convocantsystems.com

Agora runs community meetings end to end

Agora turns multi-sector community meetings into structured, accountable action. That means opioid task forces, recovery coalitions and behavioral health planning bodies. It runs the whole life of an event: the outcomes agreed before the room convenes, consent-gated capture and facilitation support while it is in session, and afterwards a human-reviewed record where every claim traces back to the transcript span it came from. Each published report compounds into a longitudinal store scoped to that community, under a privacy architecture written for HIPAA and 42 CFR Part 2 environments from the start.

Request a demo Start with the lifecycle

One event, from consent to permanent record

Seven stages, in order. Each one is a real component rather than a box in a sales diagram. A human gate sits between what the system derives and what the community publishes, and there is no path around it.

Illustrative

0%
The lifecycle drawn as one pass. The band advances through the six stages on its own; choose a stage to hold it there. Illustration of the production pipeline — no participant data.

  1. 01

    Consent-gated capture

    Recording starts only within the consent posture set for the event, and a visible recording indicator stays up for as long as it runs. Nothing is captured silently.

  2. 02

    Streaming editable transcript

    The transcript appears as the table talks and the facilitator corrects it in place. A dropped connection buffers locally and reconciles on reconnect, so a dead spot in a church basement costs the record nothing.

  3. 03

    Provider-swappable transcription and embedding

    Both sit behind interfaces. That makes the vendor a deployment decision rather than an architectural one, and it puts the BAA gate somewhere it can actually be enforced.

    TranscriptionProvider EmbeddingProvider impl Voyage store pgvector

  4. 04

    Rolling signal extraction

    While the conversation runs, the analysis engine derives emerging priorities, named barriers, action proposals and a conversation trajectory. The engine proposes; the facilitator keeps, edits or discards.

  5. 05

    Human review gate

    Drafted outputs stay drafts until a named person reviews them and publishes under their own name. There is no auto-publish path anywhere in the system.

    Draft In review Published

  6. 06

    Commitments and reminders

    Accepted proposals become commitments with owners, due dates, reminders, overdue escalation, and completion-rate reporting by sector. An action item stops being a line in minutes nobody reopens.

  7. 07

    The longitudinal record

    Every published report writes into a community-scoped store: recurring barriers over time, priority evolution, completion trends by sector. The strata are what a briefing is later assembled from.

Before

Organizers design the event: tables by sector, a facilitation framework, invitations, a consent posture, and a Desired Outcomes Document stating what the meeting is for. The outcomes are parsed by AI and then reviewed and approved by the organizer. It is the approval that makes them the reference frame for everything that follows, not the parsing.

During

Each table’s facilitator runs encrypted, consent-gated capture while signals accumulate and get curated. The meeting facilitator watches every table at once and sees the patterns no single table can see. A separate display wall shows only what has been explicitly published.

After

Agora drafts the executive summary, ranks barriers alongside the sectors that raised them, and produces sector findings and de-duplicated proposals. Every claim traces back to its transcript span. A named person reviews and publishes, accepted proposals become tracked commitments, and the report joins the community’s longitudinal record.

What the room can see while it is still in the room

Three surfaces run at once during an event, and they are deliberately kept separate. What a table sees, what the meeting facilitator sees, and what the wall shows are divided in the architecture rather than in a settings panel.

The table cockpit

Each table runs its own view. The trajectory panel never shows a bare score. It shows a trajectory word and the transcript spans that justify it, so a facilitator can disagree with the machine using the same evidence the machine used.

Priority, barrier and proposal chips are derived by the engine and then kept, edited or discarded by the person running the table. Curation is a deliberate human step rather than a review queue nobody has time for.

Stylised illustration in our design language, not a screenshot. Trajectory words are always shown with the evidence behind them.

The facilitator dashboard

The meeting facilitator sees every table at once. Cross-table pattern detection is the reason it exists. When three tables independently name transportation, the facilitator finds out while the room can still act on it, instead of reading about it in a report three weeks later.

Alerts flag the three states a facilitator would otherwise have to walk the room to find: a table that has stalled, a table that has gone silent, and a table that has asked for help.

Stylised illustration in our design language, not a screenshot. Table names and states are invented for the drawing.

The display wall

The wall is fed by a separate, structurally isolated publishable path. Only explicitly published, non-sensitive aggregates can enter it.

Raw transcript, un-curated signals and speaker attribution have no route to the wall at all. This is not a permission they are denied. There is no code path that would carry them there.

Stylised illustration in our design language, not a screenshot. The isolation it draws is architectural.

Outcome-guided intelligence

A meeting is only on track relative to what it was called for. Agora makes that comparison continuously and mechanically, against a document a person approved before the room convened.

The reference frame

The approved Desired Outcomes Document is the reference frame. It is parsed by AI and then approved by the organizer. Nothing becomes an outcome because a model proposed it.

Continuous comparison

As the conversation runs, chunks are embedded and compared to those outcomes in the same semantic space. The embedding layer is provider-swappable; the current implementation is Voyage on pgvector.

