The honest starting point
If you're evaluating Civic Filament as a "Pol.is alternative," it means you already care about the right things. You've seen surveys and comment boxes fail. You've read about vTaiwan. You want deliberation that scales, and you want to see the shape of what a community actually thinks — not just a number for and against.
Pol.is did that first. Any honest comparison has to start there.
This post lays out what we take from Pol.is, where we differ, and — most importantly — when the difference actually matters for your deployment. If Pol.is is the right tool for your situation, we'll say so.
What Pol.is does, in one paragraph
Participants write short statements about an issue. Everyone votes agree / disagree / pass on other participants' statements. A dimensionality-reduction algorithm (PCA / UMAP) plots participants in 2D based on how they voted. Groups emerge; statements that cross groups are surfaced as "group-informed consensus." The output is an interactive map plus a list of statements ordered by cross-group agreement.
This is a genuine advance over surveys. It captures nuance a Likert scale can't. And it made the mechanics of "the majority thinks X" versus "the majority in this group thinks X" legible in a way legislators can actually reason about.
What Civic Filament inherits from Pol.is
Three things, without qualification:
- ◇Consensus is cross-group, not average. The Pol.is insight — that the interesting agreement is the one that cuts across otherwise-different clusters — is the same insight that drives our Living Requirement Document. We separate Strong Consensus from Genuine Divergence rather than averaging them into a "45% support" headline.
- ◇Divergence is signal, not noise. Pol.is preserves the map. So do we. Averaging positions to make disagreement disappear is not a supported operation in either tool.
- ◇The output is legible. A Pol.is opinion map and a Civic Filament LRD are both artifacts you can hand to a legislator or a resident. They are not black boxes.
We also credit the broader vTaiwan lineage — Pol.is-plus-in-person-plus-government-commitment — which changed the political imagination about what "structured deliberation" could produce. Our Prior Art page walks through that lineage in more detail.
Where the two tools diverge
The difference is the input layer.
Pol.is captures written statements and votes on statements. Civic Filament captures a conversation and derives structured signal from it.
That sounds small. It changes what you can actually see about a community.
1. Fidelity of the input
A Pol.is statement is written by the participant. It has to survive their own self-editing (do I want to be seen saying this?), their writing comfort (English fluency, dyslexia, screen size), and their attention span. Statements that are ambivalent, conditional, or emotionally complicated tend not to get written.
A Civic Filament interview is a 3–7 turn conversation guided by Motivational Interviewing and Nonviolent Communication principles. The mediator asks open questions, reflects back what it heard, and probes for the value or need beneath a stated position. The participant doesn't have to arrive with a fully-formed opinion. They can arrive with a feeling, and the interview surfaces what's under it.
That fidelity matters most when the community's "no" is actually a conditional yes — and vice versa. The classic case: a resident who says "I oppose the bike lane" but actually just cares about parking access for the small businesses they love. Pol.is can't always distinguish those from a hard "no." A structured interview is designed to.
2. What gets clustered
Pol.is clusters positions. Civic Filament clusters values.
Both work. After each interview, Claude rates the perspective on 16 named value dimensions, and we cluster those value profiles (DBSCAN, with a k-means fallback). Value-level clustering is what lets us name a consensus that's hidden under different vocabularies. "Defund the police" and "back the blue" can both contain people who fundamentally want accountable neighborhood safety. If you cluster on the stated position, they land far apart. If you cluster on the underlying values, they can land next to each other — and when enough participants share that concern, the LRD names the overlap as a real point of consensus, not a paradox.
3. Access and equity
Pol.is is written participation. It skews toward comfort with English text and typing on a screen. Community organizers know this pattern; it's the same access filter that town-hall attendance has.
Civic Filament is a conversational interface that runs in the browser, including on a phone, with voice input. The mediator currently works in English only — multi-language support isn't built yet. That doesn't erase access bias — it changes its shape. A shy resident, a tired parent at 11pm, a non-writer — our bet is that a conversation is an easier way in for them than writing statements, though we haven't yet measured that in a real deployment.
4. Auditability of the interview itself
The Pol.is algorithm is open source and its clustering is legible. But the statements — the raw input — are generated by whoever happens to write. There's no methodology attached to the input.
Civic Filament's mediator has an explicit, versioned system prompt. What the AI is instructed to do — and never do — is published on the Methodology page. Every synthesis version publishes a Merkle root over its inputs, and each participant gets a receipt they can use to verify, in the browser, that their perspective was included. If you're a procurement officer, this is the difference between "an emergent AI" and "an AI operating under rules you can inspect."
When Pol.is is the right tool
Genuine answer: when your community is comfortable with written participation, when your issue is well-suited to short statement-voting, and when you have the process wrapper (facilitators, in-person meetings, government commitment) that made vTaiwan work.
If you have a Pol.is deployment running well, don't break it. The two tools can compose — you could run Civic Filament interviews to seed the initial statement pool for a Pol.is round, or run Pol.is to prioritize which issues get the deeper interview treatment.
When Civic Filament is the right tool
- ◇You're a municipality where written-comment participation has stalled, and you want a way to reach residents who don't show up to town halls or comment on the portal. See our free pilot for how to try this with one issue.
- ◇You need an artifact for procurement. A Living Requirement Document is a versioned document that legal, engagement, and elected officials can all reason about.
- ◇You have compliance and records-law constraints. Data minimization is built into the product: interview transcripts aren't stored on our server — they stay in the participant's browser and auto-delete after 30 days — and the server keeps only an extracted summary linked to a hashed anonymous session ID. Participants can withdraw or delete their data. Retention, FOIA posture, and consent language for a governmental deployment would still need to be agreed in writing (see Participant Rights).
- ◇You want the buyer FAQ answered up front. Ours is here.
What we don't claim
We don't claim Civic Filament is a strict superset of Pol.is. We don't claim adoption we don't have — we're a v0.1.0 beta with no live deployments or pilot results yet, and the changelog is public.
We do claim that if you're asking "what does my community actually want?" and the previous answer has been either "we ran a Pol.is round" or "we ran a survey," an LLM-native interview layer is worth 30 minutes of your attention. The best way to see whether it is for you is the rate-limited real interview, not this post.
---
This is part of the Prior Art lineage series, which credits the tools Civic Filament stands on and names where we differ. If we've mischaracterized Pol.is anywhere, please tell us via the contact page — we'll correct it here.