Elicit, never persuade.

A Claude Sonnet 4.6 mediator asks open questions using Motivational Interviewing (OARS: Open questions, Affirmations, Reflections, Summaries) and Nonviolent Communication (observation → feeling → need → request) techniques. Length adapts to what has been captured so far — roughly 3 to 7 participant turns — and when loaded language is detected a debiasing instruction is added to the mediator’s prompt. Both are computed server-side on every turn.

Before any participant text reaches the model or is stored, structured identifiers are redacted by pattern: emails, phone numbers, SSN-style numbers, card numbers, street addresses and URLs. Names cannot be reliably detected, so the extractor is instructed not to carry names or identifying details into the summary, and participants are asked not to share them. The transcript itself stays in the participant’s browser; it is not stored on our servers.

Every mediator reply carries a server signature over the conversation so far, bound to the issue and to the participant’s anonymous session. A perspective is only extracted from a completed interview whose every mediator turn verifies, so transcripts can’t be fabricated, edited, or replayed under another session. Per-network daily limits (and an optional bot check) cap how many interviews one connection can complete.

A single Claude Sonnet 4.6 structured (tool-use) call extracts the stated position; stance (support / oppose / conditional / undecided); values, each tagged with a moral foundation; needs; and conditions, each with how much meeting it would shift the participant’s support. Two scores are added without AI: a Discourse Quality Index — a keyword heuristic inspired by Steiner’s DQI — and Schwartz value tags from a keyword classifier.

Claude Haiku 4.5 rates each perspective from 0 to 1 on 16 named value dimensions. If rating fails for any perspective, the entire run falls back to TF-IDF text vectors — the two are never mixed within a run. These are ratings on named dimensions, not semantic embeddings.

Perspectives are clustered with DBSCAN using an automatically tuned epsilon. If DBSCAN finds no usable clusters, the run falls back to k-means. No group with fewer than 5 members is published.

Claude groups the needs and values participants named into concerns that express the same underlying point, and returns which items belong to each. Support is then counted from those participants — the model never asserts a percentage. Strong consensus is a concern raised by at least 70% of participants and at least 5 people. Conditional consensus compares “support without” (participants whose stance is already support) with “support with condition” (those plus participants who named that condition and said it would shift them substantially, ≥ 0.5 on a 0–1 scale); it is reported only when the condition was named by at least 5 people and the difference is at least 5 points.

Genuine divergence is reported for pairs of groups that each hold at least 15% of participants (and at least 5 people) and whose value profiles differ by more than 0.5 on some named dimension. It is not computed when the TF-IDF fallback was used, because those vectors have no named dimensions.

Representation compares who participated with who lives in the community. After their interview, participants may optionally give an age band, housing tenure and area; when the deploying organization has entered population figures for those groups, each version flags underrepresented groups and reports the stance split reweighted to the population (raking, with weights trimmed so no small group dominates). Groups of fewer than five people are never shown. Weighting corrects who took part, not self-selection within a group, so it narrows but does not remove uncertainty. Without both inputs the section reads “Not measured”.

Each counted share carries a Beta-Binomial 95% credible interval (normal approximation). The margin of error shown is simple-random-sample sampling error. Participation is self-selected, so that margin is a lower bound on the real uncertainty, not a full account of it.