Pipeline Nowcast
Are we on track, drifting, or in trouble — and are our predictions proving right?
Two jobs, one algorithm. The nowcast synthesizes five operational signals into a composite variance score. The prediction evidence system instruments our thesis predictions with live data so conviction scores update from evidence, not opinion.
The Gap
76 predictions in a database. Conviction scores assigned once. Zero evidence feeds updating them. The Bayesian protocol exists on paper — no automated triggers. Meanwhile, the commissioning dashboard tracks 74 features at L0-L4 with no timestamps on state transitions, so velocity is invisible.
The nowcast closes both gaps: operational variance detection (are we on track today?) and prediction validation (are our bets on the future proving right?). Same normalization, same decay, same composite scoring. Different time horizons.
Context
- Sales CRM & RFP — Pipeline + activity signals flow in
- Agent Platform — Agent velocity from Convex event stream
- Commissioning Dashboard — L-level progress this algorithm reads
- ETL Data Tool — Upstream data readiness + market signal collection
- Prediction Database — 76 predictions awaiting evidence
- Data Footprint — The commissioning instrument for data maturity
- Scoreboard — Where nowcast output surfaces
Questions
What happens when the algorithm disagrees with your gut — and what happens when your predictions disagree with the data?
- If the composite score says "on track" but one signal is critical, is the weighting wrong or is the signal unimportant?
- At what signal coverage does confidence become actionable — and what do you do below that threshold?
- When nowcast catches drift 3 days early, what process fires that wouldn't have fired otherwise?
- If a prediction you scored 4/5 has zero evidence entries after 6 months, is the prediction wrong or is the collection broken?
- What's the minimum evidence that justifies changing a conviction score — one data point or a trend?