Every glowing sphere is one ML project attempt entering the pipeline from the left. Each of the three gates — POC, MVP, Production — rolls that project's fate against a pass probability built from your inputs: better data quality and team/MLOps maturity raise every gate's odds, and pushing more investment into a stage buys a further edge but raises its cost too. A project that fails a gate turns red and drops out — its stage costs are sunk. Only projects that clear all three gates reach Production and pay back the deployment value.
P(gate) = clamp(base + w1·dataQuality + w2·teamMaturity + w3·(investment−1), 0.03, 0.97)
ROI = (Σ revenue − Σ cost) / Σ cost × 100%
- Data quality — clean, sufficient, representative data. Weak data quietly fails projects at every gate, especially Production.
- Team & MLOps maturity — deployment pipelines, monitoring, domain expertise. This matters most at the MVP and Production gates, where "it works on my laptop" stops being enough.
- Stage investment — more budget per stage (better tooling, more pilot users, more careful validation) nudges pass odds up, but every stage attempt is charged that multiplier — so overspending on doomed projects erodes ROI just as fast as underspending does.
- Value per deployment — the business return booked only when a project survives all three gates and goes live.
This mirrors why "start small" pilot projects are the standard advice for ML adoption: killing a weak project at the cheap POC gate protects the budget from being spent on a doomed Production rollout.