en-US/about_AITriadDebate.help.txt

TOPIC
    about_AITriadDebate

SHORT DESCRIPTION
    The multi-agent debate system — BDI-structured context, AIF argument
    mapping, commitment tracking, synthesis, and harvest.

LONG DESCRIPTION
    AI Triad includes a structured debate system where AI agents argue from
    different POV perspectives, grounded in taxonomy data. Debates produce
    synthesis results that identify agreements, disagreements, and resolution
    preferences — which can then be "harvested" back into the taxonomy.

  DEBATE OVERVIEW
    A debate is a structured multi-agent conversation about a specific topic
    or thesis. Three AI agents (one per POV: accelerationist, safetyist,
    skeptic) argue their positions, moderated by a fourth AI agent that
    manages turn-taking, cross-responds, and synthesis.

    Debates progress through phases:
    1. CLARIFICATION — the moderator ensures the topic is well-defined.
    2. OPENING STATEMENTS — each POVer presents their initial position.
    3. CROSS-RESPOND — agents respond to each other's arguments (configurable,
       1-5 rounds).
    4. CLOSING STATEMENTS — each POVer summarizes their position.
    5. SYNTHESIS — the moderator analyzes the debate and produces structured
       findings.

  BDI-STRUCTURED TAXONOMY CONTEXT
    Each debater's prompt includes taxonomy nodes from their POV, organized
    using the Beliefs-Desires-Intentions (BDI) framework:

    BELIEFS (Data & Facts)
      Nodes categorized as empirical claims, statistics, research findings.
      These ground the agent's factual premises.

    DESIRES (Goals & Values)
      Nodes categorized as normative positions, values, policy goals.
      These define what the agent is arguing toward.

    INTENTIONS (Methods & Arguments)
      Nodes categorized as strategies, methods, argumentative approaches.
      These define how the agent constructs its case.

    VULNERABILITIES
      Known weaknesses in the POV's position (from steelman_vulnerability).
      Agents are aware of their own vulnerabilities but do not volunteer them.

    CROSS-CUTTING CONCERNS
      Cross-cutting nodes relevant to the debate topic, showing how each POV
      interprets shared concepts differently.

    The formatting function is formatTaxonomyContext() in
    taxonomy-editor/src/renderer/utils/taxonomyContext.ts.

  INCREMENTAL ARGUMENT NETWORK
    After each debater's turn, a background AI call extracts claims and maps
    relationships to prior claims. The resulting argument network
    (debate.argument_network) tracks:

    - Individual claims made by each agent
    - Relationships between claims (supports, contradicts, concedes)
    - Cross-references to specific prior statements

    The network is validated by word overlap and reference checks to ensure
    claims are grounded in actual debate content. The moderator uses the
    network to select cross-respond pairings.

    Prompt: taxonomy-editor/src/renderer/prompts/argumentNetwork.ts

  PER-DEBATER COMMITMENT TRACKING
    Derived from the argument network, the commitment store tracks each
    agent's epistemic state:

    - ASSERTED claims — positions the agent has put forward
    - CONCEDED claims — points the agent has acknowledged from opponents
    - CHALLENGED claims — positions the agent has actively disputed

    Commitments are injected into debater prompts via formatCommitments()
    with an explicit consistency rule: "do not silently contradict prior
    assertions." This prevents agents from flip-flopping or making
    inconsistent arguments across rounds.

    Stored in debate.commitments, keyed by POV.

  SYNTHESIS
    After all debate rounds complete, the moderator produces a structured
    synthesis (SynthesisResult) containing:

    agreements Points all three POVs agree on
    disagreements Points of genuine conflict, with analysis
    key_insights Novel insights that emerged from the debate
    argument_map AIF-structured claim network (optional)
    preferences Resolution analysis for each disagreement
    bdi_layer Which BDI layer the disagreement operates at
    resolvability Whether the disagreement is resolvable

  AIF ARGUMENT MAP
    The argument_map field in synthesis uses AIF (Argument Interchange Format)
    structure:

    Claims have:
      claim_id Unique identifier
      claimant Which POV made the claim
      text The claim text
      attack_type rebut | undercut | undermine
      scheme COUNTEREXAMPLE | DISTINGUISH | AUTHORITY | etc.
      targets Claim IDs this claim attacks or supports

    Attack types:
      Rebut — directly denies the conclusion
      Undercut — challenges the inference from premises to conclusion
      Undermine — challenges the truth of a premise

  PREFERENCE RESOLUTION
    Each disagreement in the synthesis is evaluated for which argument
    prevails, judged by specific criteria:

    Criterion Meaning
    ------------------ -----------------------------------------------
    empirical_evidence Supported by more/better empirical data
    logical_validity Logically sounder argument structure
    source_authority Backed by more authoritative sources
    specificity More specific and concrete than the opponent
    scope Broader applicability or wider implications

    Preferences include a rationale explaining why one argument prevails.
    Rendered as "Resolution Analysis" in synthesis cards.

  DEBATE HARVEST
    After synthesis, the "Harvest" button allows promoting debate findings
    back into the taxonomy. Five harvestable item types:

    Item Type Target
    --------------------- -----------------------------------------------
    Conflicts conflicts/*.json (new conflict records)
    Steelman refinements node steelman_vulnerability field
    Debate references node debate_refs array
    Preference verdicts conflict verdict field
    New concepts taxonomy proposals

    All items are user-reviewed with AI-generated descriptions validated
    before being written. Harvest manifests are stored in
    ai-triad-data/harvests/.

  DEBATE-SPECIFIC AI MODEL
    Each debate can override the global AI model via debate_model on the
    DebateSession object. Set in the New Debate dialog, persisted and
    restored on load. getConfiguredModel() checks debate model first,
    then falls back to the global setting.

  FACT-CHECK WITH WEB SEARCH
    The Taxonomy Editor includes fact-checking via Gemini's google_search
    tool (generateTextWithSearch). Web evidence is stored in metadata and
    viewable via the "Show Web Evidence" button on fact-check cards.

  POVER CHAT (1-ON-1)
    Distinct from multi-agent debates, POVer Chat is a 1-on-1 conversation
    with a single POV agent. Three modes:

    brainstorm Exploratory, high-creativity — generates novel ideas
    inform Pedagogical, grounded — explains concepts from the POV
    decide Analytical, Socratic — helps evaluate options

    Mode is switchable mid-chat via header pills. Chat sessions persist in
    chats/chat-<id>.json. Uses the same taxonomy grounding and BDI context
    as debates, but lighter — no clarification phase, no multi-agent
    coordination.

  POWERSHELL DEBATE CMDLETS
    Invoke-AITDebate Run a debate from the command line
    Show-DebateDiagnostics Inspect debate internals and quality metrics
    Show-DebateHarvest Review and apply harvest items
    Repair-DebateOutput Fix malformed debate JSON output

  EXAMPLE: RUNNING A DEBATE
    # From the command line
    Invoke-AITDebate -Topic 'Should AI development be paused?' `
        -Rounds 3 -Model 'gemini-2.5-pro'

    # From the Taxonomy Editor
    # 1. Click "New Debate" in the toolbar
    # 2. Enter a topic or thesis
    # 3. Select the AI model (or use global default)
    # 4. Configure cross-respond rounds (1-5)
    # 5. Click "Start Debate"
    # 6. After synthesis, click "Harvest" to promote findings

SEE ALSO
    about_AITriad
    about_AITriadTaxonomy
    Invoke-AITDebate
    Show-DebateDiagnostics
    Show-DebateHarvest
    Repair-DebateOutput
    Show-TriadDialogue