Skip to content

Repository navigation

research_lifecycle.md

Latest commit

 

History

History
52 lines (41 loc) · 1.74 KB

research_lifecycle.md

File metadata and controls

52 lines (41 loc) · 1.74 KB
Raw
status ACTIVE
owner research
applies_to_commit 8692afefe1f6cc82ab1f276fca788888f9f30f3e
last_verified 2026-08-28

Research Lifecycle

immutable inputs
    -> frozen feature/label/split/model/portfolio contracts
    -> per-fold fitted-state isolation
    -> ordered non-overlapping OOS predictions
    -> continuous research portfolio account
    -> Research Gate
    -> research artifact graph
    -> at most RESEARCH_PROMOTED

Every research run binds DataRelease, DatasetVersion, FeatureSnapshot, AlphaPack, LabelSpec, SplitSpec, model profile, strategy/portfolio policy and source revision. Changing any governed input creates a new run identity.

Isolation

  • fit processors and models per fold; never share fitted state across folds;
  • enforce purge, embargo and label-lookahead gaps;
  • stitch only ordered, non-overlapping OOS predictions;
  • keep final holdout sealed from feature/model/hyperparameter/portfolio selection;
  • preserve one continuous portfolio account across rolling fold boundaries;
  • fail closed on cache, checkpoint, snapshot or lineage drift.

Gate logic

The production stability branch is:

(ICIR >= 0.50 OR RankICIR >= 0.50)
AND ExcessIR >= 0.50
AND all other research, portfolio and lineage gates

RESEARCH_REVIEW requires ICIR >= 0.30 OR RankICIR >= 0.40 plus all other hard conditions. It preserves evidence but does not publish an execution candidate.

Active program

The stability-diagnostics program performs only temporal-stability and causal-regime diagnostics over frozen rejected candidate-program anchors. It creates no formal candidate, performs no model selection, opens no final holdout and authorizes no publishing. See Alpha Research Stability Diagnostics.