Politics & Geopolitics Desk — Final Ruling

Panel: Opus, Sonnet, Grok 4.2 Reasoning, Gemini Pro 3.1 Judge: Opus (this document) Date: 2026-03-26

Overall Grade: A

Strong panel. All four delivered specific, implementable ideas with URLs, schemas, and formulas. No hand-waving. Minor deductions: Grok's cron schedule was too sparse (daily-only for things that need hourly), Gemini's Flashpoint Index is creative but the NOTAM/MARAD/ADS-B sources may be brittle in practice. Otherwise excellent.

Consensus Points (all 4 agree — CONFIRMED)

1. No single "Pinnacle" exists for politics

Every panelist independently concluded this. Sports has Pinnacle. Crypto has Deribit. Politics has nothing equivalent. We must build a synthetic sharp line from multiple sources.

2. Congress.gov API is the anchor for bill/legislation contracts

Free API, 1000 req/hr with free key from api.data.gov. Track bill stages (introduced → committee → floor → passed → enrolled → signed). All 4 proposed nearly identical bill pipeline trackers.

• URL: https://api.congress.gov/v3/bill?api_key=KEY&limit=250&sort=updateDate+desc
• Sign up: https://api.congress.gov/sign-up/ (free, instant)

3. Federal Register API is the anchor for EO/executive action contracts

Free, no auth needed. Tracks EOs, proclamations, memoranda.

• URL: https://www.federalregister.gov/api/v1/documents?conditions[presidential_document_type]=executive_order&conditions[publication_date][gte]=YYYY-MM-DD

4. GDELT is the primary geopolitical signal source

Free 15-minute feed. CAMEO-coded events with Goldstein scale scores. Filter QuadClass 3-4 (material conflict) for target countries.

• URL: http://data.gdeltproject.org/gdeltv2/lastupdate.txt

5. RCP + 538 approval scrapers are essential

These are the direct data anchors for the highest-volume recurring political contracts (KXAPRPOTUS, KX538APPROVE).

• RCP: Scrape polling table at realclearpolling.com
• 538: CSV at projects.fivethirtyeight.com/polls/data/approval-averages.csv (if still available) or scrape

6. CBP border data for encounter contracts

Monthly CSV from CBP. Predictable release schedule (mid-month for prior month).

• URL: https://www.cbp.gov/newsroom/stats/southwest-land-border-encounters

7. Don't build NLP sentiment, ML forecasting models, or real-time streaming

All 4 said this independently. The prediction market consensus already incorporates sentiment. ML models need GPU farms we don't have. Cron is sufficient — political data doesn't move at microsecond speed.

8. Focus on the 65 recurring series, largely skip the 3,158 one-offs

Recurring series have patterns, are backtestable, and compound over time. One-offs require manual context and don't repeat.

Dissent Points + Rulings

Dissent 1: Polymarket as primary sharp anchor vs. ensemble-only

Sonnet & Opus: Polymarket IS the sharp book. It has higher volume, more sophisticated (international) traders, and covers most Kalshi political markets. The Polymarket-Kalshi delta is the primary edge signal.

Grok & Gemini: No single sharp book — use a Bayesian/weighted ensemble of Metaculus + Manifold + historical base rates + proprietary metrics.

RULING: Sonnet/Opus win. Polymarket is the political Pinnacle.

Polymarket is real-money with deep liquidity, especially on political contracts. It's the closest thing politics has to a sharp book. The ensemble approach is also correct — but Polymarket should be the heaviest-weighted component, not one-of-many. We already scrape Manifold and Metaculus; adding Polymarket as the primary anchor with the others as cross-validation is the right architecture.

Implementation:

• Polymarket CLOB API: https://clob.polymarket.com/markets and https://gamma-api.polymarket.com/markets?closed=false&limit=500
• Free, no auth required
• Scrape every 30 min (same cadence as our sports edge scanner)
• Weight: Polymarket 0.45, Metaculus 0.25, Manifold 0.20, our model 0.10

Dissent 2: Approval momentum model approach

Opus: Velocity + projected Friday endpoint. Front-run pollster house effects (know which pollsters are about to publish and estimate their impact on the average).

