

Research Program

The design copies the structure of Secure Upgrade of Hardware Security Modules in Bank Networks. In that paper, upgraded bank security modules form trusted "islands," and designated border modules translate traffic between islands and legacy nodes. It also copies the dual-track validation of A Novel Balanced Reporting Innovation with Data Governance Evolution (BRIDGE) Method, where the old workflow keeps running as a reference while the new one is checked against it. Neither paper considers strategic participants. We apply their architecture to an adversarial consensus setting.

The verified core plays the role of the island. It is a small panel, like the six-expert panels that The Human-AI Hybrid Delphi Model found sufficient for thematic saturation. The open track is the legacy side: large, unverified and potentially infiltrated. A gate sits between them and does three things:

  1. It clusters open-track rationales by semantic similarity.
  2. It forwards at most a fixed number of representatives per cluster to the core.
  3. It passes no open-track numbers into the core's feedback statistics.

The final decision uses the reality-aware rule, with the core's consensus as the status quo. The open track overturns that consensus only with a δ-supermajority computed after Suppress Outer-σ trimming.

We will build this protocol and a round-by-round attack harness, and fit the anchoring rate $w$ from logged revisions. The attack harness covers three attacker types. Split attackers create many identities with templated or paraphrased rationales. Collusion attackers use a few identities with coordinated, diverse arguments. Hybrid attackers switch between the two across rounds. For each attacker type and budget, we will map the frontier of safety versus liveness under each defense.

Baselines

The naive reference point is a standard single-track Delphi. All participants are pooled, the untrimmed median and interquartile range are fed back each round, and a plain median is taken at the end. It shows how bad the undefended amplification is. Its clean-panel accuracy also fixes the ceiling that any defense should approach when no attacker is present.

The strongest existing aggregation method is Belief-Calibrated Consensus Seeking from Belief-Calibrated Multi-Agent Consensus Seeking for Complex NLP Tasks. It weights agents by their generation confidence and pairs uncertain agents with confident dissenters. That paper reports 80.60% accuracy on MATH and 78.47% on MMLU in a clean, cooperative setting. It is the most capable iterative consensus procedure in the file, and it has no adversary model. We will reimplement it with our own base models and compare within our setup, not against the published figures. The comparison is fair only if it uses the same agents, the same number of rounds and the same query budget.

The strongest Sybil-aware baseline is the reality-aware rule of Sybil-Resilient Reality-Aware Social Choice, in two forms. The first applies it only to final ballots, as the paper does. The second, "robust feedback," applies Suppress Outer-σ trimming to the statistics fed back in every round. No reported result was found on Lacuna for any Sybil-resilient rule under multi-round Delphi feedback. Beating robust feedback is the real bar for this proposal.

Two ablations isolate the key ingredient. The first runs the dual-track design but lets gated open-track numbers into the core's feedback. It tests whether quarantining the numbers, rather than gating the text, is what matters. The second runs the design with no cluster cap. It tests the Sybil side of the frontier.

Every method gets the same hyperparameter search over its trimming, supermajority and weighting parameters. That search is tuned on clean and attacked development items, and results are reported on held-out items. Our design beats the baselines if, at matched clean accuracy, it cuts attacker-induced outcome shift by a clear margin across attack budgets up to σ = 1/3. It must also keep liveness, meaning the rate at which a correct open majority overturns a wrong core, within a few points of the ablation without quarantine.

Why not just…?

Why not just apply robust trimming to the feedback in every round? This is the serious threat. Our amplification argument predicts it still drifts. Sybils placed just inside the trimming boundary pull the median a bounded amount each round. As honest ratings move, the trimming window moves with them, so the bounded pulls add up over rounds. That prediction may be wrong for small $w$ or few rounds. The robust-feedback baseline is run at every attack budget and round count. If it matches the dual-track design, we will report that per-round trimming suffices, and the quarantine becomes unnecessary machinery.

Why not just keep the panel verified-only? A verified-only core is immune by construction but can be wrong. Its error rate is exactly what liveness measures. We will build the core from weaker or fewer agents than the open pool, so that the open track carries information the core lacks. We then report how often the open track corrects the core under each defense. If the open track never corrects the core, opening the panel has no value, and the honest conclusion is not to open it.

Why not just weight by confidence, as BCCS does? Confidence is claimed by the agent itself, and a Sybil can report maximal confidence for free. BCCS's collaborator assignment and leader selection would then preferentially route honest agents toward the attacker. The attacked BCCS runs test this directly.

Why not just run a single round with no feedback? Without feedback there is no amplification. The Human-AI Hybrid Delphi Model reports consensus on 92.3% of items in one round with six senior experts. This is a strong alternative, and we include it with final-ballot trimming. Its cost is the correction that iteration provides. On items where honest panelists start split or wrong, single-round accuracy should fall below iterative accuracy. If it does not, the case for iterative Delphi under adversaries weakens considerably.

Proposed Experiments

The first testbed uses MMLU and MATH, as in Belief-Calibrated Multi-Agent Consensus Seeking for Complex NLP Tasks. Ground truth exists there, so safety and liveness are measurable. We will recast each item as a Delphi rating task, in which agents rate their confidence in candidate answers over rounds. Attackers push a fixed wrong answer.

The second testbed uses the European Parliament deliberation scenarios from EuroCon and PoliCon. Those papers already show consensus success falling from 87–93% under simple majority to 52–63% under two-thirds majority, along with a bias toward the dominant party. Here safety is measured as divergence from the outcome of a clean run, and Rawlsian scores track whether the gate silences minorities.

The decisive checks are three:

  1. We will verify that final-ballot trimming alone fails as rounds increase. The honest-population shift should track the $(1-(1-w)^R)\Delta$ prediction using the fitted $w$.
  2. We will show whether the dual-track design strictly dominates robust per-round feedback on the safety–liveness frontier for split, collusion and hybrid attackers separately. We expect it to lose ground specifically against diverse-argument collusion, and we will report that loss.
  3. A small human pilot will run an e-Delphi on insomnia-style clinical rating items, with confederate Sybil accounts, after the HAH-Delphi prospective design. It will test whether human anchoring rates resemble LLM rates. If humans anchor much more weakly, the amplification threat is mostly a property of AI panels. That would change where the defense is needed.