CHIMERA: An Autonomous Trading Research & Execution Organism

Chimera is a working system. It ingests order-book and trade data at one-second resolution across crypto venues, options chains, and prediction markets. It trains and audits a fleet of thousands of models. It proposes new strategies from a knowledge graph of what has been tried. And it promotes a strategy to real capital only after the strategy survives six deterministic gates. Every number in this paper was measured on the running system, and every promotion decision is reproducible from frozen, hash-stamped artifacts.

§1System map: the Edge Factory

The factory pipeline is strictly ordered. Data flows down, capital decisions flow only through the gate layer, and the ledger under all of it is append-only. Each phase was built against a pre-registered gate that was written and frozen before any run started.

§2The neuromorphic base: navigation, not prediction

Chimera's decision core is a closed-loop neuromorphic active-inference architecture. It does not predict the future and wait. It acts, perceives the market's response, and updates its world model, navigating the probability surface the way biological organisms minimize surprise. A hub-and-spoke orchestrator routes market context to seven cognitive modules, collects their signals, applies the veto hierarchy, synthesizes the surviving decision, and writes the full audit record.

The three-clock topology

Fast execution and slow analytics are separated by computational cost and never mixed. Clock A (100 to 250ms) maintains orders and tracks queue position with zero database calls. Clock B (0.5 to 2s) runs hot features and the seven-module orchestration. Clock C (10 to 60s) handles heavy feature computation, reconciliation, and training. The models themselves form a prediction-horizon ladder: VPIN flow toxicity and circuit breakers act instantaneously, CNN-LSTM direction models work at one minute, the Hive at 15 to 30 minutes, ZoneX exit timing at 15 minutes, the regime classifier at 7 days, and the LPPL bubble detector at days to weeks. Every decision is informed at its own natural timescale, and an Ornstein-Uhlenbeck mean-reversion module owns the ranging-market regime that trend models cannot see.

§3The Greeks brain: one risk space for every venue

Direction is a dead wall; volatility is where the measurable edge lives. Chimera therefore runs one logical book across all products, expressed in 1st and 2nd order Greeks, defaulting to delta- and gamma-neutral so the volatility edge is held pure. Every instrument reduces to the same four numbers, so an equity option, a perpetual future, a VIX ETP, and a Kalshi bracket all live in one risk space and net against each other. A prediction-market binary is a digital option with computable delta and gamma, which is how event contracts enter the same book as listed options.

The engine has six parts: cross-venue position aggregation into a common numeraire; correlation-aware netting through a covariance matrix, converting cross-asset deltas to a common SPY-beta factor; a constrained quadratic-programming neutrality optimizer that finds the cheapest hedge set driving portfolio delta inside epsilon at target gamma; a capital-efficiency allocator ranking expected vol-edge P&L per unit of margin across venues, sized by fractional Kelly; a dynamic gamma-scalping rebalancer whose hedging is the P&L when the book is long gamma; and a sell-the-Greeks controller that flips the book short when implied vol is rich and calm is predicted. The honest bar is written into the design: PnL ≈ vega·(realized − implied) + gamma-scalp + theta − hedging cost. The book only wins when predicted realized vol beats implied, and it stays in paper and shadow until the gates say otherwise.

§4The feature state space: one catalog, every scale

The system computes roughly 400 features at 1-second and 1-minute grain: microstructure at 100ms updates, on-chain, derivatives positioning, NLP sentiment, and 48 FRED macro series, aggregated across timeframes from 10 seconds to 1 day, regime-conditional and cross-asset. Two disciplines keep this honest:

• The Harmonic Tensor Stack. A single feature registry (33 live specs across 9 dimensions, with period grids like EMA 10/15/20/25 as first-class instances) across 62 symbols gives 21,762 declared tensor instances, each with warmup accounting, readiness flags, and a dual-cadence declaration. Every experiment declares features by registry ID, and the zoo's parquet stores are mapped against the registry as canonical or legacy, so two parts of the system can never drift into private dialects of the same feature.
• The dual-cadence rule. 1s and 1m, always both, one definition. A single canonical resampler derives every timeframe from the same source through the same code path, and the trial registry records the cadence of every experiment. The F0 transfer check proved why this matters: production 1-second z-scores and 1-minute-bar z-scores agree on trigger sets only 42 to 51% of the time, which means materially different trades. Every exit and veto study is re-verified at the production cadence before promotion.

§5The model zoo and champion freeze

The zoo trains 14 model families across 496 underlyings and 27 labeled targets: gradient boosting, linear stacks, sequence models (GRU, Mamba, TFT, N-BEATS, PatchTST), cross-sectional graph networks, and tabular foundation models (TabPFN v2/v3). That is 2,629 trained models in the current fleet and 2,775 scored board cells. The operating rule: always serve the proven champion, never the latest experiment.