Sonnet: Simpler — 7-day velocity, threshold crossing detection.

Grok: Second derivative (acceleration) + cross-market beta against PRESINDEXD.

Gemini: Time-series forecast with cosponsors and bipartisanship (wrong section — this was for bills, not approval).

RULING: Opus's approach is the most novel. Build velocity first (Sonnet), add house effects later (Opus).

The poll release timing exploitation (front-running which pollsters enter the RCP average using house effect tables) is genuinely alpha. But it's Phase 2 — Sonnet's simpler velocity model works immediately and can be enhanced. Grok's acceleration idea is interesting but harder to calibrate with limited data.

Implementation:

• Phase 1: approval_velocity = (current - 3d_ago) / 3, project to Friday
• Phase 2: Build pollster house effect table, predict upcoming poll releases
• Compare projected endpoint to Kalshi bucket boundaries using Normal CDF

Dissent 3: Geopolitical signal sources

Opus, Sonnet, Grok: GDELT events (CAMEO-coded, Goldstein scale)

Gemini: GDELT PLUS FAA NOTAMs, MARAD maritime alerts, and ADS-B/AIS military flight/ship tracking

RULING: Gemini's Flashpoint Index is the most creative idea in the entire panel. APPROVED for Phase 3.

NOTAMs, MARAD alerts, and military flight tracking are genuine alpha that virtually no prediction market trader uses. A new MARAD alert for the Persian Gulf combined with increased US tanker-aircraft activity IS a concrete signal for UNSC veto / conflict contracts. However:

• ADS-B Exchange API may require paid access for military tracks
• NOTAMs are hard to parse (specialized format)
• This is Phase 3 material — build the basics first

Implementation (Phase 3):

• MARAD: https://www.maritime.dot.gov/msci/msci-alerts — scrape alert list
• NOTAMs: https://notams.aim.faa.gov/ — scrape by region
• GDELT as primary (Phase 1), Flashpoint Index as enhancement (Phase 3)

Dissent 4: Gemini's "Say-Do Gap" (FEC filings vs Truth Social)

Gemini only: Track where campaign money is actually being spent (FEC filings) vs what Trump posts about. The money is the sharp signal.

RULING: Interesting but LOW PRIORITY. FEC data is quarterly and heavily lagged.

FEC filings are valuable for long-term election markets but update too slowly for our weekly/monthly recurring contracts. The idea is sound in principle — money reveals intent better than tweets — but the data cadence doesn't match our contract cadence. Defer to Phase 4 if ever.

Dissent 5: Truth Social scraping

Opus: Mastodon-compatible API or third-party trackers. Poisson model for post velocity.

Grok: LLM-classified post types (rage/policy/personal) for entropy metric.

Sonnet/Gemini: Acknowledge it's hard. Gemini suggests logged-in scraping.

RULING: Use third-party trackers first. Don't scrape Truth Social directly.

Truth Social's API is hostile and direct scraping risks bans. truthsocialarchive.com or similar third-party trackers that already aggregate post counts are lower-risk. The Poisson extrapolation model (Opus) is the right math for the KXTRUTHSOCIAL contract. Grok's LLM entropy classifier is over-engineered for a post-counting contract.

Dissent 6: PredictIt inclusion

Sonnet: Include PredictIt as a source. API at predictit.org/api/marketdata/all/

Others: Don't mention PredictIt.

RULING: Include PredictIt but low weight.

PredictIt is less efficient than Polymarket but still free data. Include it as a cross-validation source at low weight (0.05-0.10). The API is simple and stable.

Dissent 7: Separate database vs existing research-pipeline.db

Grok: Separate politics.db

Others: Implied use of existing db

RULING: Use existing research-pipeline.db. No new database.

All our other desks use the same database. Splitting creates maintenance burden and breaks cross-desk queries. Add new tables to research-pipeline.db via getPipelineDb().

