The product

Not another meeting summarizer.

Threadline is a real-time, privacy-aware collective reasoning environment. Zoom or another meeting platform remains the media plane. Threadline becomes the understanding, validation, provenance, and continuity plane.

Speech-to-text entry plane

Conversation enters naturally through local transcription, typed notes, reactions, and structured UI controls.

Complementary AI roles

A personal thought partner helps the individual. A shared facilitator helps the group. They operate across separate trust domains.

Decision and evidence fabric

Ideas, risks, questions, findings, decisions, and artifacts retain explicit links to the signals that produced them.

A meeting is an event stream where people produce evidence, interpretations, ideas, commitments, and decisions.

Threadline preserves each layer separately. AI generates proposals and projections over the record. It does not silently redefine the record.

Experience model

Capture. Think. Understand. Preserve. Choose.

The experience follows the natural shape of conversation. Structure is available when useful. It is never imposed as the destination of every discussion.

1Capture naturallyLocal speech-to-text, typed input, reactions, imported context, and exact timestamp anchors.
2Think privatelyPersonal AI can consider notes, memories, dissent, and selected prior conversations without exposing them.
3Understand togetherThe shared facilitator surfaces topics, gaps, risks, opportunities, evidence, and positive motion.
4Preserve whyFindings remain traceable to transcript spans, participant signals, evidence, and human validation.
5Choose what followsIdeas, agenda items, experiments, decisions, actions, RFCs, ADRs—or no formal outcome yet.

Brainstorming stays divergent

A successful session may end with an idea landscape, unresolved tensions, an accepted uncertainty, or a future agenda. Threadline does not force convergence.

Facilitation stays complementary

The AI notices gaps, suggests questions and methods, and highlights progress. It evaluates the state of the conversation—not the quality of the people.

Trust model

Private influence is not shared evidence.

Users control whether prior conversations, one-to-ones, personal notes, or vent sessions may influence their private reasoning. Those sources do not enter shared provenance unless the participant deliberately publishes an admissible contribution.

Personal cognition plane

Private notes, one-to-ones, personal memory, selected conversations, private dissent, and participant-selected AI providers.

Use privatelyMay influence personal reasoning.
Allowed
Increase shared confidencePrivate evidence cannot silently strengthen a group finding.
Blocked
Generate candidate contributionSafe abstraction may be proposed locally.
Review
Participant-controlled
promotion boundary

Collective sensemaking plane

Shared transcript events, reactions, explicit stance, evidence, ideas, gaps, findings, summaries, and optional governed outcomes.

Shared transcript and UIInspectable by authorized meeting participants.
Admissible
AI proposalClearly marked until validated.
Tentative
Validated findingSupported by inspectable signals and human review.
Traceable
Speech is natural. UI establishes intent.

The system may know which device captured an audio span without knowing who spoke. Participants can claim, confirm, correct, react to, or comment on exact spans. Explicit declarations outrank inferred verbal agreement.

Purpose and boundaries

What Threadline must accomplish—and what it refuses to become.

Primary goals

  • Capture conversation with minimal friction.
  • Help individuals think without exposing private context.
  • Help groups see topics, gaps, risks, ideas, and progress.
  • Preserve evidence-backed findings with provenance.
  • Support human validation, correction, dissent, and action acceptance.
  • Recommend useful SDLC structures at the right time.
  • Promote memory deliberately rather than accumulating it automatically.

Explicit non-goals

  • Replace Zoom or the media platform.
  • Score employee behavior, dissent, or reasoning quality.
  • Act as an autonomous meeting chair or decision-maker.
  • Assume device identity equals speaker identity.
  • Store every participant’s model credentials centrally.
  • Automatically create Jira or Confluence records in the first release.
  • Force every discussion into decisions, actions, or documents.
Capabilities

A conversation model broad enough for real work.

Threadline represents more than transcript paragraphs and action items. It captures the knowledge structures that groups naturally create while preserving the freedom of ordinary conversation.

Sensemaking

Ideas, observations, questions, hypotheses, assumptions, constraints, themes, tensions, and open branches.

Evaluation

Evidence, counterevidence, risks, opportunities, confidence states, unresolved gaps, and validation.

Coordination

Agenda items, initiatives, decisions, accepted actions, dependencies, ownership, and review triggers.

