Preserve chronology
Research records are organized so that each decision can be evaluated using only information that was available at that time. Later outcomes do not alter the original candidate state.
Separate market regimes
Structurally different periods are identified before primary evaluation. Older data may support context and causal feature warm-up, but it does not automatically represent the current market regime.
Keep a sealed evaluation set
Development results are not treated as final evidence. A lockbox period is held apart from repeated tuning and used to test whether an observation survives contact with unseen data.
Record rejection
Research that fails to remain stable is retained as a negative result. A rejected hypothesis is useful when it prevents the same weak assumption from being rediscovered later.
Publish boundaries, not execution logic
Public documentation explains the research question, data class, validation state and limitations. Proprietary rules, thresholds, current candidates and system internals remain private.