The rules the models are held to.

Primary sources, stamped with when they were knowable.

Every series on CutPoint comes from a government or regulatory primary source: EIA petroleum data, FRED macro series (WTI and Brent spot, the fed funds rate), SEC EDGAR capital-expenditure filings, and TSA passenger throughput as a jet-fuel demand proxy.

Each row is stored with two timestamps: when the observation happened, and — separately — when it became publicly available. A weekly EIA figure describing last Wednesday isn't knowable until its release day; our schema records both facts and never confuses them. Backfills can only move the 'knowable' stamp earlier if we learn it was published earlier — a database trigger enforces this.

Models only see what the market could have seen.

When we build a training row 'as of' a given day, every feature is joined on feature_available_at — the knowable stamp — not the observation date. This is the point-in-time discipline that kills look-ahead bias, the most common way published trading models cheat (usually by accident).

The join is tested: our SQL test suite includes a look-ahead leakage check that fails if any feature row could leak future information into a training frame.

Walk-forward, fold by fold, against a baseline that must be beaten.

A model is trained on a window of history, then scored on the period immediately after it — data it has never seen. Then the window rolls forward and the process repeats. Each of these folds is an independent out-of-sample test; the numbers we publish are aggregates across all folds, never a single lucky split.

Every model is also scored against a naive 'predict no change' baseline. If it can't beat that, the card says so.

Intervals, not points — and we report when they break.

Forecasts ship as three quantiles: p10, p50, p90. The band between p10 and p90 is the model's 80% confidence claim, and the model card reports its actual coverage — how often reality landed inside the band. A calibrated model hovers near 80%; overconfidence shows up as red dots on the chart, in public.

Current production models are LightGBM quantile regressors. The registry is target-agnostic — as coverage grows, new targets and model families land in the same public format.

Model cards are public and permanent.

Every trained model gets a card in the lab: its feature configuration, hyperparameters, training windows, every fold's metrics, and every individual prediction with its realized outcome. The lab's glossary explains each metric in plain language — MAPE, RMSE, coverage, folds — so you can audit the claim, not take our word for it.