FOR AGENTIC AI
The white box behind the black box.
LLM wrappers, deep RL agents, autonomous research crawlers, and execution bots all share the same upstream dependency: statistically validated signal parameters. None of them solve discovery. We do. Enumerated. Deterministic. Explainable. The statistical infrastructure your agents can rely on.
Your models are brilliant at forming strategy. Let us be the calculator.
Large language models excel at helping users articulate research questions, reason about asset behavior, and assemble strategy logic. What they cannot do is run walk-forward survival tests, permutation null distributions with FDR correction, SPA bootstrap, or probability of backtest overfitting checks at scale. Not for a single permutation. Certainly not for millions in under a minute.
Allow S1 to become the white-label compute layer behind your conversational interface. Your model handles the reasoning. Our API handles the enumeration. The user receives statistically validated outputs that your platform can explain, cite, and audit. Beyond the signal itself, you can prove data destruction to both users and their regulators: ephemeral compute, ingress and egress certificates, cryptographic verification that raw data was purged after delivery.
Your PPO agent overfits because the feature space is wrong.
Standard deep RL pipelines feed raw calculated indicators into the observation space: RSI(14), MACD(12,26,9), Bollinger(20,2). These are arbitrary defaults. The agent learns to exploit statistical artifacts of those specific parameter choices rather than structural properties of the underlying price process.
S1 outputs are the result of exhaustive statistical tests performed on indicators, not the indicator values themselves. Each event row in the output is a fully labelled feature vector in Croissant ML format:
These are fully labelled, structured datasets. Every event carries its regime context, temporal metadata, and statistical measurements. Feed them directly into any architecture.
Any system that consumes tabular feature vectors can ingest S1 outputs without transformation. The statistical substrate is architecture-agnostic.
Typical deep RL observation space
With S1 enumeration outputs
Clawbots, copilots, swarm agents, and research orchestrators.
Autonomous research agents operate on your own infrastructure. They ingest data, form hypotheses, test conditions, and compile reports without manual intervention. The constraint is that quantitative math at enumeration scale requires purpose-built compute. These agents can call the Dojo API for on-demand enumeration without provisioning their own statistical engine.
The interaction pattern is API-to-API. The research agent decides which assets, families, and period bands to enumerate. S1 returns the complete solution set. The agent interprets, filters, or feeds results downstream. No human in the loop at the computation layer.
Clawbots
Autonomous scraping and research agents that continuously scan asset universes. Submit enumeration jobs per asset on a schedule. Flag regime changes when previously stable configurations break. Transact with metered compute plans so your bots pay per call without manual intervention.
Research Copilots
Assist human analysts by pre-computing enumeration across watchlists. Surface configurations that warrant further investigation.
Swarm Agents and Multi-Agent Systems
Coordinated agent fleets where each node handles a different asset class or research family. Distribute enumeration jobs across the swarm. Aggregate results centrally.
Strategy Orchestrators
Multi-step agentic workflows that discover signals, validate them, assemble candidate strategies, and pass them to separate backtesting or execution layers.
AutoGPT / BabyAGI / Task-Driven Agents
Self-directed planning agents that decompose research goals into subtasks. Call S1 as a tool within their action space for quantitative validation at each step.
Data Pipeline Agents
Airflow DAGs, cron-triggered notebooks, and scheduled compute jobs that integrate enumeration into existing quantitative infrastructure.
MT4, MT5, and autonomous execution systems.
Automated trading systems on MetaTrader 4, MetaTrader 5, and proprietary execution frameworks handle order routing, position management, and risk controls. They are built to execute. They are not built to discover which signal parameters are statistically valid across time.
These systems typically operate with hard-coded indicator settings chosen during initial development or manual optimization. When market regimes shift, the parameters may no longer reflect the statistical properties of the current price process. Periodic re-enumeration provides an updated parameter landscape that execution systems can reference.
S1 does not execute trades, manage positions, or provide timing signals. We provide the statistical substrate: which configurations satisfy specified constraints on a given asset at a given point in time. What the execution system does with that information is outside our scope. All services are non-discretionary.
THE STRUCTURAL GAP
None of these systems solve statistical discovery.
LLM wrappers retrieve information. Deep RL agents optimize policies. Research crawlers automate workflows. Execution bots route orders. Each operates within its own domain. None of them enumerate the full parameter space of technical indicators against empirical price data to determine which configurations exhibit statistically stable behavior.
This is the layer we occupy. Downstream of raw OHLCV data, upstream of any strategy, model, or trading decision. We compute across periods 1 to 14,000 on a one-minute canonical axis. Users select their own period bands, allowing weekly and fortnightly anomalies that influence intraday behavior to surface within a single enumeration run. The output is deterministic, explainable, and reproducible.
Enumeration Spec
Integrate enumeration into your agent stack.
Request API credentials. Inspect the engine in Dojo before committing.