The agent proposes economic hypotheses; deterministic tools evaluate them. Attempt history, statistical selection bias, and strategy stability inform the next round of exploration. Research diagnostics and final candidate validation have separate roles.
Knowing what has been tested and why it failed helps the agent decide what to change. It reviews previous outcomes, the current research scope, and available data and analytical tools before preparing a new hypothesis.
What happens here
01
Distinguish completed experiments from interrupted work to avoid repeating the same evaluation.
02
Check the research scope and available data and analytical tools.
03
Verify the link to previous outcomes and learning records.
What goes to the next step
The agent reads the previous result and decides what to change in this cycle.
01 / 06
Learning from each experiment
Conditions where returns disappeared and assumptions that failed guide what to test next. The agent reads the latest evaluation and research record before writing the next hypothesis.
RESEARCH METHODOLOGY
Research safeguards
A broader search creates more opportunities to misread a favorable result. The loop considers how a result was obtained and the conditions for validating it, alongside the return itself.
01
MULTIPLE TESTING / DSR
Selection bias after repeated trials
Testing many strategies increases the chance of finding an impressive Sharpe by luck. The loop preserves attempt history and uses the Deflated Sharpe Ratio (DSR) during exploration to account for selection bias and estimation uncertainty.
Method and interpretation
Diagnostics use daily return statistics, sample length, trial counts, and return-distribution skewness and kurtosis. Evaluated variants, interruptions, and partially exposed attempts are recorded under defined rules alongside original proposals. DSR informs exploration; it is neither a guarantee of future returns nor a substitute for final validation.
02
LOCAL ROBUSTNESS / DEDUPLICATION
Parameter sensitivity and near-duplicates
Local parameter variations help reveal results that depend on a single favorable setting. The candidate-selection protocol distinguishes variants of the same idea and strategies with near-identical return paths to reduce duplication.
Method and interpretation
The comparison neighborhood is bound to the executable program, preventing favorable neighbors from being chosen after observing results. The protocol considers stability across neighboring settings, selects lineage representatives, and checks return-path similarity. Neighbors also count as attempts; candidate ordering does not use final target thresholds as a search reward.
03
QUALITY DIVERSITY / FEEDBACK
Preserving diversity in the search
Optimizing a single score can repeatedly produce similar strategies. A Quality Diversity (QD) archive retains strategies with different trading characteristics and supplies their outcomes as context for further exploration.
Method and interpretation
Classification uses characteristics measured from evaluations rather than strategy names or the agent’s description. Learning state is reconstructed from verified experiment records, and the agent consumes the previous outcome before authoring the next hypothesis. Learning here means evidence changing the next research decision, not retraining model weights.
04
FROZEN SELECTION / SEALED HOLDOUT
Separating research from validation
Research, candidate selection, and terminal validation use separate data roles. Candidates and decision rules are frozen before new validation results are observed, keeping iterative hypothesis revision apart from final assessment.
Method and interpretation
Selection takes the first candidate in the frozen order that meets the predefined survival rules. Terminal validation evaluates that same candidate once after human approval; failure does not permit switching to another candidate for a retry. Favorable research diagnostics cannot bypass this path.
05
NULL CONTROLS / REPRODUCIBILITY
Checking the evaluator itself
The code judging a strategy also needs scrutiny. The validation framework includes null controls and fixed test cases to check for false positives and changes in evaluator behavior.
Method and interpretation
Links between compilation, evaluation, and attempt records preserve the identity of the program and inputs needed for reproduction. Controls provide evidence about the evaluation procedure, not proof of a trading strategy’s profitability. This page describes the system’s design, not a validation result for a particular campaign.