Champion freeze makes that rule a contract. Each frozen champion stores the model artifact, preprocessing, feature contract, training window, and seed, keyed by (asset, target), and is registered as a trial so every board score can be re-derived. A weekly rotation gate runs pre-audit, rotation scan, per-lane refreeze of only the changed symbols, then post-audit, and it fails loud: a mismatch stops the pipeline instead of silently serving a stale model. The audit exists because it already paid for itself. It caught 169 champions silently frozen as the wrong model class, correct-looking contracts that would have served garbage scores, and all 169 were refit and re-audited to zero mismatches (§12).

The signal service exposes all 14 serve paths, bitwise-repeatable on repeat calls. One family carries a special guard. TabPFN v3 inference showed sub-per-mille GPU jitter on the previous stack, so every v3 serve now requires dual-predict agreement within 1e-9 or the request fails with HTTP 500. A jittered score can never reach a deterministic gate silently. All 228 TabPFN3 champions are live-capable under this guard.

§6Execution quality: where the edge actually lives

The factory's first evidence pack settled the hierarchy: entries are weak levers, exits and states are strong levers, and fills carry the edge, since nearly every arm dies at naive taker costs. The execution stack is engineered as a first-class system on academic foundations (Avellaneda-Stoikov market making, Almgren-Chriss optimal execution, VPIN flow toxicity, order-book imbalance):

• TruthMesh price oracle: multi-venue aggregation with robust-median outlier rejection and staleness checks, so no single venue can lie to the system.
• Pac-Man maker-first entry: passive placement with timeboxed escalation; the F2 fill model tunes patience per state instead of a fixed timeout.
• Chandelier exits: volatility-adjusted trailing stops (2 to 4× ATR regime multiplier), a championship family member tuned per cell, never a default.
• Toxicity-aware action ladder: pre-trade adverse-selection detection (microprice drift plus order-book imbalance) with pause, reduce, and standard tiers.
• Inventory velocity engine: Avellaneda-Stoikov skew moves the reservation price against growing inventory, so the system is paid to flatten, not to bag-hold.
• Economic gate, intent dedupe, target hysteresis: no order fires unless expected edge exceeds measured friction, duplicate intents collapse, and targets do not flap on noise.

The position lifecycle state machine

Every position moves through an explicit machine, FLAT → ENTRY_PENDING → LIVE_POSITION → BANKING → UNWINDING → COOLDOWN, with a DEFENSIVE branch, and every transition is logged with its trigger and full context. Health monitors act without humans: a time-to-live cap forces exits, a rent meter charges opportunity cost per second and exits when rent exceeds P&L, a toxicity score drops the book to defensive mode, and a maximum-adverse-excursion breach unwinds. Options order management self-heals the same way: stale working entries re-price by cancel-replace at fresh transactable credit under a bounded budget, failed legs retry, and an end-of-day settlement pass closes the expiry-window coverage gap.

§7The learning system: paid for velocity, not just profit

Chimera optimizes the velocity of money rather than win/loss: reward = (PnL × velocity_multiplier) − (max_drawdown × 3), where a sub-15-minute trade earns a 2.0× multiplier and a 4-hour tie-up is penalized at 0.5×. A fast +2% beats a slow +5%. Learning runs at each component's natural timescale: BDH Hebbian feature-outcome associations in milliseconds, Hive and strategy trust per trade, Shield calibration per decision, and meta-controller weights daily. Three structures make learning cumulative:

• Unified experience memory: every trade, veto, and counterfactual lives in one append-only schema, so any component can learn from any other's experience.
• Counterfactual tracking: the system scores what it did not do. Every veto and every skipped entry is tracked forward, producing reward and regret streams that feed both state maps and the swarm's hypothesis seeds. Gates that cost more than they save are demoted by their own ledger.
• The learning heartbeat: the pipeline monitors itself. Fills happening without reward, regret, or weight updates raises a stalled alarm. If learning stops, the system says so instead of silently coasting.

§8Validation: four layers, six gates

The distance between a good-looking backtest and live capital is where trading systems die. Chimera stacks four validation layers and makes the last one deterministic.

The six-gate ladder, stamped to an evidence basis

Every one of the 40 ladder strategies is re-scored on every 15-minute heartbeat against six deterministic gates: stage0 · dsr · dsr_daily · walk_forward · lcb · beats_1N. The dual-DSR basis (stamp v2-dual-dsr, minimum 20 days) requires significance on both the per-trade and the daily-aggregated return series, because a strategy can look significant per-trade while its daily P&L is noise, and vice versa. The gate set is pre-registered, the before/after snapshots are frozen and hashed, and wiring it in changed zero existing verdicts.

