Ignition — Predictive Scoring Engine

The Problem

I wanted to predict which products would achieve explosive adoption — not just “good” growth, but the kind of rapid, viral traction that makes or breaks a product. The question: can you score a product across measurable dimensions and predict its adoption trajectory with real accuracy?

The Architecture

Ignition scores products across 5 weighted pillars, each containing multiple features:

Pillar	Weight (v7)	What It Measures
Foundation	0.00	Team, funding, timing
Product	0.10	UX, speed, delight, activation
Distribution	0.11	Virality, channels, network effects
Environment	0.31	Market timing, frustration, wave alignment
Moat	0.38	Network effects, switching costs, data lock-in

The surprise: Moat is the strongest predictor (weight 0.38, pillar correlation r=0.73). Who the team is (Foundation) matters almost nothing for adoption prediction — what matters is whether the product creates structural advantages that compound.

18-table SQLite schema with full audit trail, change tracking, and calibration history.

7-signal wave detection: Tracks search velocity, incumbent frustration, technology maturity, regulatory shifts, investment patterns, creator density, and cultural indicators. Fires alerts when 4+ signals accelerate simultaneously >15% week-over-week.

Weight Learning

The system doesn’t use static weights. It learns from outcomes:

Score a product across all pillars
Record the actual outcome (dead, slow_burn, spike_and_die, steady_growth, explosion)
Run weight optimization to minimize prediction error
Accept new weights only if correlation improves by 3%+ (prevents overfitting to noise)

v7 calibration (current): correlation 0.799, tier accuracy 65%, ±1 tier accuracy 93%.

Outcome distribution across 671 cases:

151 steady_growth, 137 spike_and_die, 137 dead, 126 slow_burn, 120 explosion

Key Decisions

Why environment-normalized scoring: LLM-generated batch cases systematically inflated certain features (incumbent_frustration +20pp, wave_strength +8-20pp). Rather than discarding the data, I applied global scale normalization (0.67x and 0.87x) anchored to hand-scored ground truth. This preserved the valuable signal while correcting the bias.

Why honest accuracy reporting: The system reports 65% exact tier accuracy, not 93%. The 93% is ±1 tier. Both numbers are useful, but leading with the exact number keeps the system honest. If I only reported the forgiving metric, I’d lose the pressure to improve.

Why slow_burn is structurally hard: slow_burn outcomes overlap with both steady_growth and spike_and_die in their IQR ranges. This isn’t a model failure — it’s a real property of the data. Products that grow slowly look statistically similar to products that either die or take off. The model is honest about this ambiguity.

What I Learned

Calibration must compute from scratch. An early bug read stale accuracy values from a calibration_summary table instead of recomputing. The system appeared to be improving when it wasn’t. Apples-to-apples comparison on every calibration run is non-negotiable.
Feature importance surprises you. I expected Product quality to be the top predictor. It’s not even close — Moat and Environment dominate. The data says: it matters less how good your product is than whether the market conditions and structural advantages are in your favor.
671 cases is enough to learn, not enough to be certain. The correlation is strong (0.80) but the confidence intervals on individual predictions are wide. I could add more cases, but the marginal accuracy gain decreases with each batch. The honest answer is: this system is useful for relative ranking, not absolute prediction.