Continuity

Cross-session memory, supersession, decision lineage, future context packs, SDLC artifacts, and learning reviews.

Evidence-backed findings

A summary is composed from structured findings. Each material statement exposes supporting signals, counter-signals, uncertainty, human validation, and source lineage.

Real-time gap register

Definition, evidence, assumption, authority, ownership, risk, scope, validation, exploration, and closure gaps move through explicit lifecycle states.

SDLC guidance plane

Suggests one primary artifact or method—RFD, RFC, ADD, ADR, experiment, threat assessment, ORR, runbook, or learning review—without forcing workflow.

Red-team view

The hard problems are trust problems.

The technology is feasible. The product succeeds only if it remains honest about attribution, privacy, inference, uncertainty, records management, and human authority.

ChallengeFailure modeRequired control
Consent and surveillanceMeeting capture becomes invisible monitoring or performance evaluation.Visible capture state, explicit consent, pause controls, scoped retention, no participant scoring.
Unknown speakerDevice capture is treated as proof of human identity.Unknown-by-default attribution; explicit claim, confirmation, correction, and cross-talk states.
AI false authorityA polished summary appears more certain than its evidence.Findings and signals remain inspectable; uncertainty and counter-signals remain visible.
Private-context leakageA private one-to-one or vent session appears in shared synthesis.Isolated indexes, restrictive derivation labels, local disclosure review, explicit promotion gate.
Repetition biasFrequently repeated ideas appear better supported.Separate mention count, unique arguments, independent evidence, authority, and decision influence.
Forced ownershipAI assigns actions from ambiguous speech.Separate proposed assignment from explicit owner acceptance.
Prompt injectionConversation text manipulates models or tools.Treat transcript as untrusted data; constrain tools; require confirmation for side effects.
Records exposureHigh-fidelity meeting history becomes an unmanaged legal record.Retention classes, deletion, legal hold, regional storage, administrative audit.
Prompt fatigueFacilitation becomes a distracting second meeting.Materiality gate, quiet and checkpoint modes, facilitator-private suggestions, progressive disclosure.
Reference architecture

A layered system with explicit trust boundaries.

The browser is the collaboration surface. A native or local companion provides reliable capture, personal model access, private memory, and the participant trust boundary. The central platform stores shared events, projections, evidence, and policy-governed memory.

Threadline reference architecture Participant devices connect through a real-time session plane to an append-only event and provenance plane. Shared AI, trust controls, and integrations operate over authorized context. Personal AI remains inside the participant boundary. Participant device plane Local capture, private context, personal AI Microphone + local STT Audio span anchors, local buffering Private workspace Notes, memory, one-to-ones, dissent Policy-aware retrieval Personal AI bridge Bedrock, Claude CLI, Copilot CLI Local credentials and tool policy Participant promotion gate Keep private Publish abstraction or question Attach selected shared evidence Withhold antecedent provenance Real-time session plane Identity, sockets, ordering, replay WebSocket gateway Presence, authorization, rate limits Sequencing + deduplication Canonical server order, causal links Reconnect + replay Local buffering, acknowledgments Identity model Person ≠ account ≠ device Entra ID + scoped guests Unknown-speaker support Evidence, provenance, and shared record Append-only events with derived projections Event ledger Immutable history Supersession Auditability Projections Transcript, topics Findings, summaries Actions, decisions Lineage graph Source → signal Finding → summary Artifact → outcome Intelligence, control, and integration AI proposes; policy constrains; humans authorize Shared facilitator Topics and themes Gaps and progress Findings and summaries SDLC guidance Control plane Consent and authz Context admissibility Retention and audit Model and tool policy Integrations Zoom transcript Jira + Confluence Repositories Export + write-back
Reference architecture, version 0.1. The central server receives shared events and authorized context. Participant-private memory and personal model credentials remain inside the local or enterprise-controlled participant boundary.
Product requirements document

The contract for what we build.

The PRD is organized around outcomes, product invariants, functional epics, and measurable release gates. The initial release proves trust before expanding automation.

Primary users

Individual participants, facilitators, subject-matter experts, architects, product leaders, and security or governance stakeholders.

MVP objective

Validate that participants trust and use a shared, threaded conversation record with explicit private and shared boundaries.

Core invariant

No material summary statement exists without a traceable finding, and no material finding exists without inspectable signals.

