Flow Diagrams
Three maps. Each answers a question. Each answer feeds the next. The loop is the CONDUCTOR in the inner-loop P&ID — it chains every other instrument.
What does success look like?
Four outcomes. Each proves a different layer.
OUTCOME MAP — Autoresearch Loop ═══════════════════════════════════════════════════════════════ Outcome Measure Baseline → Target Timeline ───────────────────────── ───────────────────────── ────────────────── ──────── North Stars queryable PRDs with Queryable: Yes 0/7 → 7/7 30 days Full cycles completed PRDs pain→proof/month 0 → ≥1/month 60 days Return signal wired PRDs with actual_metric 0 → ≥1 45 days Morning reports generated Sessions with struct log 0 → ≥1/week 30 days
Outcomes name what we're measuring. The value stream shows where the work flows.
Where does value flow?
Current state: 7 manual triggers. Target: 1 trigger, full cycle.
VALUE STREAM — Current State (7 manual triggers)
═══════════════════════════════════════════════════════════════
pain-signal create-prd engineering proof-to score-prds
-extractor → (manual) → -handoff → -story → (manual)
│ │ │ │ │
▼ ▼ ▼ ▼ ▼
Pain card PRD with Spec + Meta Reranked
(Comms) 5P scores tests article board
BUT prose BUT no BUT no BUT no
metrics queryable frontmatter confidence
check update boost
───────────────────────────────────────────────────────────────
Wait time between steps: hours to days (human availability)
Total cycle time: weeks to never
VALUE STREAM — Target State (1 trigger)
═══════════════════════════════════════════════════════════════
LOOP-ORCHESTRATOR (one trigger)
│
├─→ metric-definer ──→ measure-north-star ──→ outcome-validator
│ (LOOP-001) (LOOP-002) │
│ Enforce formula Return scalar ▼
│ proof-to-story
│ (LOOP-003)
│ Write to frontmatter
│ │
│ ▼
│ score-prds upgrade
│ Confidence boost
│ │
├─→ session-experiment-logger (LOOP-004) ◄──────┘
│ Aggregate → morning report
│
└─→ HALT if: budget exceeded | metric regressed 2x | forbidden outcome
───────────────────────────────────────────────────────────────
Wait time between steps: seconds (agent handoff)
Total cycle time: hours (overnight session)The value stream shows the flow. The dependency map shows the build order.
What depends on what?
Build order: Phase 1 (gauge) → Phase 2 (return wire) → Phase 3 (conductor).
DEPENDENCY MAP — Build Order
═══════════════════════════════════════════════════════════════
PHASE 1: THE GAUGE (can we measure?)
┌─────────────────┐ ┌──────────────────────┐
│ LOOP-001 │────→│ LOOP-002 │
│ metric-definer │ │ measure-north-star │
│ Enforce formula │ │ Return scalar via CLI│
└─────────────────┘ └──────────┬───────────┘
│
PHASE 2: THE RETURN WIRE (do outcomes compound?)
▼
┌─────────────────┐ ┌──────────────────────┐
│ LOOP-003a │────→│ LOOP-003b │
│ proof-to-story │ │ score-prds upgrade │
│ Write to front- │ │ Confidence boost for │
│ matter │ │ validated PRDs │
└─────────────────┘ └──────────┬───────────┘
│
PHASE 3: THE CONDUCTOR (does one trigger run all?)
▼
┌─────────────────┐ ┌──────────────────────┐
│ LOOP-004 │────→│ LOOP-005 │
│ session-logger │ │ loop-orchestrator │
│ Morning report │ │ Chain all + safety │
└─────────────────┘ └──────────────────────┘
First target: Sales CRM (won_deals / total_deals > 0.30)Questions
What's the minimum viable loop — the smallest set of wires that produces one full cycle?
- Can LOOP-001 + LOOP-003 deliver value without the full orchestrator?
- Is Sales CRM the right first target, or should we start with a simpler metric?
- What safety rails prevent the loop from producing false greens overnight?