Live · mission control
Watch the platform run.
A read-only view of the system’s own health: container fleet, data spine, validation, self-healing, machine regime, honestly measured performance, and the live ML research matrix (walk-forward AUC across models and targets). No positions or account data are exposed.
loading snapshot…
11 sleeves, 2 live · 6 in shadow · deliberately different animals
Each sleeve is named for the Greek letter of the risk it monetizes. Alpha · beta-neutral statistical edge in event contracts: small, repeatable, independent of market direction. Beta · directional exposure to the underlying, with daily size set by a statistical regime classifier built on published regime-detection academics. Charm · Δ-decay (dΔ/dTime): deltas drift as expiry approaches; 0DTE pin dynamics. Delta · dPrice/dSpot: direction; index-futures trend exposure from range-break models. Gamma · dΔ/dSpot: convexity; long premium when realized vol is forecast above implied. Rho · dPrice/dRate: rates; event contracts on rate decisions and macro prints. Theta · dPrice/dTime: decay; defined-risk short premium in forecast calm. Vega · dPrice/dVol: volatility level; term-structure calendars. Vanna · dΔ/dVol (equivalently dVega/dSpot): deltas shift as volatility jumps; skew structures. Veta · dVega/dTime (vega decay): vol exposure bleeds as expiry approaches. Vomma · dVega/dVol (volga): vega convexity; exposure that grows as vol itself jumps. All are 100% data-driven; every strategy earns capital through purged walk-forward validation, deflated Sharpe ratios, and null-control hypothesis tests, and is demoted the day the data says stop. They share no signal, no sizing, no failure mode.
| Sleeve | Strategy | Venue | Stage |
|---|---|---|---|
| alpha | Cross-venue lead-lag xven_leadlag_shadow · live/shadow 2421 closed · +$202.43 · backtest N/A · blocked: dsr, lcb, beats_1N | Kalshi | shadow |
| alpha | Hourly strike hourly_strike_paper · live/shadow 6791 closed · −$3.79 · backtest N/A · blocked: stage0, dsr, walk_forward, lcb, beats_1N | Kalshi | shadow |
| alpha | Range binary (breach vs stay 1σ) range_binary · live/shadow 676 closed · +$439.19 · backtest N/A · blocked: dsr, walk_forward, lcb | Kalshi / opts | shadow |
| beta | Winners Book (stacked) __winners_book__ · live/shadow 5 strategies · +$10,076.96 · backtest N/A · blocked: aggregate of member strategies | multi-venue | shadow |
| charm | 0DTE defined-risk condor volexp_condor_0dte · live/shadow 34 closed · +$986.84 · backtest N/A | Tastytrade | canary ($1) |
| delta | 15m direction (ML) uf_direction_shadow · live/shadow 322 closed · +$2.03 · backtest N/A · blocked: stage0, dsr, walk_forward, lcb, beats_1N | Kalshi | shadow |
| gamma | Long premium / tail convexity heatmap_long_vol · live/shadow 2 closed · +$2,025.50 · backtest −$149.17 (long strangle upper bound (mid), DoltHub) · blocked: stage0, dsr, walk_forward, lcb, beats_1N | Tastytrade | shadow |
| rho | Rates & macro event contracts no strategy has cleared into this sleeve yet · live/shadow N/A · backtest N/A | Kalshi | no candidate |
| theta | Short premium (naked strangle) heatmap_short_premium · live/shadow 79 closed · +$8,729.00 · backtest −$150.01 (exp/trade crossed, DoltHub 2020-26) | Tastytrade | canary ($1) |
| vega | Vol term-structure calendars no strategy has cleared into this sleeve yet · live/shadow N/A · backtest N/A | Tastytrade | no candidate |
| vanna | Skew & vol-acceleration structures no strategy has cleared into this sleeve yet · live/shadow N/A · backtest N/A | Tastytrade | no candidate |
| veta | Vega-decay structures no strategy has cleared into this sleeve yet · live/shadow N/A · backtest N/A | Tastytrade | no candidate |
| vomma | Vol-of-vol convexity no strategy has cleared into this sleeve yet · live/shadow N/A · backtest N/A | Tastytrade | no candidate |
What each sleeve trades · instrument, thesis, and the models that feed it
alpha: Beta-neutral statistical edge in event contracts. Small, repeatable structures whose payoff does not depend on market direction; the flagship is cross-venue lead-lag on Kalshi, with maker and mean-reversion variants admitted only where net-of-fee expectancy survives real-fill accounting.
beta: The Winners Book: a stacked portfolio of validated directional strategies whose daily gross exposure is set by the statistical regime classifier. Each constituent earned its place through purged walk-forward validation and is demoted the day the data says stop.
charm: 0DTE pin and delta-decay structures. Trades the pull of price toward high-open-interest strikes into expiration, gated by the fleet’s inside-move and gamma-state models.
delta: Directional index futures on CME micros. Range-break and trend continuation; the ML daily lane gates entries, and the sleeve makes no edge claim until that gate proves itself in shadow.