The scorecard

End-of-meeting analysis produces an Outcomes Scorecard: a per-outcome achievement judgment with cited transcript evidence, plus cross-table theme clustering. Facilitators never see a bare score.

Illustrative

Conversation chunks as points in semantic space, clustered into themes, with the approved outcomes as anchors. Drag to orbit. Synthetic points, generated in your browser — no real meeting content appears here.

The conversation state engine watches process health while the process is still running

The engine is built and has been evaluated against fixtures taken from real sessions. It rests on the Cynefin framework: a domain classifier places the conversation in the domain it is actually in, hybrid outcome alignment runs alongside it, and eight risk detectors watch for documented facilitation failure modes.

These are observations about the process rather than judgments about people. Every one of them is something an experienced facilitator would notice, assuming they could sit at all six tables at once.

Guidance surfaces

Nudge cards in the facilitator view and guidance panels on the dashboard are the delivery layer for everything above. They are still in development and have not shipped.

A nudge advises the person running the table. It never claims the AI understands the room.

The eight risk detectors

Select or hover a detector to read the failure mode it watches for. Both faces are always present for a screen reader.

Phragma: barrier intelligence as a versioned taxonomy

A barrier described differently at every meeting cannot be counted, compared or resolved. Phragma gives barriers stable names and stable codes, so a record from one county in 2024 is comparable with a record from another in 2027.

33 Phragmata

Named barrier classes, each with a stable versioned code.

7 Families

The kinds of constraint a coordination failure can be.

9 Poroi

Resolution vectors. Every phragma routes to a primary one, most to a secondary as well.

3 Lenses

Change, entry and understanding: the same barrier read three ways.

Illustrative

The taxonomy as a map: seven family hubs, thirty-three barrier classes, and the nine resolution vectors they route to. Hover or focus a vector to light the barriers it resolves. Codes, names and routings from the canonical taxonomy v0.1.

The seven families

  • LLegal & Regulatory
  • FFinancial & Incentive
  • IInformational & Technological
  • TTrust & Relational
  • CCultural & Paradigmatic
  • SStructural & Operational
  • PPolitical & Accountability

The nine poroi

  • POR-1 CodifyWrite the agreement down.
  • POR-2 BrokerInsert a human connector.
  • POR-3 AlignRestructure incentives and metrics.
  • POR-4 InstrumentBuild the technical bridge.
  • POR-5 ConveneCreate shared governance.
  • POR-6 TranslateBuild shared language.
  • POR-7 ProbeRun safe-to-fail experiments.
  • POR-8 AdvocateRoute the barrier upward.
  • POR-9 RitualizeTrust through repeated small commitments kept.

Shadow barriers

A shadow barrier has a precise definition: present in the transcript evidence, absent from the formal outputs. It was named in a side comment, in a qualifier, or in the sentence somebody said before the note-taker started typing.

These are systematically under-captured, and they are frequently the constraint that actually binds. Surfacing them is what Phragma is for.

Workaround signals are treated as two things at once: evidence that a barrier is real enough for someone to have routed around it, and a seed for the resolution, because somebody in the room has already built part of the answer.

The human-confirmation rule

No barrier finding is attributed to a named organization in any output before a human analyst confirms it. The system can record that a constraint was described. Deciding whose constraint it is stays with a person, permanently.

PH-L1 PH-F2

Codes are immutable. A retired barrier is deprecated, never renumbered, so a 2024 record still means in 2027 what it meant when it was written. Adding a class takes three confirmed instances that fit nothing already in the taxonomy, across two communities.

Where Phragma actually is

Only the first of these four phases exists. The other three are on the roadmap, in the order we intend to build them.

  • Phase A Taxonomy as versioned configuration, plus the analyst workflow. Built and in use.
  • Phase B Automated batch detection over completed events. Not built.
  • Phase C Data signatures and longitudinal barrier tracking. Not built.
  • Phase D Live in-meeting barrier surfacing. Not built, and hard-gated.

We gate our own features. Live surfacing ships only after proven precision on real deployments and a recorded human decision to go ahead.

What the coalition still has in three years

A single meeting is not the unit of value. What matters is what accumulates: which barriers recur, which priorities moved, which commitments closed, and which sector closed them.

The longitudinal store

Every published report writes into a community-scoped intelligence store: recurring barriers over time, priority evolution, and completion trends by sector.

The briefing generator

It answers one question: what has this coalition accomplished, and where is it stuck? Every claim in a briefing resolves to the intelligence records and events behind it, and a claim with nothing behind it does not get made.

Grant Platform export

A versioned export contract turns community evidence into grant evidence. An export can be generated only from a briefing that has already been published, and the export itself is logged as a disclosure.

The community keeps what it learned without keeping the recording

Transcripts carry a retention window set by the event’s consent posture. Before that window expires and the transcript is crypto-shredded, a preservation pipeline pulls out the durable knowledge: decisions, barriers, commitments and insights, each with its verbatim evidence. That goes into an equally protected archive with its own per-record keys. The retention sweep fails closed, so if the preservation run has not completed, the shred does not happen. When the deletion machinery breaks, it breaks by refusing to delete, which is the only safe direction for it to fail in.

