Historical skill and prospective evidence are easy to mix accidentally. A model can be frozen today, then appear to collect “prospective” observations that were already present in the research dataset before the freeze. This protocol was designed to prevent that.
Freeze state
The Conditional Spread research branch created a prospective freeze with a recorded frozen_at_utc and a matching activation_not_before_utc. The forward state also recorded accepted and settled prediction counts.
At freeze, already-existing historical bars were checkpointed and explicitly marked as ineligible to become prospective observations later.
Frozen calibration
The freeze preserved the calibration policy used by the model rather than allowing a quiet recalibration after prospective evidence began:
- Quantile calibration:
ADD - Probability calibration:
ISO
Historical evidence remains historical
A later certificate bound development-skill evidence to the frozen research object while keeping the prospective status pending. That distinction is visible in the verdict:
HISTORICAL_DEVELOPMENT_SKILL_EVIDENCE_BOUND__PROSPECTIVE_CERTIFICATION_PENDINGThe point is not the length of the verdict. The point is that it refuses to let “evidence exists” silently become “future performance has been demonstrated.”
Why this matters
Prospective validation is not created by changing a filename or by rerunning a historical dataset after a freeze date. The observation itself must become available after activation under the frozen contract.
Limitations
- A clean freeze protocol does not prove that the model has useful predictive skill.
- Prospective evidence was still pending in the captured research state.
- Operational clocks, feed availability and settlement logic must remain auditable for the protocol to retain meaning in production.