Decisions

Load-bearing algorithmic choices — thresholds, gates, blend weights, source-selection rules. One page per decision.

Status: Phase 3 placeholder

The full set of ~28 decision nodes lands in Phase 3 — this is the highest-value pass of the KB. Each decision links to the field it gates, the metric it affects, and (where relevant) the forecasting method it routes to.

Decisions to be documented (grouped by source archive doc):

From hybrid_growth.md / hybrid_growth_timeframes.md

• Volume > 200 display gate
• Growth metrics > 200 strict gate
• Tier-1 / Tier-2 / Tier-3 history-window boundaries
• YoY vs rolling vs MoM tier routing
• Seasonal-keyword detection threshold

From confidence_score.md

• Coverage source weighting (GSC + 2-trend > GSC + 1-trend > GSC-only)
• Freshness staleness ladder per source
• Forecast reliance scoring
• Geometric-mean subscore floor

From refresh_detection.md / REFRESH_PIPELINE.md

• Surge / staleness / divergence / missing-data signals
• GT tier assignment for extraction priority

From processing.py directly (no archive doc)

• ADI/CV² trend-type quadrant classifier (routes to forecast method)
• GT validity (max(gt_trend) > 50)
• GT/GSC correlation gate (suppress GKP when redundant)
• GKP blend weights (0.8 trend / 0.2 GSC, with proportion-based hard override)
• GT 2× soft cap
• GT/GKP overlap-cut rules (three sub-rules)
• Bot spike threshold (peak / baseline ≥ 500)
• Cross-source spike corroboration (GT ≥ 3× median, GKP ≥ 3× median)
• Recovery-window gate (≤ 0.2× peak within 4 months)
• Completeness² extrapolation blend
• GT collapse override
• Forecast model selection cascade (score < 0.2 → score < 1 → lowest count-zero)
• Forecast outlier filter (max forecast > 1.5× max training)
• Forecast decay floor (0.85^h × historical_min)
• Isolated GT peak forecast skip
• Device-distribution threshold (latest month ≥ 1000 impressions)

(Final list and grouping may shift as Phase 3 probing surfaces additional or reclassifies existing decisions.)