White Paper · Internal · 6 July 2026

CHIMERA: An Autonomous, Self-Auditing Trading Research Platform

Multi-venue crypto derivatives platform: architecture, evidence record, and honest economics

STATUS: forward-validation phaseOPERATION: 24/7 autonomous, self-healingHOST: dual-GPU workstation (98 GB + 32 GB VRAM)

Chimera is a four-generation build: a microsecond-latency multi-regime ML trader (Kraken futures), a GPU genetic strategy factory (JAX Forge), a config-driven slot system on Coinbase perpetuals (V13), and now an edge factory: a fully automated research pipeline that discovers, stress-tests, and promotes trading strategies under institutional-grade statistical validation. Every strategy claim is proven against real execution costs, out-of-sample, before capital touches it. Today the platform runs ~117 containers unattended, holds a 3.5 TB proprietary research corpus, and operates a fully pre-registered forward experiment book with a fixed promotion ladder from research to live deployment.

Build lineage

GenerationCore ideaWhat it contributed
v7–v8 · multi-clock ML traderRegime-routed ensemble (CNN-LSTM, 10-zone ZoneX exit nets with per-zone attention, LPPL, VPIN, 12-agent Hive consensus) driving μs-grade exit loops on Kraken futuresThe hot-loop architecture (GPU exit <100 μs, HFT exit <50 μs, entry 100–250 ms), the Redis pre-compute pattern, capital guardian / DEFCON risk stack, counterfactual & episodic-RL learning layer
Forge / JAX Forge · strategy factoryGPU genetic algorithm evolving per-volatility-zone parameters (10 zones × 3 assets), DSR/WFE fitness, continuous 4-hour reoptimization daemonMassively parallel GPU-vectorized backtesting and sweep infrastructure, thousands of strategy evaluations per minute
V13 · slot systemAsset × vol-zone × direction slots on Coinbase CFM perps, numba backtester, auto-optimizer, shadow runnerProduction venue integration, twin paper-provers, exchange reconciliation, and a live execution-fill corpus used to calibrate every simulation
Edge factory · currentFalsification pipeline: calibrated costs, pre-registered forward book, promotion ladderThe scientific operating discipline everything now runs under (§3)
117Docker services, self-healing
3.1 TBTimescaleDB research corpus
447 GBParquet feature lake
4venues instrumented
~6,600Python modules
2GPUs (98 GB + 32 GB)

Venues: Coinbase CFM perpetuals, Kraken, Bitnomial, Kalshi perpetuals. All figures measured 2026-07-06.

§1System map

Five layers, strictly ordered: data flows down the stack, and no strategy reaches Layer 4 without surviving Layer 3. Layer 5 watches everything, including itself.

L1Data spine
20+ collectors: Coinbase L1/L2 depth, Kraken, Binance spot+derivatives (WS-persisted), Kalshi perp order books, funding, open interest, liquidations, ETH/SOL on-chain, news firehose (Tiingo/RSS/Reddit), sentiment. Stores: TimescaleDB hypertables, unified 1-second features 731 GB, raw trades 197 GB, order-book deltas 133+57 GB, 1-second silver layer (funding/OI/on-chain/liquidations/sentiment, ~30 GB each) · Redis hot state · daily parquet lake.
L2Feature & ML plane
GPU feature computer (LPPL, regime, vol zones, VPIN, Hive consensus → Redis, <100 ms budget) · vector assembler (1-second fast features, TTL-guarded) · JEPA ensemble, honest walk-forward AUC: volatility 0.82, big-move 0.73 · ZoneX: 10 per-volatility-zone exit networks with learned attention, ~600 μs inference · XGBoost sleeves · regime classifier · leakage-guard and feature-health watchdogs. Every model is continuously benchmarked against honest walk-forward baselines; only components that prove predictive power stay in the decision path.
L3Research factory
(the core)
Numba backtester with calibrated real costs (per-zone stop-fill slippage from 237 exchange-confirmed fills) · GPU sweep infrastructure inherited from Forge/JAX-Forge (80k-cell searches in minutes) · scientific gate: purged+embargoed walk-forward, deflated Sharpe (multiple-testing), three null controls, both-halves, stress perturbation, feature-basis robustness ("B-gate") · forward book: 21-config factory battle, 5 regime-engine A/B variants, Kalshi twin runners, all with verdict rules frozen before first data.
L4Execution
Multi-clock hot path on host networking: GPU exit <100 μs · HFT exit <50 μs · entry clock 100–250 ms, hot loops read Redis only, all compute is async pre-compute · V13 shadow runner (live+paper, venue mode hot-loaded from Redis) · execution engine with maker/taker routing, order splitting, retry & fill reconciliation · venue clients: Coinbase CFM, Kraken futures (multi-collateral), Bitnomial, Kalshi · Kalshi paper runner ($500 book, position/day-loss caps). Risk stack: capital guardian + DEFCON stress gates + drawdown throttle (equity-scaled leverage). Currently paper-only on every venue. Learning loop: counterfactual tracker (11 exit horizons), episodic-RL memory (MEMRL), exit-outcome learner feeding strategy weights.
L5Ops & self-healing
Infra-heal (auto-restarts critical containers, ~38 s recovery) · 25/25 heartbeats green · table-freshness & parity watchdogs · Kalshi fee-promo tripwire · nightly ops loop · a live operator monitor with true-forward vs backtest-seed columns per experiment.