Final Build Plan (Binding)

Phase 1: Days 1-5 — Polymarket + Approval Anchors

1. Polymarket politics scraper — CLOB API, every 30 min, store in external_market_prices
2. PredictIt scraper — simple REST API, every 30 min
3. RCP approval scraper — daily at 6AM ET
4. 538 approval scraper — daily at 6AM ET
5. Kalshi politics series collector — collect prices for the 65 recurring series
6. Manual mapping table — map top 20 Kalshi series to Polymarket/Manifold/Metaculus/PredictIt equivalents
7. Consensus computation — weighted average, output edge signals
8. Telegram alert — when abs(edge) > 0.08 on any recurring series

Tradeable output by Day 5.

Phase 2: Days 6-14 — Government Data Anchors

9. Congress.gov API scraper — bill pipeline with readiness scoring
10. Federal Register API scraper — EOs, proclamations, memoranda
11. WH schedule scraper — for lid prediction
12. CBP encounter scraper — monthly data
13. Approval momentum model — velocity + projected Friday endpoint
14. EO frequency model — Poisson with contextual multipliers

Phase 3: Days 15-30 — Geopolitical + Advanced

15. GDELT event ingestor — filter CAMEO 14-20, store by country
16. Geopolitical conflict velocity metric — 72h count vs 90d baseline
17. MARAD alert scraper (Gemini's Flashpoint Index)
18. Pollster house effect table (Opus's poll timing exploitation)
19. Truth Social post count via third-party tracker

Phase 4: Ongoing — Refinement

20. Backtesting against historical Kalshi resolutions
21. Weight optimization for consensus ensemble
22. NOTAM parser (if MARAD proves valuable)
23. FEC integration (if relevant for longer-dated contracts)

Table Schema (Binding — add to research-pipeline.db)

-- Approval ratings (RCP, 538)
CREATE TABLE IF NOT EXISTS approval_ratings (
  id INTEGER PRIMARY KEY AUTOINCREMENT,
  source TEXT NOT NULL,          -- 'rcp', '538', 'silver_bulletin'
  metric TEXT NOT NULL,          -- 'approve', 'disapprove', 'net', 'favorable'
  value REAL NOT NULL,
  captured_at TEXT NOT NULL
);

-- Individual polls (for house effect calculation)
CREATE TABLE IF NOT EXISTS polls_individual (
  poll_id TEXT PRIMARY KEY,
  pollster TEXT NOT NULL,
  field_date_end TEXT,
  sample_size INTEGER,
  population TEXT,               -- 'RV', 'LV', 'A'
  question_type TEXT NOT NULL,
  result_positive REAL,
  result_negative REAL,
  captured_at TEXT NOT NULL
);

-- Pollster house effects (computed weekly)
CREATE TABLE IF NOT EXISTS pollster_house_effects (
  pollster TEXT NOT NULL,
  question_type TEXT NOT NULL,
  house_effect REAL NOT NULL,
  sample_count INTEGER NOT NULL,
  updated_at TEXT NOT NULL,
  PRIMARY KEY(pollster, question_type)
);

-- Congressional bill pipeline
CREATE TABLE IF NOT EXISTS congress_bills (
  bill_id TEXT PRIMARY KEY,
  title TEXT,
  bill_type TEXT,
  congress INTEGER,
  introduced_date TEXT,
  last_action_date TEXT,
  last_action TEXT,
  status TEXT NOT NULL,          -- 'introduced','committee','floor','passed_one','passed_both','enrolled','signed'
  readiness_score INTEGER DEFAULT 0,
  captured_at TEXT NOT NULL
);

-- Executive actions (EOs, proclamations, memoranda)
CREATE TABLE IF NOT EXISTS executive_actions (
  document_number TEXT PRIMARY KEY,
  action_type TEXT NOT NULL,
  title TEXT NOT NULL,
  signing_date TEXT,
  publication_date TEXT,
  eo_number INTEGER,
  captured_at TEXT NOT NULL
);