Why dual-DSR matters: the flip table

Source: frozen dual-DSR study (dual_dsr.json, sha256 fb6466a5…), reproduced from the ladder's own dated-net series. Of the strategies studied, 5 pass per-trade, 6 pass daily, and 3 flip verdict. A single-basis gate would have misjudged all three.

Trial accounting: the machine that catches overfit

Discovery is encouraged to run to the edge of overfitting because validation carries the accounting. Every backtest is a row in research.trials (config hash, data window, seed, family). Promotion thresholds deflate with the logged trial count (z≥2 first-of-family, z≥3 to 4 for pick-of-thousands). Pass/fail bars are pre-registered before runs, and every report ships IS-to-OOS rank correlation, permutation nulls, and per-decade stability. Two worked examples: the sliding-Z entry-timing hypothesis failed its frozen bar (best improvement z=+1.61 against a +2.0 bar, with one cell showing significant harm) and was ledgered dead; the naive predict-the-winning-trade meta-model scored AUC 0.498 to 0.523, a coin flip, and stays a discovery screen rather than a gate.

§9Loop engineering: every engine has a clock

Nothing in Chimera runs when someone remembers to run it. Each loop has a named schedule, an append-only run record, watermark-based refire after downtime, and fail-loud logging. The operating cadence:

The nightly swarm loop follows the published loop-engineering discipline: triggered agent runs bounded by machine-checkable stop conditions, persistent state files, a verifier sub-agent that re-runs winners on different seeds and rejects without appeal, a skeptic that writes the kill-memo for every survivor, budget caps, and human escalation only for cells that pass both. Approved cells arrive as morning decision cards, not as log lines.

§10Graph engineering: the factory's memory

The 166-edge strategy knowledge graph makes the research process itself a queryable object. Nodes are entry rules, targets, exit families, sizing rules, state tiles, and cost models. Edges are validated combinations, weighted by deflated OOS decision value and tagged with regime coverage and venue. What this buys:

• Composition without re-search. When the chandelier exit validated on SOL high-vol, the graph immediately proposed it for every compatible cell; those five proposals are the swarm's current hypothesis queue.
• No zombie retests. A dead idea is an edge with an outcome, so the swarm cannot waste a night re-discovering a known failure.
• Decay propagation. A demotion applies a 30-day half-life weight decay and flags every dependent strategy automatically. "Why is this arm off?" becomes a query, not an investigation. Three demotion events are live in the graph today.
• Ledger round-trip as the gate. The graph passed its promotion gate by proving that a graph demotion propagates to an arm flag and then to portfolio-manager action, end to end.

§11Geometric reasoning: portfolios as shapes

Chimera treats correlation and risk as geometry, not scalars. The trial-correlation matrix is an ellipsoid, and its eigenvalue participation ratio is the portfolio's effective N: how many independent bets the book actually holds, which is the honest denominator for every significance test. The regime map is a state space (asset × vol-state × trend × session) where each strategy owns the cells it proved. Canary exits trail a chandelier in favorable-excursion space (3.0×ATR(14)) rather than in raw price, so the exit rule lives in the geometry of the trade's own excursion distribution. Options exposure is managed on Greek surfaces across all monitored underlyings. The Greeks brain (§3) is geometric reasoning applied to the book: every position a point in (delta, gamma, vega, theta) space, with the optimizer moving the whole portfolio along the cheapest path to the target posture.

§12Recursive self-improvement: failures become permanent guards

The system converts its own failures into structure. Three worked examples from a single 48-hour window:

The stale arm. The reconciler flagged ARM_UNJUSTIFIED: volexp_condor_monthly held a live arm while its ladder stage was SHADOW. The root cause was 60 managed exits batch-appended to the ledger inside a two-minute window, which diluted per-trade Sharpe from 1.10 to 0.545 and flipped DSR below gate. It was a measurement artifact on a book standing at +$12,436 over 94 trades with a 74.5% win rate. The fix shipped the same day as three permanent changes: the reconciler now auto-disarms stale arms with an audit row, the strategy was registered in the meta-gate registry so the omission cannot recur, and its scheduler was repaired to fire inside its intended session window. The dual-DSR sixth gate (§8) exists because of this incident class.

