Logo

Quorum

Tell Quorum the issue. It generates your advocacy playbook.

A civic advocacy generator built in Jac for JacHacks SF 2026. Give it a civic action — a project, proposal, agenda item, or ballot measure — and Quorum generates what a community actually needs to be heard: a plain-language issue brief, a cited playbook, bilingual outreach drafts, and strategic advice on where that advocacy should actually be aimed.

Inspiration

When a government action harms a local community, the community usually loses — not from apathy, but from going down the wrong paths. Two real San Francisco cases show the two ways it happens.

Portsmouth Square, Chinatown. A ~$73M renovation included demolishing the historic park and bridge. Real opposition existed: the 176-year-old CCBA says it was never invited to project meetings; a five-decade board member opposed the removal; 150+ merchants protested. Yet the District 3 supervisor's office reported no communication in support of keeping the bridge. Outreach never reached the associations connected to the elderly residents who used the park daily. The community held a position; the decision-maker's record held nothing. Nobody generated the ask: submit this comment, in this format, to this person, before this date.

Prop K / Great Highway, Nov 2024. The measure to close the highway passed citywide at 54.7%, while the neighborhoods that used it voted heavily to keep it open — 60/40 in the Sunset, 70/30 in the Richmond. The Great Highway supporters weren't disorganized, they spent $269,000, filed a CEQA challenge, and ran an active campaign. But their own post-mortem was blunt: they never went east. Nobody told them the one thing that mattered: your outreach was reaching people who already agreed, while the persuadable audience was the citywide electorate.

Both sides had motivation, networks, even money. What they lacked was output and advice — a drafted, deadlined, correctly-addressed submission, and a strategic read on who actually needed persuading.

What it does

Point Quorum at a civic action. It generates:

A plain-language issue brief — what's happening, who's affected, what's contested — pulled from local and in-language press, forum threads, and public records.
A procedural playbook — who decides, who they answer to, where to submit, by what deadline. Every fact carries a source URL, or is flagged as an explicit gap: "No verified source found. Quorum will not guess."
Three recourse paths — recorded_comment, decision_zone_contact, and escalation. Each is source-backed; none is ever marked legally reviewed.
Targeting advice — Quorum compares the impact zone against the decision zone and flags when they diverge: "You are not the electorate for this decision — here's who is." The sentence the Prop K opposition paid $269,000 to learn too late.
Precedent-based strategy — matched by how a community lost, not what got built. Boston's Parcel C garage fight informs Portsmouth Square because both involved an in-language community excluded from an official process.
A bilingual outreach packet — per-organization English and Chinese drafts built around the playbook's real, deadlined ask.
A campaign dry-run spec — an outreach and media plan, staged for human review, never sent automatically.

The target user is a volunteer at a family association who shares her community's position, knows the elders who share it, and has no idea a comment period closes in eleven days. Quorum hands her the finished draft and tells her where to send it.

How we built it

Generation grounded in a fact graph. A confident wrong deadline costs an organizer the one resource they can't recover — so Quorum splits the work. An LLM writes briefs, proposes evidence bindings, and drafts outreach copy. Every procedural fact underneath is retrieved, not generated, into a persistent Jac graph — with a source URL or an explicit gap.

The pipeline is one-way: structured APIs and scrapers feed a cache, which feeds extraction, which feeds one schema-validated file (fixture.json), which feeds the Jac graph, which walkers traverse to produce the playbook and outreach packet. Nothing downstream calls back upstream.

Fetch: Socrata (DataSF 311) and Legistar go first since they're structured and free; Firecrawl handles pages with no API (agendas, press); Apify pulls Reddit for context only, never testimony. Everything is cache-first and budget-guarded, so the whole pipeline runs offline.
Extract: four extractors turn raw fetches into records, and their real job is to refuse bad data — dead links get filtered, unclear dates become "unknown" rather than guessed, private individuals get reduced to roles.
Contract: fixture.json is the only interface downstream — seven collections, every record source-tagged and timestamped. An empty comment-channel field isn't missing data, it's a finding.
Graph: the fixture materializes into a Jac workspace — 15 node types, 13 edge types, across seven semantic layers. Fetched edges (located_in, decided_by) are automatic; judgment edges (accountable_to, inside_process) are hand-curated and never model-set. Where a link can't be verified, the graph gets an explicit ProcedureGap node instead of a guess.
Walk: six walkers assemble the output with no HTTP or model calls inside the traversal. DivergenceCheck is the core trick — it forks one traversal toward impact zones and another toward decision zones, and the mismatch is the advice; no single node stores it.
Validate: RecommendRecourse checks any LLM-proposed paths against a strict schema (three paths, three steps each, real source IDs) and discards the whole thing for a deterministic fallback if anything fails. It never returns fewer than three paths.

The honesty boundary is structural, not a prompt: EvidenceMatcher can only write potentially_relevant_to — no satisfies or proves edge type even exists — and legal_reviewed is hardcoded false. The generator can't claim a legal standard is met, structurally.

A Jac fullstack UI renders the output verdict-first, with citation chips and live/cached badges. Nothing is hardcoded, and nothing sends automatically.

Challenges we ran into

Real procedure resists resolution. Rec & Park's and the D3 supervisor's comment channels never resolved — every candidate agenda page was dead by scrape time. DecisionWalker reports a ProcedureGap instead of guessing, and in the lead case, the gap proves the thesis.
311 data lies by category. "Portsmouth Square" returns 994 cases, 609 of them routine maintenance. SF's 311 taxonomy has no category for project opposition, so we shipped an honest zero.
Frozen upstream data. Legistar's API is stuck at ~Dec 2018; we built a portal scraper for current data and disclosed the split.
Building substantially in Jac. Keeping graph and validation logic in a young language, while hitting a 40%+ Jac source-line share, meant compiling constantly and reading compiler output instead of assuming.

Accomplishments we're proud of

Zero fabricated procedural facts: 3 verified comment channels, 9/9 organizations resolved or explicitly blanked, 22 sourced testimony records.
Targeting advice that's computed, not scripted — it comes from a graph fork, not a field.
Gaps reported as output, not hidden — the refusal is what makes the rest trustworthy.
Offline-by-design: 42/42 tests pass with no network.
New cases are YAML config, not code.

What we learned

Communities don't need another awareness tool, they need the finished output: the drafted comment, the named recipient, the deadline, and one honest piece of strategic advice.
Generated advice needs a retrieval floor. Structural constraints (no edge to write) beat prompt-level promises.
Object-Spatial Programming fits when the advice lives in the topology — divergence is a comparison of traversal endpoints, not a field.

What's next for Quorum

Wire remaining generators fully into the UI.
Richer per-path advice: expected effort, historical success signals, what "enough" input looks like.
DataSF live mode with cache fallback.
A Chinese-language entry point and text zoom.
More cases as verified manifests.