Epic A — Identity, sessions, and participation
Entra ID and scoped guests; stable participant identifiers; distinct person, account, session, and device identities; explicit consent; roles for organizer, facilitator, contributor, observer, decision owner, and artifact owner.
Epic B — Conversation entry plane
On-device microphone capture and STT; local buffering; WebSocket synchronization; canonical server sequencing; overlapping speech; capture-source attribution; transcript versioning; pause and resume.
Epic C — Threaded provenance
Timestamped audio spans become stable thread anchors. Participants can react, comment, claim speaker identity, correct text, add evidence, support meaning, dispute interpretation, raise questions, or attach private notes.
Epic D — Personal thought partner
Private workspace; selected prior context; personal model adapters; safe question formulation; novelty detection; disclosure review; explicit publication of abstractions, questions, or selected evidence.
Epic E — Context admissibility
Independent controls for retrieval, reasoning, transformation, publication, citation, provenance disclosure, retention, and memory promotion. Derived content inherits the most restrictive source policy until reviewed.
Epic F — Shared facilitator
Incremental summary, topic detection, idea clustering, question and risk extraction, gap register, positive-motion tracking, suggested inquiry, suggested method, and optional SDLC recommendation.
Epic G — Evidence-backed findings
Findings expose supporting and countervailing signals, uncertainty, source lineage, model provenance, human validation, freshness, and supersession state. Confidence applies to the finding, never the participant.
Epic H — Decisions, actions, and SDLC guidance
Decisions record authority, alternatives, criteria, rationale, uncertainty, and review triggers. Actions require explicit acceptance. SDLC guidance recommends the smallest sufficient structure and explains why now, why not another artifact, and what remains missing.
Epic I — Memory and integrations
Memory tiers from private participant context to reviewed organizational battle scars; deliberate promotion; read-only Jira, Confluence, Zoom, and repository integrations first; governed write-back later.
MVP principle: prove trust before autonomy.

The first release focuses on local transcription, threaded spans, participant signals, private notes, incremental summaries, gap proposals, evidence-backed findings, provenance exploration, and read-only integrations.

Delivery strategy

Earn each layer of intelligence.

Threadline progresses from a trusted conversation record to shared facilitation, private thought partnership, SDLC integration, and cross-session memory. Each phase has a distinct trust gate.

Phase 1

Trusted conversation record

Sessions, identity, local STT, sockets, timestamped spans, reactions, comments, corrections, private notes, and exports.

Exit gateParticipants trust the record more than an unreviewed transcript.
Phase 2

Evidence-backed understanding

Topics, knowledge objects, findings, supporting and counter-signals, validation, and provenance exploration.

Exit gateParticipants can explain why every material summary statement exists.
Phase 3

Shared facilitation

Gap register, positive motion, topic checkpoints, suggested inquiry, and facilitator modes.

Exit gateSuggestions improve understanding without prompt fatigue.
Phase 4

Personal thought partner

Private retrieval, local AI bridge, context admissibility, disclosure review, and safe abstraction.

Exit gatePrivate context remains technically and perceptually isolated.
Phase 5

SDLC guidance

Artifact discovery, RFD/RFC/ADD/ADR and experiment recommendations, architecture context, and source-linked outlines.

Exit gateThe system recommends the smallest useful structure at the correct lifecycle point.
Phase 6+

Governed write-back and memory

Previewed writes, delegated authorization, project memory, supersession, context packs, and reviewed battle-scar promotion.

Exit gateEvery side effect and promoted memory is attributable, authorized, and auditable.
Validation

Measure trust, clarity, and follow-through—not transcript volume.

Record trustParticipant correction, dispute, and provenance-inspection patterns.
Attribution honestyPercentage of uncertain speaker spans represented without false identity.
Finding qualityProposed findings accepted, corrected, or kept unresolved.
Gap closureMaterial gaps resolved, deferred, or accepted explicitly.
Human controlActions with explicit owner acceptance and private publications reviewed.
Decision recoveryTime required to recover why a direction was selected.
SDLC linkageMaterial decisions connected to the appropriate lifecycle artifact.
Privacy integrityPrevented leaks, policy denials, and unauthorized context exposure.

Understand together.
Preserve why.

Threadline gives every participant a private place to think, every group a clearer way to understand, and every consequential finding a visible path back to its evidence.

Return to the vision Revisit the architecture