Resilience proven in production: hard host reboots and infrastructure faults are absorbed automatically, and each incident class is converted into a permanent automated guard (TTL-guarded caches, disk monitors, self-re-seeding buffers, ~38 s container recovery).


§2Validation engine & certified results

Execution truth at institutional rigor. Every simulation is calibrated to measured reality: stop-fill slippage modeled per volatility zone from exchange-confirmed fills, per-contract fees from the platform's own live fill corpus, spread crossing priced from its own recorded order books. A proprietary screening taxonomy of a dozen simulation failure modes, fill modeling, selection bias, look-ahead leakage, survivorship, adverse selection, is applied to every result before it is trusted. This is the discipline that separates research-grade numbers from marketing numbers, built in as automation rather than policy.

Exhaustive certification of the search space. Every candidate strategy faces the full scientific gate: purged and embargoed walk-forward, deflated Sharpe against the complete search multiplicity, three independent null controls, split-sample stability, and cost-stress perturbation. The platform has screened over 88,000 strategy configurations across four venues under this gate, a complete, reproducible map of where edge exists and where it does not, so capital concentrates only on validated ground. Strategies that cleared the gate:

CandidateEvidenceStatus
Regime engine, long majors only in detected uptrend (ADX/DI, daily)Sharpe 1.28→1.80, maxDD −89%→−55%, stable OOS, 12 yr, 37 pairsForward A/B, 5 variants, ~30-day gate
Episodic funding harvest, spike-gated basis trade13.5% of 11-mo observations >15% annualized, clusteredActive research
Kalshi multi-slot runner, cheap-fee venue expressionValidated mechanism (slow reversion, momentum exits)Forward paper, risk-capped

The flagship structural discovery: crypto downtrends are detectable in time to act, regime detection fires within 0–7 days against 6–10-day downtrend persistence, so the engine sidesteps the bulk of every major drawdown while staying long the trend. That asymmetry is what converts buy-and-hold's −89% worst-case into −55% at higher Sharpe, validated across 37 pairs and 12 years.

§3Method: the scientific operating discipline

§4Economics & outlook

The portfolio is engineered around two complementary return profiles. The scalable core (regime engine): 20–60%/yr expectation at roughly half of buy-and-hold's drawdown, capacity-unconstrained on major pairs, with vol-targeted variants available when drawdown headroom is spent deliberately. Capacity-capped opportunistic edges (cheap-fee venue windows, episodic funding harvest): bounded dollars per day on minimal collateral, harvested while their conditions hold. The architecture runs both simultaneously, the core carries the account; the opportunistic sleeve compounds on top.

What makes it distinctive. A falsification-grade validation pipeline of a rigor usually found only inside institutional quant desks; a proprietary aligned data moat, multi-venue perpetuals tape, order-book depth, and funding at 1-second grain, 3.5 TB and growing, that cannot be purchased off the shelf; and infrastructure that operates, monitors, and repairs itself 24/7 without human intervention.

Road ahead. The forward book matures its lead candidates through pre-registered windows toward capped live deployment; the promotion ladder then scales capital in measured steps as live performance confirms paper. Planned hardening: multi-host redundancy and continued expansion of the venue set as new markets open.