Principles
The core research principles and the engineering philosophy behind them.
Clawlas is opinionated on purpose. Two sets of principles govern it: the core principles (how the laboratory thinks) and the engineering philosophy (how the code is allowed to grow). Both exist to keep every decision explainable, measurable and remembered.
The operating principle
Deterministic systems detect and control. AI interprets and assists. The risk engine decides whether execution is permitted. Every layer escalates to the next only when it must, and no layer can bypass the one below it — deterministic filters run before any model call, and the model never reaches the exchange.
Core principles
These four shape every part of the system.
AI is a research partner
The AI does not replace quantitative analysis — it collaborates with it. The agents generate hypotheses, propose indicators and scoring models, challenge previous assumptions, and document every decision. The AI helps build the system itself.
Mathematics validate ideas
Every hypothesis should eventually be measurable. Whenever possible, intuition becomes mathematics: volatility models, liquidity metrics, holder growth, social velocity, anomaly detection. No assumption is trusted until the data validates it.
Human governance
The human always owns the strategy. Only the human validates strategic changes, deployment, capital allocation, execution permissions, scoring changes and new learning rules. The platform proposes; the human governs.
Memory first
The platform remembers every market state, hypothesis, trade, score, reasoning, execution, failure and improvement. The goal is not only to learn from markets, but to learn from previous decisions. Nothing is forgotten.
Research discipline
These are encoded, not documented-only — they protect capital and keep the research honest.
Evidence before capital
Nothing reaches real money before it survives the lifecycle: backtest, then shadow mode on live data, then simulation, then a progressive production phase with very small capital.
Deterministic execution
The execution layer is pure and replayable; the LLM layer is isolated from it. Risk caps live in code, never in the model’s judgment.
Everything logged
Every decision, its reasoning, and its outcome is recorded immutably. If it isn’t logged, it didn’t happen — and it can’t be learned from.
The platform proposes, the human governs
The agents run the research loop continuously; the human owns every move that reaches the market. When unsure, the system does less, not more.
High risk is part of the process
The lab intentionally hunts asymmetric, short-lived opportunities. Large losses are considered part of the research — provided each one is measured and remembered.
Continuous re-evaluation
No trade is sacred. At every market update each position must re-justify itself against every new opportunity, or the default action is to prepare an exit.
Engineering philosophy
The codebase follows KISS, YAGNI, DRY, SOLID — the most direct solution that works, no abstraction “just in case”.
- No premature abstraction. Prefer a local duplication over a wrong coupling. Mutualize only after three real usages.
- Deleting code is as valuable as adding it. The smallest diff that solves the problem wins.
- Short, intention-named functions. Business logic lives in pure, testable units; handlers and components stay thin.
- Feature-first vertical slices. A feature is autonomous and imported only through its public interface.
Why the discipline is structural, not stylistic
A research lab that can place real orders is one careless refactor from an unmanaged loss. Encoding the rules — allowlists, typed tool registries, per-resource authorization, the no-real-money key during research — is what lets the agents explore freely without ever being trusted on the dangerous actions.