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 |