Delphi Process¶

Iterative anonymous elicitation method — instantiates Consensus Convergence

Converges a dispersed expert panel through repeated anonymous rounds, feeding back the group's distribution and reasons between each so judgments revise without status pressure.

A Delphi Process converges a panel of experts who never argue face to face. Each participant submits a judgment anonymously; a coordinator aggregates the responses and feeds back the group's distribution — the spread of estimates plus the anonymized reasons, especially the outliers' — and everyone submits again, revising in light of what the others thought and why. Its two distinctive commitments are anonymity and iteration: stripping identity from each judgment removes the status, seniority, and rhetorical dominance that distort a live room, and the repeated rounds let convergence emerge from reasons rather than from social pressure. Because outliers must justify themselves rather than fall silent, the method is also a test of whether apparent agreement is genuine or merely the loudest expert's view echoing.

Example¶

An energy agency needs a planning estimate for when grid-scale storage will fall below a cost threshold that makes a certain generation mix uneconomic. The experts who could answer are scattered across labs and firms, and in a live meeting the two best-known names would anchor everyone. So the agency runs a Delphi.

Round one: twenty experts each submit a year, anonymously, with a one-paragraph rationale. The estimates range across ≈12 years — far too wide to plan on. The coordinator returns the distribution and the reasons, without names, flagging the earliest and latest guesses and their arguments. Round two: several experts, having read a supply-chain argument they hadn't considered, pull their estimates in; two outliers hold their late position but now explain why — a materials constraint the others had underweighted. By round three the spread has narrowed to ≈3 years, and the two holdouts' reasoning is preserved as a documented risk rather than averaged away. The panel converged, and the surviving disagreement is one the planners can actually see.

How it works¶

Delphi's design does work that a meeting cannot:

Anonymity de-links judgment from identity. No one knows whose estimate is whose, so revision responds to arguments, not to who made them — the antidote to a room that converges on the senior voice.
Controlled feedback between rounds. Participants see the group's distribution and the anonymized rationales — not a running debate — so each round is a fresh, considered revision rather than an argument to win.
Iteration until movement stops. Rounds repeat until estimates stabilize; convergence is something the panel walks toward across rounds, and a persistent outlier who keeps justifying a position is a signal, not noise.

Tuning parameters¶

Number of rounds — typically two to four. More rounds tighten the spread but risk fatigue and can grind legitimate disagreement into false agreement.
Anonymity strength — fully blind versus partially attributed. Full blindness maximizes candor; some attribution can raise accountability for sloppy estimates.
Feedback richness — bare statistics (median and quartiles) versus statistics plus the full anonymized arguments. Richer feedback moves people for the right reasons but is heavier to run.
Stopping criterion — a fixed round count versus a stability threshold (rounds stop when the spread stops shrinking). A stability rule guards against converging just because the schedule said to.
Outlier obligation — whether extreme estimates must be defended in writing. Requiring it preserves minority reasoning and prevents silent capitulation.

When it helps, and when it misleads¶

Delphi is the tool of choice when the right experts are geographically or politically scattered, when status differences would poison a live discussion, and when the object is a judgment or forecast rather than a negotiated commitment. Its signature risk is that iteration can manufacture convergence: a panel pulled toward the running median across rounds may produce a tight distribution that reflects regression to the group's center rather than genuine agreement, and a coordinator who curates the feedback can steer the result.^[1] The guard is to make convergence earn itself — stop on a stability criterion rather than a round quota, require outliers to state their reasons rather than merely revise, and report the final spread alongside the central estimate so a narrow band isn't mistaken for certainty it hasn't earned.

How it implements the components¶

Delphi realizes the archetype's iterative, pressure-free convergence machinery:

iteration_cadence — the round structure is the mechanism: convergence is produced by controlled repetition, and the cadence (how many rounds, when they stop) is its central dial.
false_convergence_check — anonymity plus the obligation for outliers to justify themselves tests whether agreement is genuine reasoning or suppressed dissent; a spread that refuses to close flags real disagreement instead of hiding it.

It does not frame the question the panel answers (shared_question — Consensus Workshop), set what evidence is admissible (evidence_standard — Evidence Briefing Packet), or declare the estimate officially closed (closure_rule — Consensus Vote).

Instantiates: Consensus Convergence — Delphi is the anonymous, iterative route to convergence.
Sibling mechanisms: Scientific Consensus Process · Disagreement Register · Consensus Vote · Consensus Workshop · Consent Decision Rule · Evidence Briefing Packet · Facilitated Decision Meeting · Minority Statement Protocol · Stakeholder Alignment Session · Standards Committee Process

References¶

[1] The Delphi method was developed at the RAND Corporation in the 1950s–60s (Dalkey and Helmer) for technological forecasting. Its known weakness is that the round-and-feedback structure can pull a panel toward the running median regardless of merit, so a tight final distribution must be read as convergence of reasoning, not proof of accuracy. ↩