Total account value · Kalshi
$2,448.20
+$3.67 (+0.15%) past 24h
Kalshi mark · as of 8/22 2:38 PM PT · refreshes each load (≤1 min)
Cash
$1,270.40
settled, ready to deploy
In positions
$1,177.80
6 open markets
Deposited
$1,675.00
lifetime, since Mar 2026
Net P&L
+$773.20
+46.2% on deposits
+159% annualized

Value over time

Total value Contributed (deposits) Deposit Sale — realized gain the gap between the two lines is trading gains
Solid line: live hourly marks (since Jul 2, 2026). Dashed: reconstructed from trade history — positions at cost, value stepping only at deposits, trades, and settlements (May 4 onward; the Mar–Apr era, under $300 total, is beyond Kalshi's export window). Reconciles with the live balance to the cent.

System Design

Canopy — an ensemble research engine that prices prediction markets: three independent AI analysts, a conditional reconciler, and a market blend, all validated in a 12-month walk-forward simulation before real dollars move. Nothing trades automatically.

DATA IN Polymarket prices · full resolution rules · order flow · settlements Live news, date-gated GDELT global news index via BigQuery — the agents run their own searches Market universe & evaluation triggers ~1,200 tradeable markets scanned cheaply every cycle — a market earns a full (expensive) evaluation only on a trigger: new listing · price drifted from our estimate · news scout detects a coverage spike · random audit (the control group) PROVEN IN A 12-MONTH WALK-FORWARD FIRST ① Research Ensemble — three model families price the market independently each analyst is blind to the market price AND to the others — diversity of model family is the ingredient that works Claude Sonnet Anthropic · frontier GPT-OSS 120B OpenAI · open-weight DeepSeek V3.1 DeepSeek · open-weight ② Aggregate — geometric mean of odds pure code, no model · the members' SPREAD is kept as a signal — disagreement flags danger ③ Reconciler — a judge that fires only when it's needed trigger: analysts >25 points apart, or the ensemble >30 points from the market price reads all three write-ups, finds the crux, runs its own targeted searches, issues its own number — never averages calm markets skip the judge ④ Blend with the market price — log-odds, 50/50 the crowd gets a vote: extreme divergence becomes a small position instead of a blind all-in ⑤ Quarter-Kelly sizing + risk rails bet a quarter of the mathematically optimal fraction · caps per market, per theme, and on market liquidity ⑥ Trading Desk — a human places every wager hold to resolution · nothing auto-trades Settlement & scoring every wager scored: closing-line value · Brier vs the market · which trigger paid for the estimate OUTCOMES RECALIBRATE RUNS ON Google Cloud Rundashboard + scheduled jobs Cloud SQLalerts · predictions · trades BigQueryGDELT news · SQL date firewall Secret Mgr + Storagekeys · market archive · images
Why an ensemble: re-asking one model gives near-identical answers, but different model families disagree in genuinely useful ways — cross-model agreement carries measurable predictive value, and their spread doubles as a danger gauge. Why price-blind: an analyst who can see the market price anchors to it; the price enters only later, in deterministic code — as the Reconciler's alarm and the 50/50 blend that turns extreme disagreement with the crowd into a small position rather than a bet-the-house moment. The Trading Desk is a human — nothing trades automatically. The proving ground: before any configuration touches real dollars it replays 12 months of history in a walk-forward simulation with an honest bankroll — finite cash, positions lock until resolution, and every model re-evaluation must be justified by a trigger and paid for. Python + Flask, containers on GCP.
Previous design — single-principal era (kept for comparison)
DATA IN Kalshi REST API prices · full resolution rules · candlesticks · positions · settlements Web Search current facts — defeats the model's training cutoff Orchestrator — runs the sweep end to end sequences every stage · fans out the agent tiers · retries failures · persists verdicts and funnel stats CONTROLS EVERY STAGE Discover the universe enumerate ~3,800 series per-category → filter price / time-to-close → the tradeable universe ① Deterministic Intake — filter & rank plain code, no model · correlation-aware sibling caps · category round-robin → ranked shortlist ② Managing Director — dispatch prioritizes the shortlist and fans out the research tiers below UNDER CONSTRUCTION ③ Associate Research Agents — cheap triage tier a fast, low-cost LLM pass over a much LARGER candidate set — recent-news aware scores each opportunity so only the strongest earn an expensive Principal agent ④ Principal Research Agents — one LLM agent per market reads the FULL resolution rules (the auto-title is untrusted) and states what resolves YES BEFORE pricing · web-searches current facts · blind to the market price returns a structured verdict: probability, trigger type, nuance, confidence MCP tool layer get_resolution_rules get_market · orderbook search · context read-only Kalshi data; the agent chooses which tools to call Score the edge our probability vs. the live price → conviction from projected ANNUALIZED return · near-certain-trigger guard Alert Queue Cloud SQL · ranked by conviction · live-repriced with drift flags · sectioned by sweep date ⑤ Trading Desk — human triage Take · Watch · Dismiss, then the wager is placed manually on Kalshi — nothing auto-trades Track record positions auto-linked to the prediction that drove them → CLV + calibration (Brier vs. the market) OUTCOMES RECALIBRATE RUNS ON Google Cloud Rundashboard service + scheduled jobs Cloud SQL (Postgres)alerts · predictions · trades · history Secret ManagerAPI keys · RSA-PSS signing Cloud Storage + Registrymarket archive · container images
What's actually an agent: the Principal Research Agents are the real LLM agents today — one per shortlisted market, each web-searching independently and returning a schema-validated verdict. The Associate Research Agents are the cheap triage tier that will let us screen far more markets than 24; that stage is under construction and nothing runs through it yet. Orchestrator, Deterministic Intake and Managing Director are deterministic code, not autonomous agents — a workflow runner, a filter/rank pass, and the fan-out. The Trading Desk is you. Both agent tiers reach market data through a read-only MCP tool layer (Model Context Protocol) — the agent chooses which tools to call, which is what makes the research step agentic rather than fed a fixed context blob. Python + Flask throughout, deployed as containers on GCP.