A deletion promise gets kept, and the community still keeps what it learned.

Illustrative entries

chain unverified

Press Verify chain to re-derive every hash in this browser.

Every hash below is a real SHA-256 of the entry shown, computed in your browser. Verify re-derives the chain; tamper alters one entry and re-derives it again. Metadata only, as in the product: sequence, action, actor, hash. No content, no participant, no utterance.

How the privacy guarantees are actually enforced

Meeting audio at recovery-sector events is treated as 42 CFR Part 2-protected from the first line of code. Below are the mechanisms that do the enforcing, named individually. The full architecture, including the threat model and what we deliberately do not claim, is on the trust page.

  • AES-256-GCM

    Application-layer envelope encryption with per-record keys. Protection travels with the record rather than with the disk it happens to be sitting on.

  • deny-by-default

    Every sensitive read passes a policy decision point: gate, log, return. A read that is not explicitly permitted does not happen, and the attempt is written down anyway.

  • SHA-256 hash chain

    An append-only audit log where each entry commits to the one before it. Tamper-evident, metadata only, and verifiable by anyone holding the log through GET /audit/verify.

  • BAA gate

    The Business Associate Agreement is enforced as code. At runtime the system refuses to send protected data to any vendor without a recorded signed agreement on file.

  • consent gate

    Capture is gated on the event’s consent posture and shows a visible recording indicator for as long as it runs. Nothing is captured silently.

  • crypto-shred

    Deletion destroys the per-record key, which makes the ciphertext unrecoverable. Deletion is a physical fact about the data, not a flag on a row.

  • derived sensitivity

    Embeddings of sensitive text inherit the sensitivity and the deletion cascade of their source. A vector is not a loophole.

Read the full privacy and trust architecture

What is built, what is not

A deliverable is marked done only when its automated evidence checks pass. This table reflects verified progress rather than intentions.

Delivery status
Milestone What it delivers Status
Core milestone ladder M0–M8 The platform spine: event design, consent-gated capture, live analysis, human review and publication, commitments, and the longitudinal store. Built
Collaborative Strategic Planning (Exhibit C) The planning workbook, complete and in use as a facilitation framework. Built
Knowledge preservation and retention enforcement Retention windows, crypto-shredding, and the preservation run a shred cannot proceed without. Built
Desired Outcomes framework The outcomes document an organizer approves before the event, which becomes the reference frame everything is compared against. Built
Embedding foundation Provider-swappable embedding pipeline; Voyage implementation on pgvector. Built
End-of-meeting analysis Outcomes Scorecard with cited transcript evidence, plus cross-table theme clustering. Built
Conversation State Engine Cynefin domain classifier, hybrid outcome alignment, and eight risk detectors, evaluated against fixtures from real sessions. Built
Phragma Phase A The barrier taxonomy as versioned configuration, with the analyst confirmation workflow. Built
OGI guidance surfaces Nudge cards in the facilitator view and guidance panels on the dashboard. In development
MOA mode The settlement-specific mode for Memorandum of Agreement reporting. In development
Longitudinal Planning Series chaining Linking a sequence of planning events into one continuous series. In development
Phragma Phase B Automated batch barrier detection over completed events. Roadmap
Phragma Phase C Barrier data signatures and longitudinal barrier tracking across events and years. Roadmap
Phragma Phase D Live in-meeting barrier surfacing. Hard-gated behind proven precision on real deployments and a recorded human go decision. Roadmap
Reasoning Graph track Explicit representation of how a conclusion was reached across a series. Roadmap
Poros, Convocant Grant Platform, Phoreo, Oikos The rest of the portfolio. The Grant Platform export contract is already built into Agora. Roadmap

What you can see in a demo today

The whole lifecycle, running. Design an event with tables, a facilitation framework and a consent posture; write a Desired Outcomes Document and approve it; run consent-gated capture on a table with a streaming, editable transcript, and pull the network out to watch it buffer and reconcile.

Then: curate signals as a facilitator, open the all-tables dashboard and see cross-table pattern detection with stalled, silent and help alerts, and project the display wall. End the meeting and generate the Outcomes Scorecard with cited evidence and cross-table theme clustering. Review the drafted summary, ranked barriers and de-duplicated proposals, follow any claim back to the transcript span it came from, and publish it as a named reviewer.

Afterwards: turn accepted proposals into tracked commitments, generate a briefing from the longitudinal record and export it against the versioned Grant Platform contract, verify the audit chain, and run a retention sweep that refuses to shred until preservation has completed.

Request a working session

How “done” is decided

A milestone is not complete because someone says it is. It is complete when its automated evidence checks pass in the build, and the status column above reads from that result.

Anything this page describes as in development or on the roadmap is exactly that. We are not selling it as shipped.