gamma: Vol-expansion long premium. Buys straddles and strangles when the vol-spike and IV-expansion models say realized vol will outrun implied. Long premium only pays when the forecast is right, so every structure earns its way through shadow first.
rho: Rates and macro event contracts on Kalshi. A venue and a hypothesis; candidates are being built.
theta: Short premium. Sells rich implied volatility, gated on a vol-spike forecast rather than on high IVR alone. The proven track is a NAKED strangle book (undefined risk); defined-risk expressions of the same signal are still accruing evidence and are not yet proven.
vega: IV term-structure calendars. Trades the shape of the volatility curve across expiries, driven by the fleet’s IV-native targets (crush, expansion, premium survival).
vanna: Skew and vol-acceleration structures. Vanna-flow tags from the dealer-exposure engine time the entries; candidates in validation.
veta: Vega-decay structures. Studies how vega decays across the term structure; features live in the Greeks brain, candidates being spec’d for walk-forward.
vomma: Vol-of-vol convexity. Trades the price of vol-of-vol when it sits rich or cheap versus realized; tags context while candidates validate.
No per-day numbers here by design: they change nightly, the theses do not. Current evidence lives in the validation record.
Promotion is per-strategy through gate → shadow → canary → live, and demotion lands the day re-validation fails. No P&L decomposition, positions, or account data is shown.
model zoo
Every family, what it is good at
Forecast fleet · 12 architectures, one honesty protocol
Gradient boosting (XGB · LGBM · CatBoost): tabular workhorses; CatBoost’s best cells include SOL vol-spike 0.807 and IV crush 0.888. The IV-native crowns themselves belong elsewhere (IV crush 0.907 TabPFN, IV expansion 0.860 Mamba, premium survival 0.801 XGB).
GRU: the lightest neural net; owns SOL vol-spike (0.828) and leads the 0DTE inside-move and spread-safe tables; widest 0DTE coverage (69% of cells).
Mamba (state-space): owns the fleet’s best IV-expansion cell (0.860); runs crypto 5m and the daily universe.
JEPA (pretrained encoder): owns crypto crash on all three assets and BTC vol-spike (0.805 walk-forward AUC).
Challengers (TFT-lite · N-BEATS · TiDE · PatchTST): TFT owns ETH vol-spike (0.801) and the best inside-move cell (0.911); N-BEATS holds the best non-proxy cell in the fleet (HYG vol-spike 0.906).
TabPFN (tabular foundation model, no per-task training): the fleet's centre of gravity. The family (2.5 + 3.0) holds 804 of 2,717 outright championships (29.6%, 2.0× the next family) and stands on the podium in a majority of all rows (53.7%); v2.5 alone still leads every individual architecture (607 wins, 47.8% podium). v3 covers 80% of the board with 197 championships of its own; the v2-vs-v3 head-to-head stands 1,513 to 644, so the upgrade verdict waits for full coverage. Family bests include ^NDX vol-spike 0.899 and the fleet’s best IV-crush cell (0.907).
GNN (cross-sectional graph): owns BTC big-move (0.693); its 0DTE variant honestly failed net-of-cost and that negative is published.
Production stack
XGBoost meta-controller: turns specialist signals and 322 unified features into the approve and size decision.
Regime classifier: CRASH_WATCH > SPIKE_WATCH > RICH_SELL > CALM_THIN; sets gross exposure for every sleeve.
Regime tags (candidate, advisory): second-order Greek axes layered over the four master regimes: dealer-gamma flip (pinning vs trending tape), vanna flow (vol-direction hedging pressure), charm window (decay-hedging drift into the close), vol-of-vol state (convexity rich vs cheap). Direction is real in research; net economics are unproven at real costs (one of four axes cleared its standalone bar at thin coverage, the combined program did not), so they tag context only and gate nothing until they pass the same walk-forward bar as everything else.
Vol forecaster: CatBoost forward realized vol, OOS R2 0.574 at 7 days, rank-IC 0.80; drives strike selection.
0DTE target models: inside_move 0.911, spread_safe 0.884, gamma_state 0.986 (flagged proxy, disclosed).
ZoneX exit networks: 10 per-volatility-zone exit nets, 47 walk-forward-validated exit slots in shadow.
Conformal gate: per-regime abstention; the machine is allowed to say no.
Options and risk engine
Pricing lab: BSM, Bachelier, binomial and Monte Carlo pricers cross-check each other; inter-model spread is the model-risk gauge.
Greeks brain: net-book delta, gamma, vega, theta plus second order (vomma, vanna, charm, veta) with a neutrality optimizer.
LPPL bubble model: kept as regime context; its crash predictor honestly failed the quality gate (AUC 0.46 to 0.51) and is not traded.
Governance and memory
CaT constitutional pipeline: a strategist proposes, a judge enforces the constitution and risk limits before any order.
TUMIX deliberation: deterministic multi-agent rounds with consensus early stopping.
ArcMemo: concept and episode memory; lessons survive restarts.
Every number above comes from the same purged walk-forward protocol as the matrix; published failures count as results too.