There is a version of this page that talks about model architectures. It would be the least useful page on this site. The architecture is rarely what decides whether a deployment works; sequencing is. So this is about sequence.
1. Spring — Instrument
You measure before you move. Nothing autonomous happens first. First the ground gets instrumented — every signal that a human operator was quietly reading with their eyes becomes a number with a timestamp. Most organisations discover at this stage that they have been running on folklore.
Spring is unglamorous and it is where the whole thing is won or lost. A model trained on an unmeasured process learns the folklore too.
2. Summer — Predict
The model gets better than the average operator. With enough instrumented history, prediction stops being a demo and starts being cheaper than the alternative. Not better than your best person on their best day — better than the average of everyone, on every day, including the bad ones.
This is the threshold that matters commercially. The question is never "can it beat an expert." It is "can it beat the median, at scale, at 3am."
3. Fall — Act
Prediction turns into action without asking. The model stops producing recommendations for a human to approve and starts closing the loop itself. Fall is the season of organisational argument, because this is the point where the job actually changes rather than getting a new dashboard.
Everything before this was a pilot. This is the part companies announce and then quietly postpone for two years.
4. Winter — Endure
It keeps going when the humans go home. Winter is the only season that proves anything. Endurance — through the night, through the cold, through the quarter where nobody is watching. The human in the artwork is not being defeated. He is walking out at the end of a shift that no longer requires him to stay.
The W-2 soldier does not lose. The W-2 soldier gets to leave.
Why the order cannot be rearranged
Every phase consumes the output of the one before it. A model trained on an uninstrumented process learns the folklore that was standing in for measurement. A closed loop on an unvalidated model automates a mistake at machine speed. And endurance on an unclosed loop is just an expensive machine waiting for a human to press a button.
The failure is almost never the model. It is attempting Fall before Spring is finished.
Where the data comes from
Prediction is only as good as the history behind it, and most organisations have far less usable history than they believe — years of records that describe what was decided but never what was observed. Predictive Data & Technology supplies the external data layer that fills those gaps, and R0cketShip provides identity resolution and the deployment plumbing so a model scored in one place can act in another without a six-month integration project.
What this costs you in practice
- Spring is months, not weeks. Anyone selling you a two-week instrumentation phase is selling you a dashboard.
- Summer is where the ROI is provable — and where most programmes should stop for a while and bank the win.
- Fall is a political phase wearing a technical costume. Budget for the argument.
- Winter is the only phase that proves anything. Everything before it is a pilot.
The full argument is in the white paper
Operation Vivaldi — the doctrine, the architecture and the industry case, as a PDF.
Request it