The backtest that lied. The F0 parity harness found that flat (pinned-book) windows collapsed a rolling standard deviation to exactly zero, producing z = ±infinity, clamped to a sentinel that forced entries regardless of signal. 98.1% of DOGE MAX-EV's backtest trades were this artifact, and 52.4% of SOL MAX-EV's. Both sides were fixed (stable rolling std plus a flat-window no-signal rule), the live executor was re-cut onto the canonical engine mid-session with three open positions adopted safely, and the final parity gate passed bitwise with 9,218 of 9,218 identical trades. Historical conclusions were re-issued on the fixed math: SOL MAX-EV flips from negative to positive out-of-sample once the artifact is removed.

The wrong-class champions. The champion audit caught 169 frozen contracts whose artifacts were the wrong model class entirely. They would have served garbage scores under correct-looking paperwork. The freezer now raises on unknown model classes (fail-loud, no silent fallback), the 169 were refit family-correct, and the audit gate is wired into the weekly rotation so this class of bug cannot return. Incident, root cause, gate, frozen artifact: that sequence is the improvement loop, and it runs on the system's own mistakes.

§13Strategy evolution: alpha that fights back

Published alpha decays on a known curve: a Sharpe of 2.5 becomes 0.7 within a few years as the crowd replicates the trade. Chimera's answer is to make strategy generation adversarial and continuous. In the self-play framework, a Challenger agent invents adversarial market scenarios (a stop hunt with a fake volume spike, a liquidity mirage) and is penalized when the Solver survives too easily. The Solver evolves strategy logic to survive the attacks. The two co-evolve, so strategies emerge that no human designed, already battle-tested against manipulation patterns not yet seen in live markets. Complementing it, the Social Sentinel ingests strategies published by traders across financial media, filtered for audience and capital, and tests each one against the same frozen gates. Most die in validation; the survivors are assimilated as new tensor cells. The evolutionary loop never sleeps: decay detection on live edges, demotion propagation through the graph, fresh hypotheses from the swarm. The factory treats alpha as a renewable resource precisely because it refuses to believe in permanent alpha.

§14The operator surface: every page is a factory instrument

The command center is the factory's private operating surface rather than a dashboard, with a public live view at daniel-roberts.com/live. The Overview tape shows real money only: open TastyTrade structures and Kalshi positions marked live, with no shadow or paper P&L blended in. The Strategies page renders every one of the 40 ladder strategies with all six gates visible. The Decision Queue surfaces any canary that clears its frozen bar as a GO/NO-GO card with its full evidence pack and skeptic memo, and a persistent banner warns until actioned, so the human never has to poll logs to discover a decision is needed. The fleet board, the fill-adjusted alpha gates, the veto reward/regret ledger, the money-leak ledger, and per-source data freshness each have their page, and each feeds a factory input.

§15Live evidence and honest economics

Three Kalshi mean-reversion arms are live on the v2 canonical engine. Three more were built, shadowed, and demoted by the gates when they could not prove themselves. The demotions are as much the system working as the promotions.

Live arms: operator P&L card, week of 31 Aug 2026. Demotions: ladder stage events. DOGE SMOOTH's guarded baseline shows why honesty matters: +6.33bp/week gross, but −3.67bp/week net at the maker-only tier-0 fee schedule (n=753 fills, 46.6 days, 5,946 sentinel ticks dropped). A strategy that looks alive gross and is dead net never gets armed.

On the options side, two volatility-expansion strategies (0DTE c25 and the 0DTE condor) hold CANARY stage at $1 size with frozen kill criteria. The SOL chandelier challenger runs its live A/B canary against a frozen promotion bar (n≥20 paired closes, challenger beats incumbent, Welch z≥1, top-day concentration ≤50%) and will surface as a Decision Queue card if it clears. Autonomous realized P&L across TastyTrade and Kalshi stood at roughly +$100 at the end of August 2026, small by design, because the canary layer caps size until the evidence says otherwise. All economics are quoted at post-promotion fee schedules: Kalshi's fee-free promotion ended in July 2026, and every number in this paper is net of the tier a real account actually pays.

§16Doctrine and road ahead

Five rules the system enforces on itself: backtests never promote; every loop has a clock and a ledger; every model served is a frozen, audited champion; every gate is deterministic and reproducible from a frozen artifact; and every failure becomes a permanent guard. The factory proposes, the swarm tests, the verifier rejects, the skeptic writes the memo, the gates decide, the human authorizes size, and the graph remembers.

Next, in gated order: the F5 swarm's first full unattended week (gate: every survivor traceable to registry, verifier log, and skeptic memo); the SOL chandelier canary's verdict; the state map's per-symbol causal extension; the Greeks brain's neutrality optimizer graduating from paper to shadow; and venue expansion across new TastyTrade underlyings, new Kalshi series, and Polymarket's captured-but-unprobed tape, which stays frozen until the probing phases say the ideas are good. The machine is running.