Active Positions Summarized

Over the past month the desk booked its first realized wins fading low-probability events. The standout was a NO position against Samuel Alito retiring from the Supreme Court before September — entered near 79¢ and exited at 96¢ as the market repriced toward our thesis, for a realized +$139.52 (+18.6%). Smaller settled fades (Jerome Powell’s governor seat, plus World Cup and PGA markets) added incremental gains. The book is currently positioned in the same spirit — a set of NO fades against unlikely outcomes: that Warner Bros. avoids its expected takeover, that senior officials (Hegseth, Kash Patel) exit their posts before 2027, and that the President is impeached before 2028. Each is a high-probability wager that a ‘crazy thing’ won’t happen, sized small; by design the portfolio grinds slowly upward, with the risk concentrated in occasional single-position drawdowns on the rare occasion we are wrong.
MarketSide ContractsCost ValueP&L
Will Miguel Díaz-Canel leave First Secretary of the Communist Party of Cuba before Jan 1, 2027? NO 421 $345.63 $357.85 +$12.22
Will the President be impeached before Mar 1, 2027? NO 300 $228.00 $237.00 +$9.00
Will Jerome Powell leave Member of the Board of Governors of the Federal Reserve System before Jan 1, 2027? NO 460 $193.82 $363.40 +$169.58
Will None's takeover of Warner Brothers succeed Before July 2027? NO 189 $171.89 $154.98 $-16.91
Will Kash Patel leaves Director of the FBI in before 2027? NO 50 $24.50 $37.50 +$13.00
Will the President be impeached before Jan 1, 2028? NO 50 $20.00 $21.50 +$1.50
Total — $983.84 $1,172.23 +$188.39
MATES · Multi-Agent Trading Execution System
Illustrations: J.J. Audubon (1837) · M.V. Walcott, North American Wild Flowers (1925), public domain · data refreshed 8/22 2:38 PM PT