-- WH schedule + lid calls
CREATE TABLE IF NOT EXISTS wh_schedule (
  event_date TEXT NOT NULL,
  lid_called_at TEXT,
  lid_type TEXT,
  event_count INTEGER,
  is_travel INTEGER DEFAULT 0,
  captured_at TEXT NOT NULL,
  PRIMARY KEY(event_date, captured_at)
);

-- CBP border encounters (monthly)
CREATE TABLE IF NOT EXISTS cbp_encounters (
  month TEXT PRIMARY KEY,
  sw_encounters INTEGER NOT NULL,
  source_url TEXT,
  captured_at TEXT NOT NULL
);

-- External prediction market prices (Polymarket, PredictIt)
CREATE TABLE IF NOT EXISTS external_market_prices (
  id INTEGER PRIMARY KEY AUTOINCREMENT,
  source TEXT NOT NULL,           -- 'polymarket', 'predictit'
  market_id TEXT NOT NULL,
  question TEXT,
  yes_price REAL,
  no_price REAL,
  volume REAL,
  matched_kalshi_series TEXT,
  captured_at TEXT NOT NULL
);

-- Kalshi-to-external market mapping (manually curated)
CREATE TABLE IF NOT EXISTS prediction_kalshi_map (
  kalshi_series TEXT NOT NULL,
  external_source TEXT NOT NULL,
  external_id TEXT NOT NULL,
  match_quality TEXT NOT NULL,   -- 'exact', 'close', 'related'
  last_verified TEXT NOT NULL,
  PRIMARY KEY(kalshi_series, external_source, external_id)
);

-- GDELT conflict events
CREATE TABLE IF NOT EXISTS gdelt_events (
  id INTEGER PRIMARY KEY AUTOINCREMENT,
  event_date TEXT NOT NULL,
  country_code TEXT NOT NULL,
  cameo_code TEXT NOT NULL,
  goldstein_scale REAL,
  num_mentions INTEGER,
  quad_class INTEGER,
  captured_at TEXT NOT NULL
);

-- Computed politics consensus + edge
CREATE TABLE IF NOT EXISTS politics_consensus (
  kalshi_series TEXT NOT NULL,
  kalshi_price REAL,
  consensus_prob REAL NOT NULL,
  consensus_sources INTEGER,
  edge REAL,
  polymarket_prob REAL,
  predictit_prob REAL,
  manifold_prob REAL,
  metaculus_prob REAL,
  model_prob REAL,
  computed_at TEXT NOT NULL,
  PRIMARY KEY(kalshi_series, computed_at)
);

-- MARAD maritime alerts (Phase 3)
CREATE TABLE IF NOT EXISTS marad_alerts (
  alert_id TEXT PRIMARY KEY,
  region TEXT,
  alert_date TEXT,
  advisory_text TEXT,
  severity INTEGER,
  captured_at TEXT NOT NULL
);

-- Indexes
CREATE INDEX IF NOT EXISTS idx_approval_source_time ON approval_ratings(source, captured_at DESC);
CREATE INDEX IF NOT EXISTS idx_gdelt_country_date ON gdelt_events(country_code, event_date DESC);
CREATE INDEX IF NOT EXISTS idx_external_market_source ON external_market_prices(source, matched_kalshi_series, captured_at DESC);
CREATE INDEX IF NOT EXISTS idx_congress_status ON congress_bills(status, last_action_date DESC);
CREATE INDEX IF NOT EXISTS idx_politics_consensus_series ON politics_consensus(kalshi_series, computed_at DESC);

Cron Schedule (Binding — all ET)

Edge Detection Formula (Binding)

consensus_prob = (0.45 * polymarket) + (0.25 * metaculus) + (0.20 * manifold) + (0.10 * model)

// For data-anchored contracts (bills, border, EOs):
// model weight increases to 0.40, others decrease proportionally

edge = consensus_prob - kalshi_implied_prob
trade_threshold = 0.08  (8 cents)
min_sources = 2         (at least 2 independent sources must agree)

Key Data Source URLs (Reference)

Individual Panelist Grades