The failure mode
Most pretty curves we get asked to "just productionize" were fit on the same window they are scored on. The plot is a memory of the sample. It is not a forecast, and it is not a track record.
Minimum bar before a note ships
| Check | A pass looks like |
|---|---|
| Walk-forward | The out-of-sample window was never used for tuning |
| Cost stress | The edge is still there at 2x the slippage haircut |
| Parameter plateau | More than one neighboring cell is green |
| Capacity | Written down in shares and in dollars, at the volume we can actually hit |
| Clock shift | The sign stays put when the rebalance bar moves by one |
A rule that flips because Tuesday became Wednesday was never a rule. Same for a Sharpe that lives in one ticker. If one name is the strategy, label it that way and stop pretending it is a book.
What we publish here
The research note states the hypothesis, the universe, the cost model, and the out-of-sample window. It does not publish a live allocation. If the curve only works from 2017 to 2019, that sentence goes in the first paragraph, not the appendix.
Tooling
The runner is boring on purpose: point-in-time data, a frozen feature snapshot, and a report that diffs against the last accepted run. A model can draft the commentary around the table. It does not get to pick the window, drop a name, or rerun until the chart behaves.
When a result fails one of the rows above, the note still ships — as a negative. Silent discards are how a desk talks itself into a curve.