Verdict & MVP Roadmap | Geopolitical Trade Pipeline

YES, build this.

Cost-viable at $50-150/month. Adds a deep geopolitical analysis layer to the existing research ecosystem. NanoClaw provides autonomous orchestration, Supabase integration is trivial. Phase 1 (triage + alerts) deployable in 4-6 weeks, full pipeline over 12 months.

MVP Roadmap

Phase	Timeline	Monthly Cost	Capability	Models
Phase 1: Triage Only	Months 1-2	$25/mo	RSS scan + alerts	Gemini Flash
Phase 2: Deep Dive	Months 3-4	$75/mo	Full analysis on trigger	Flash + Opus
Phase 3: Auto-Rec	Months 5-6	$125/mo	Position sizing + intent	Flash + GPT-5
Phase 4: Autonomous	Months 7-12	$150/mo	Full pipeline	Flash + GPT-5/Opus

Phase 1: Triage Only (Months 1-2, $25/mo)

Builds on top of the existing RSS pipeline. NanoClaw runs a triage agent that scans geopolitical feeds and generates WhatsApp alerts. No auto-trade. The goal is to validate that the triage model catches real events with acceptable precision. Monitor false positive rate and adjust scoring threshold.

Phase 2: Deep Dive on Trigger (Months 3-4, $75/mo)

High-score events trigger NanoClaw to spawn full deep-dive research agents in container isolation. Output: structured analysis JSON pushed to Supabase. Human reviews analysis before any trade decision. Validates research quality and signal accuracy.

Phase 3: Auto-Recommendation (Months 5-6, $125/mo)

Pipeline generates trade recommendations with position sizing. Intent is generated but requires human approval before execution. Tests the full pipeline end-to-end with a human gate.

Phase 4: Full Autonomous (Months 7-12, $150/mo)

NanoClaw runs the full pipeline autonomously. Findings flow to Supabase and are picked up by the weekly research combiner and portfolio rebalancer automatically. High-conviction findings (>85%) feed directly into rebalancing. Full autonomous mode only after Phase 3 proves reliable.

What Could Go Wrong

Risk	Impact	Mitigation
False positives	Triage triggers on noise, wastes API spend on unnecessary deep dives	Phase 1 validates triage precision before enabling deep dives. Adjust score threshold based on empirical false positive rate.
Analysis quality	LLM hallucination in research leads to bad trade recommendation	Review agents (Wave 2 in Iran architecture) catch factual errors. Human gate in Phase 2-3 catches remaining issues. Objectivity audit agent verifies source diversity.
Latency	50-80 min from event to trade may miss fast-moving opportunities	Acceptable for geopolitical/macro events which play out over days to weeks. For genuinely time-critical events, the alert path (~24 min) enables manual fast action.
Cost creep	Multiple events in one day exceed daily budget	Hard daily cost cap in the orchestrator. Queue excess events for next-day processing. Priority scoring ensures the highest-impact event gets analyzed first.
Model degradation	Provider model updates change output quality or format	Structured output schemas with validation. LLM Judge benchmarks detect quality drift. Model fallback chain (Opus -> GPT-5 -> Sonnet) if primary model degrades.

Next Steps

Set up NanoClaw on a local Mac Mini (or existing machine) with 24/7 operation
Configure geopolitical RSS feed group in NanoClaw with CLAUDE.md context
Write triage agent prompt (Gemini Flash) with relevance/urgency scoring
Write deep-dive research agent CLAUDE.md files (one per research domain)
Build post-processor: IPC JSON results to Supabase geopolitical_findings INSERT
Create geopolitical_findings table in Supabase with finding record schema
Wire weekly research combiner to include geopolitical findings from Supabase
Test on historical events: would the system have caught the Iran strikes buildup?
Phase 1 deployment: NanoClaw triage + WhatsApp alerts, no auto-trade