Remaining Build Tasks — NBA Pipeline

Date: 2026-03-31 Status: System functionally complete. These items need building. Context: Full NBA analyst pipeline was built in a marathon session. 3 blind analysts (Sonnet 4.6, Gemini 3.1 Pro, gpt-oss-120b) set lines from fundamentals, comparison layer classifies convergence, Opus makes final verdict. All specs, rulings, scrapers, matchup cards, APIs, MCP server, council skill, and cron jobs are done.

PRIORITY 1: Fix Before Going Live DONE (2026-03-31)

1.1 Fix Cron Timing DONE

Changed cron.schedule('30 18 * * *') to cron.schedule('30 17 * * *') in scheduler.ts line 604. Log message updated to "5:30 PM". EdgeClaw restarted.

1.2 Clean Up Temp Files on VPS DONE

Deleted 181 Python patch files from /tmp/. Confirmed clean.

1.3 Clean Up Local Temp Files

What: Delete temp .py and .ts and .md files in C:\Users\New\ that were used for patching. Files like: fix-*.py, wire-*.py, add-*.py, improve-*.py, rebalance-*.py, etc. These are build artifacts — the actual code is on the VPS. Note: This is on boss's local PC — can't be done from VPS.

PRIORITY 2: Set Up Opus Verdict Scheduled Task

What

Create a Claude Code scheduled task that runs daily at ~6:15PM ET. This IS the Opus verdict layer.

How

Go to claude.ai/code/scheduled or type /schedule in Claude Code.

Task Configuration

• Schedule: Daily at 6:15 PM ET (after analysts finish at ~6:00 PM)
• Model: Opus (for max reasoning on final verdicts)
• MCP: Enable EdgeClaw MCP server

Task Prompt (draft):

You are the Opus Verdict Layer for the EdgeClaw sports betting pipeline.

You go FRESH every session. No memory of past verdicts.

STEP 1: Check if today has analyst picks.
Use the edgeclaw MCP tool `analyst_picks` for today's date and NBA.

If no picks exist, respond "No analyst pipeline output found for today. Nothing to rule on." and stop.

STEP 2: Read all data.
- Use `matchup_cards` to see today's matchup projections
- Use `analyst_picks` to see what Sonnet, Gemini, and gpt-oss said
- Use `edges` to see the mathematical edge scanner output
- Use `freshness` to verify data is current

STEP 3: For EVERY market (spread, total, ML) on EVERY game:
Make a final verdict:
- final_line: your line
- final_probability: your probability
- final_conviction: 1-5
- position_size: FULL / THREE_QUARTER / HALF / QUARTER / MINIMUM
- convergence_class: from comparison layer
- verdict_narrative: 2-3 sentences on why

STEP 4: Store your verdicts.
Use the EdgeClaw API at https://edge-assist.duckdns.org/api/ to store results.
POST to /api/store-verdict with the verdict data.

Rules:
- No passing. Every market gets a pick.
- You see everything: analyst reports + market prices + math. The analysts were blind — you are not.
- Position size based on convergence: A=FULL, B=THREE_QUARTER, C=HALF, D=QUARTER, E=MINIMUM
- Adjust based on your own assessment of the evidence quality

API Endpoint Needed

Build POST /api/store-verdict endpoint on EdgeClaw that accepts Opus verdict data and updates analyst_picks table with verdict_line, verdict_probability, verdict_conviction, convergence_class, position_size.

File to modify: /home/ubuntu/edgeclaw/src/pipeline/edgeclaw-api.ts

app.post('/api/store-verdict', async (c) => {
  const body = await c.req.json();
  const db = getDb();
  const stmt = db.prepare(`
    UPDATE analyst_picks SET
      verdict_line = ?, verdict_probability = ?, verdict_conviction = ?,
      convergence_class = ?, position_size = ?
    WHERE date = ? AND sport = ? AND game_id = ? AND market_type = ?
  `);
  for (const v of body.verdicts) {
    stmt.run(v.final_line, v.final_probability, v.final_conviction,
      v.convergence_class, v.position_size,
      v.date, v.sport, v.game_id, v.market_type);
  }
  return c.json({ success: true, updated: body.verdicts.length });
});

PRIORITY 3: Run It Live (Test)

What

Fire the full pipeline manually for one night to verify output quality.

How

# On VPS or via MCP:
curl -X POST https://edge-assist.duckdns.org/api/run-full-pipeline/nba

# Or via MCP tool:
run_full_pipeline

Cost: ~$0.10 for all 3 analysts on a full slate.

What to Check

• Do all 3 analysts return valid JSON?
• Are probability curves monotonically decreasing?
• Do win probabilities sum to 1.0?
• Are lines reasonable (spreads -1 to -15, totals 195-240, ML 30-85%)?
• Does the pipeline page show all 6 steps populated?

PRIORITY 4: Performance Tracker Tab

What

New tab on /data-status page alongside "Desk Coverage", "Unrouted Kalshi", "All Sources".

Shows

• Running Brier scores per analyst per market type
• Conviction calibration (do HIGH picks hit more than LOW?)
• Ensemble vs individual (does the composite beat any single analyst?)
• Convergence class hit rates
• Opus adjustment value (are overrides adding value?)

When to Build

After 1-2 weeks of graded picks (need data first).

PRIORITY 5: Monthly Calibration Digest Generator

What

Script that generates personalized calibration digests for each analyst + Opus.

Details

Per the CAPS-001 calibration spec at /home/ubuntu/edgeclaw/docs/CALIBRATION-SPEC.md:

• Compute Brier scores per analyst per market type per sport
• Compute ECE (Expected Calibration Error)
• Identify biases: compression, home court, narrative, fatigue, injury
• Generate plain-text digest with 8-15 corrections
• Include what they get RIGHT (not just wrong)
• Store at /home/ubuntu/edgeclaw/data/calibration/digests/

When to Build

After 4-6 weeks of baseline data (no corrections during baseline period per spec).

Files

• New: src/pipeline/calibration-digest-generator.ts
• Reads from: analyst_picks table (settled picks with Brier scores)
• Writes to: data/calibration/digests/{analyst}-D{version}-{date}.md

PRIORITY 6: Research Intel Display on Pipeline Page

What

The research_intel table exists and the Step 2 section on the pipeline page is wired to read from it. But it currently shows "No alerts" because research hasn't run yet.

To Test

Run research manually:

curl -X POST https://edge-assist.duckdns.org/api/run-research/nba

Cost: ~$0.50 for Grok x_search queries across all games.

After running, refresh pipeline page — Step 2 should show findings per game.

PRIORITY 7: Other Sport Desks

What

Replicate the NBA pipeline for NHL, NCAAB, MLB, Soccer, MMA.

How

1. Use the panel prompt template at /home/ubuntu/edgeclaw/docs/panel-prompt-data-audit.md
2. Run the council for each sport to identify data gaps and custom metrics
3. Build sport-specific scrapers and matchup card calculators
4. Same 3 analysts (Sonnet, Gemini, gpt-oss) — locked roster applies to all sports
5. Same comparison layer and Opus verdict

Priority Order

• NHL (goalie matchups + existing data)
• NCAAB (tournament time, high volume)
• MLB (when season starts)
• Soccer (20+ leagues, complex)
• MMA (qualitative, needs most research)

KEY FILES ON VPS

KEY SPECS ON VPS (in docs/)

KEY CONFIG

CRON SCHEDULE (all ET)