Implementation Phases

The current shipped runtime fully covers phases 0–5 and the recovery portion that phase 6 depends on. Continuous watch mode primitives exist in the codebase, but the website now reflects only the features that are actually wired into the runtime and API surface today.

Phase Summary

Phase 0: Architecture Spike

Exit criterion: A research run can be interrupted and resumed without duplicating completed stages.

• 20-node ADK 2.0 Workflow graph (sequential)
• Source and claim data schemas (14 Pydantic v2 models)
• Session persistence
• Event streaming infrastructure
• Web search tool integration
• Durable checkpoint/recovery is now implemented and wired into the persisted runtime

Phase 1: Evidence-First MVP

Exit criteria: All substantive report claims map to claim records. All source-derived claims map to evidence.

• Research Director, Perspective Planner, Question Architect
• Query Planner with 10 search strategies
• Evidence Curator with exact excerpt preservation
• Claim Builder with atomic claims and epistemic status
• Outline Architect and Section Writer
• Citation linking: report → claim → evidence → source
• REST API: create, inspect, export research runs

Phase 2: STORM-Grade Iterative Research

Exit criterion: Compute-matched evaluation demonstrates better coverage than one-shot RAG.

• Research Moderator Agent: stagnation detection, rebalancing
• Evidence-conditioned follow-up question generation
• Frontier scheduler with priority function: Priority = w_m·M + w_r·R + w_d·D + w_u·U + w_n·N − w_c·C
• Iterative stopping conditions: budget, coverage, info gain
• Perspective budget allocation
• Information-gain metrics: IG_t = α·N_c + β·N_e + γ·N_k + δ·C_r − λ·D

Phase 3: Co-STORM-Style Collaboration

Exit criterion: Users can steer a running investigation without corrupting workflow state.

• Streaming event infrastructure (16 event types)
• Live research frontier API
• User intervention API: add_question, challenge_claim, exclude_source, increase_depth
• Approval gates A/B/C/D with configurable triggers
• Knowledge Organizer Agent + concept map projection
• Mid-run scope changes
• Strict API approval pausing with persisted decisions and resume after approval

Phase 4: Epistemic Reliability

Exit criterion: Material unsupported-inference rate meets production threshold (<5%).

• Counter-Evidence Agent: contradiction search, independence checks
• Contradiction detection and resolution tracking
• Source independence clustering (deterministic)
• Confidence model: authority × independence × freshness × corroboration
• Verification Agent with independent prompt from writer
• Citation entailment with keyword overlap heuristics
• Cross-section consistency checks

Phase 5: Enterprise Governance

Exit criterion: Security, authorization, and audit release gates pass.

• Identity engine: principal context propagation through persisted runs and node execution
• Policy engine: role-to-tool allowlists, stage-based tool reduction, durable decision linkage
• Real source policy: deduplication, classification, domain filtering
• Immutable audit log with chain verification
• DB-backed workflow checkpoints, node execution records, and resumable runtime state
• Prompt injection defenses

Phase 6: Continuous Research and Reusable SOPs

Exit criterion: A saved research workflow can update prior findings without rebuilding blindly.

• The repository includes a workflow.watch module and related primitives for delta detection.
• Those pieces are not yet wired into the public API or documented as a completed product capability.
• The website therefore treats this phase as partial rather than claiming an end-to-end continuous research surface today.

Dependency Graph

Phase 0 ──► Phase 1 ──► Phase 2 ──► Phase 3 ──► Phase 5 ──► Phase 6
                         │            │            │
                         └──► Phase 4 ─┘────────────┘

• Phase 4 depends on Phase 2 (needs iterative loop)
• Phase 5 depends on Phase 3 (approval center) and Phase 4 (verification)
• Phase 6 depends on Phase 5 (governance + audit)