Today — a daily running system and a pilot as the next step
What started with one question now runs every night. The testing ground has done its work; the next step is a pilot at an external organisation where the same question is alive.
Today the platform runs every night on the actual production data of the company where it originated. The question that started everything — "why do margins diverge from what the contract promises?" — gets a concrete answer every morning: per contract, with amount, with evidence and with an explanation a controller could sign.
What the testing ground delivered
- A foundation at the source. A nightly chain that does not touch operations yet offers a controlled, historised reality to the agents.
- Orchestration with discipline. Seven explicit steps, each with its own input and output, no agent above its station.
- A quality gate with teeth. Model output that fails the figure-exact, elimination, miss and maturity checks does not reach the user.
- A thinking model that scales. Lenses measure, code decides, model phrases. New checks take days, not quarters.
- A sovereign decision layer. Explainable, versionable, and robust against model swaps.
What comes next
The testing ground has proven two things. First: the approach works on real data of a real contract-heavy organisation. Second: the approach is separable from this one company — the rules, the quality chain and the thinking model are sector-neutral even though the context in which they matured is not.
The next step is a pilot at an external organisation where the same question is alive. Not as a demo, but as a real nightly chain on real data, with the same discipline: orchestration, verification, deterministic decision, and model language only at the end.
The real lesson
The journey from question to daily running system was longer than we hoped and different from what we expected. Every mistake hurt here first. Every lesson is now a rule in the system. That is what a testing ground does — and why the platform exists.
THE LESSON The most important choices were not technical but disciplinary: which question to ask, where to let decisions live, and when to say no to a convincing answer.