Infra-bot

Infra-bot generates grounded public market-intelligence memos from Musashi's truth layer.

It reads directly from the Musashi Supabase/Postgres truth layer and produces:

• deterministic source packets
• a strict gated publish packet
• a flagship daily memo in Markdown + JSON + PDF
• optional deep case-study memos only when an event cluster clears the publish gate

Outputs are deterministic Markdown + JSON files under outputs/.

What it is for

This repo is not a trading bot.

It is a content/report automation layer that sits on top of Musashi infra and answers questions like:

• what did Kalshi reprice the most in politics / macro / crypto?
• what does Kalshi imply about AI / crypto / Fed / elections right now?
• how did one event cluster move over time?
• what do resolved markets suggest about trustworthiness by category / liquidity?

Pipeline

Infra-bot v2 is a three-stage pipeline:

1. Deterministic source packets
   • movers_packet
   • sector_packet
   • market_structure_packet
   • case_study_packet
2. Strict publish gating
   • suppresses thin/noisy sections
   • omits sectors that do not clear minimum thresholds
   • blocks weak case studies
3. Narrative synthesis
   • uses validated publish packets only
   • keeps numerically sensitive section bullets deterministic
   • uses the LLM only for constrained framing/title/summary

Setup

1. Copy .env.example to .env
2. Fill in the DB credentials for the Musashi Supabase pooler connection
3. Install dependencies

npm install
npm run check

Commands

Generate the deterministic daily packet set:

npm run generate:daily

Generate one online flagship daily memo directly from live deterministic packets:

npm run generate:deep-daily

Generate one flagship daily memo from already-written JSON packets:

npm run generate:deep-daily-from-cache -- --date 2026-05-13

Generate one deep case-study memo. This command fails if the cluster does not clear the case-study gate:

npm run generate:deep-case-study -- --event-id YOUR_EVENT_ID

Export every Markdown artifact for a given day to PDF:

npm run export:daily-pdfs -- --date 2026-05-13

Generate only movers:

npm run generate:movers

Generate a sector snapshot:

npm run generate:sector -- --sector crypto

Generate a deterministic case-study source packet:

npm run generate:case-study -- --event-id FED-SEP-2025

Generate market-structure / trust analysis:

npm run generate:market-structure

Output shape

Each command writes both:

• Markdown
• JSON

to date-scoped folders under outputs/.

The flagship public artifact is:

• deep-daily-YYYY-MM-DD.md
• deep-daily-YYYY-MM-DD.json
• pdf/deep-daily-YYYY-MM-DD.pdf

Supporting packet files are internal/source artifacts and are kept primarily for validation, reuse, and debugging.

Scheduled automation

The GitHub workflow in .github/workflows/generate-content.yml runs the daily pack on schedule and uploads outputs as an artifact.

This workflow needs these repo secrets:

• SUPABASE_DB_HOST
• SUPABASE_DB_PORT
• SUPABASE_DB_NAME
• SUPABASE_DB_USER
• SUPABASE_DB_PASSWORD

For the LLM-backed memo path, set:

• LLM_PROVIDER=deepseek
• DEEPSEEK_API_KEY
• optionally DEEPSEEK_BASE_URL
• optionally DEEPSEEK_MODEL

The scheduled workflow now:

1. generates deterministic source packets
2. synthesizes the flagship memo from cached packets
3. exports Markdown artifacts for that day to PDF

Accuracy model

The system is intentionally split so that:

• raw numbers come from Musashi truth-layer data
• the publish gate decides what is strong enough to surface
• deterministic bullets carry numerically sensitive facts
• the LLM only writes constrained framing on top of validated facts

The daily memo is therefore a market-implied intelligence product, not a claim of external-world causal certainty.