Point-in-Time
Only information that was actually observable before the decision point is allowed into a feature.
Make the test harder to fool.
We do not design tests to make a hypothesis easy to confirm. We design them so that a false conclusion is difficult to survive.
Prospective OOS
Instead of tuning rules after seeing past results, we define a new prospective start and evaluate only data collected afterward.
Falsification
We design placebo, reverse-direction, and timing tests specifically to find ways the hypothesis could fail.
Quarantine, not deletion
When a data-integrity incident occurs, we quarantine the affected records rather than deleting them, preserving the original provenance.
Blind maturity gates
Effect metrics remain blinded until the sample reaches maturity, reducing the temptation to tune the study after seeing the result.
Manual promotion only
Even strong research results never change execution settings automatically.
Consumer availability
An old event timestamp does not prove that the feature was available at decision time. We track first-seen availability from the consuming system's perspective.
Exact lineage
Where possible, decisions, references, features, and endpoint identities are linked end to end. Any upstream lineage that cannot be reconstructed remains explicitly UNKNOWN.
Verified recovery
A backup is not trusted simply because the file exists. We restore it into a separate namespace and verify row counts and schema.
Durable research journal
Hypotheses, failures, reasons for changes, and evidence locations are recorded in a separate journal and preserved in checksummed off-host backups.
Unknown is not a PASS. Anything that cannot be verified remains explicitly unknown, and historical invalid or pilot arms are never merged into a new confirmatory arm.