The mutual action plan agreed with a customer now records what was promised and what actually happened. The third piece, how far each step slipped, turned out to be unable to hold a negative number, so a step finished ahead of its date and a step finished exactly on it would have stored the same value.
- This plan is the only place in the sales model that records actions the customer owns rather than ours, so slippage split by side answers which side a deal stalls on. That is the reason for building it, and it depends on the sign. A step delivered early is evidence about how a customer works, and clamping it to zero throws that evidence away at the moment of writing.
- The column was defined with a minimum of zero. Nothing had ever written to it, so nothing had ever failed, and no report had ever looked wrong. A field that nothing writes to cannot fail, which means its defects are only ever found by reading its definition rather than its data. This one surfaced because a ruling forced the question of whether it could go negative, not because anything went looking.
- Storing the sign is reversible and flooring it is not. Any view can show lateness alone whenever that is what is wanted, but a value clamped on the way in is gone for good. The floor was lifted and then proven by writing a negative to a real row and reading it back, rather than by trusting the change to have applied.
- Setting the formula that fills the column was accepted, reported success, and silently thrown away, while every other setting sent in the same request applied correctly. What that leaves behind has the right type, the right range and the right label, and computes nothing. Reading the definition back is no help, because a correct looking definition is exactly what each failed attempt produced. The only check that works is planting rows with known answers and reading the numbers out.