Proven on our published instrument. Re-certified on your data before anything releases. Every certified model earns its grade the way that model should be judged — and when it can't, the run stops rather than ships. Here is what that looks like, family by family.
There is no single certification story, because there shouldn't be. A decomposition has to sum exactly to the change it explains; a forecast is proven by being tested on history it never saw; a score by beating a baseline on your own data. Showing you the right proof for each model is the difference between selling rigour and demonstrating it.
Certified
Certified: proven on a realistic, client-like book, run end to end through the deployed system — and re-certified on your data before anything releases.
Decomposition — the bridgeOn the shelf
Reconciled exactly, every time.
This isn't a statistical grade — it's arithmetic honesty. The decomposition is exact: every part sums to the change, with nothing left over. Its check against your own reported totals is being switched on; until a run certifies, this page says so.
Explains, never invents
It attributes a change that already happened, within the mix/rate split — it does not reach past the data into causes it wasn't given.
Refused on thin cuts
Below a minimum row count for any segment cell, the run is refused rather than run on a statistically thin slice.
ForecastingCertified
Tested on the future it hadn't seen.
Out-of-sample validation
The forecast is scored on held-out history — trained on the past, tested on the future it hadn't seen — and beaten against a naive baseline. It ships with that scorecard and a data grade from your own numbers.
Honest intervals
The confidence range is calibrated to contain the outcome about as often as it claims to, measured on your own history — not a band drawn to look reassuring.
Withheld, not guessed
Horizons and confidence levels the data can't support are withheld rather than fabricated. A series too short to forecast honestly is refused.
Anomaly detectionCertified
Judged against your own history.
Judged out-of-sample
Each period is measured against an expected band built only from data available before it — never fitted to the point it's judging.
Reports its own coverage
The band is calibrated on your history and the scan reports the coverage it achieved, so a flag carries a confidence read rather than an assertion.
Unscannable, not normal
Periods too early to have an expected band are marked unscannable, not quietly called normal. Too little history to judge, and it says so.
Propensity — scores & ranked listsCertified
Certified out-of-time, refused below the floor.
Trained on past, tested on future
The model is scored on rolling windows of your own history — trained on the past, tested on the future it hadn't seen.
Beats a baseline, and tracks reality
Ranking skill must beat a naive baseline and the scores must track observed frequency. The model earns its grade on both, or it doesn't publish.
Refused below the floor
Below a minimum of observed events, the run stops rather than certifies. A weak score is never shipped dressed as a strong one.
Survival — retention & decay curvesCertified
Certified strictly as far as your data supports.
Beats your history, or nothing ships
Every prediction is made out-of-sample, and the model must outperform the best forecast your own past rates could give — by more than your data's own run-to-run variation.
Checked where you'll read it
Predicted against observed at every horizon shown — not a sample of them — inside a tolerance sized to the plans the numbers feed, with bounds on every curve from your own book.
Certified to your horizon, not ours
Your history earns a certified window and the curve stops at its edge — a thinner book gets a shorter window at full strictness, never a softer test.
Saturation — response curves
The bend, located — or a plain refusal.
Graded on your own spend
The curve must be identifiable from your own spend variation, its shape must describe your response near the bend, and the bend must be pinned tightly enough to plan against. Miss any one and the run stops. And if your spend could have shown a bend and none appears, it says so: your spend hasn't reached saturation in the range you've run.
Only the spend you've run
What it certifies is located inside the range you have run, and the platform makes no claim about spend you have not tried.
Diminishing returns only
Validated for channels whose returns diminish from the first dollar. Channels that need a threshold of spend before they take off are not covered.
Attribution — channel contribution
Credit, with the limits stated.
Order is tested
Whether the order of touchpoints carries information is tested, not assumed — against the same model read without order, at a stated level.
Shares with honest intervals
Each channel's share of the credit carries an interval proven to cover over time, not merely across customers. A channel too thin to measure precisely has its share shown without a range, and the platform names why rather than implying a precision it does not have.
Two channels and a comparison
It needs at least two channels to divide credit between, and journeys that didn't convert to compare against. Without either, the platform refuses.
How the certification is built
The gates are tested before they're trusted. Every threshold is set before results are seen — never adjusted to let a model through. Every gate is proven able to refuse before it's believed when it passes: we construct data that should fail, and confirm it fails. And when a run stops, that's the system working — a halt is the certification doing its job, not an error.
The full record
The proof, written down.
Everything here is the plain-language version. The full certification records — the instrument closes, the validation scorecards, the methodology in detail — live in the client security-review document, scrubbed to the same standard your IT team will read. We'd rather you see the proof before you ask.