📄 icarus-phase-7-roadmap.md 7,018 bytes Apr 28, 2026 📋 Raw

🎯 Icarus Phase 7 — 'Cognition' Sprint

Status: Planned — Entry after Phase 6 UAT completion
Director: Matt
Theme: Cognitive layer, predictive intelligence, household optimization


Pre-Entry Requirements

Before entering Phase 7, the following must be validated:

Requirement Owner Criteria Status
Event Graph maturity Socrates 30 days stable operation, <1% data loss ⏳ Phase 6
Multi-user state sync Socrates Aundrea + Matt concurrent use, conflict-free ⏳ Phase 6
Phase 6 UAT complete Matt >90% document routing accuracy ⏳ In Progress
LLM token efficiency review Wadsworth Local vs cloud cost analysis ⏳ Phase 7 entry

Pillar 1: Hierarchical Briefing Model (HBM)

Source: shared/project-docs/hbm-spec-v0.1.md (deferred from Phase 6)
Owner: Daedalus (UX), Socrates (backend), Wadsworth (scheduler)

Scope

Four-level household briefing system based on cognitive psychology research:
- Month: Pattern density heatmap (aggregation only)
- Week: Anchor events + commitment cards (aggregation only)
- Day: When/Then chains + forgetting budget (Briefing Agent)
- Hour: Full action scripts (Briefing Agent)

Phase 7 Approach: Incremental

Sprint Deliverable Risk
7.1 Month + Week briefs (aggregation, no Agent) Low
7.2 Day brief — When/Then chains Medium (LLM cost)
7.3 Hour detail + Morning/evening notifications Medium
7.4 Forgetting budget optimization Low

LLM Cost Mitigation

  • Cache strategy: DayBrief cached until Event Graph changes
  • Target: >80% cache hit rate (most days = no LLM call)
  • Fallback: Simplified extraction if Briefing Agent fails

Sovereign Constraint

Daily Briefing Agent must support local execution (Gaming PC 3080 Ti):
- Primary: phi4:14b or qwen2.5-coder:7b for chain generation
- Fallback: Cloud (GLM-5.1) only if local fails
- Cost target: <$0.50/month via aggressive caching


Pillar 2: LLM Token Efficiency Review

Trigger: HBM cost analysis requirement
Owner: Wadsworth (analysis), Socrates (implementation)
Goal: Minimize cloud LLM costs via local model optimization

Analysis Scope

System Current Target Strategy
Document briefing (Icarus) Cloud (qwen3-vl:8b) Local (qwen3-vl:8b) Gaming PC via Tailscale
Recipe extraction Cloud (jina + LLM) Local Ollama phi4:14b
Newsletter parsing Cloud (qwen2.5:7b) Local Gaming PC
HBM Day Brief Cloud (GLM-5.1) Local first phi4:14b chain generation
Payment alerts Cloud (LLM classify) Local Keyword + light LLM

Cost Modeling

Scenario A: All Cloud (Current)
- Document briefing: ~$0.01 × 30 docs = $0.30/day
- Newsletter parsing: ~$0.02 × 2 newsletters = $0.04/day
- HBM Day Brief: ~$0.04 × 1 = $0.04/day
- Monthly total: ~$11.40

Scenario B: Local-First (Target)
- Document briefing: Local (Tailscale to Gaming PC) = $0
- Newsletter parsing: Local = $0
- HBM Day Brief: Local (80% cache hit) = $0.01/day
- Monthly total: ~$0.30 (99% reduction)

Implementation

  1. Benchmark local models on Gaming PC (tokens/sec, quality)
  2. Hybrid pipeline — try local first, fallback to cloud
  3. Token budgeting — per-request limits, max daily spend
  4. Cache everything — embeddings, generations, summaries

Deliverables

  • [ ] shared/project-docs/llm-efficiency-analysis.md — benchmark results
  • [ ] shared/api-specs/local-llm-routing.yaml — hybrid pipeline spec
  • [ ] Socrates: Gaming PC Ollama integration
  • [ ] Wadsworth: Cost tracking dashboard

Pillar 3: Multi-User State Synchronization

Prerequisite for HBM
Owner: Socrates

Problem

Aundrea marks permission slip signed at 21:45. Matt's 22:00 brief must reflect this.

Solution

Event Graph state versioning:

Event:
  state_version: int           # increment on change
  state_last_updated: ISO8601
  state_updated_by: person_id   # aundrea | matt
  state_history: [StateChange]  # audit trail

StateChange:
  timestamp: ISO8601
  person: person_id
  field: string                 # e.g., "permission_slip"
  old_value: any
  new_value: any

Sync Strategy

  • Optimistic updates — local change immediately, sync to Event Graph async
  • Conflict resolution — last-write-wins with notification of conflict
  • Offline support — queue changes, sync when reconnected

Deliverables

  • [ ] State version tracking in Event Graph
  • [ ] Conflict detection UI (Daedalus)
  • [ ] Sync status indicator ("Synced 2 min ago")

Pillar 4: Predictive Household Optimization

Beyond HBM — intelligent suggestions
Owner: Socrates (ML), Wadsworth (routing), Daedalus (UX)

Features

Feature Input Output Example
Carpool suggestion Event Graph + locations Optimized pickup schedule "Harper and neighbor both have OT Tuesday — carpool?"
Prep task bundling Week's prep_tasks Bundled errands "3 permission slips need signing — batch tonight"
Conflict prediction Pattern history Likely double-bookings "Friday 6pm historically overbooked"
Energy forecasting FamilyState patterns Cognitive load prediction "Tuesday looks heavy — front-load Monday prep"

Technical Approach

  • No new ML infrastructure — pattern matching on Event Graph
  • Heuristics first — only add ML if heuristics insufficient
  • Local computation — Gaming PC for any model inference

Phase 7 Success Criteria

Metric Target Measurement
HBM Week Brief Ship 7.1 Month + Week views live
HBM Day Brief Ship 7.2 When/Then chains working
LLM cost reduction 90% vs Phase 6 Monthly cloud spend
Cache hit rate >80% Briefing Agent invocations
Multi-user sync Zero conflicts State change log
Local LLM coverage >80% of calls Cloud fallback ratio

Dependencies from Phase 6

Phase 6 Deliverable Phase 7 Dependency
Event Graph (Icarus) Required for HBM
IMAP IDLE ingress Required for email-based state updates
Drop Zone share sheet Nice-to-have for document → Event Graph
Finance Intelligence Optional — expense tracking in HBM
Brain Query Optional — natural language → Event Graph queries

Timeline Estimate

Phase 7 Duration: 4-6 weeks
- Week 1-2: Month + Week briefs (Sprint 7.1)
- Week 3-4: LLM efficiency review + local model integration
- Week 5-6: Day Brief with When/Then chains (Sprint 7.2)
- Week 7+: Hour detail, notifications, optimization

Entry Criteria: Phase 6 UAT complete + Event Graph 30-day stability


The cognitive layer transforms data into actionable household intelligence.