Uncertainty is not weakness. It is a contract.
A forecast without a tested margin of error is not a promise — it is noise with confidence attached. In release R84 of the Camile World Model, the system relearns from the client's own most recent events: when the world changes pattern, the margin of error adjusts, with a minimum number of cases so it does not become unstable. And the science comes with a band, not a point: we repeated the experiment eight times, with different initial data, and the first result turned out to be a rare case — not the rule.
Uncertainty is not weakness. It is a contract.
When our system forecasts, it also states how much it does not know — with a margin tested against actual accuracy. This publication explains how that works in release R84.
Relearning when the world changes
The Camile World Model (CWM) platform publishes its R84 release. The novelty is an adjustment to the service that returns the margin of error: besides calibrating on the client's own verified history, it can now relearn from the most recent events, in a window that moves over time. To avoid oscillation, the window has a tested floor: it recalibrates only with at least 192 recent events. Window requests outside the rules are refused with an explicit error. Those who do not use this adjustment notice no difference at all: nothing existing changes behaviour — exactly as before.
The method is what sustains the number
Every hypothesis is written before any measurement; every verdict follows a criterion decided in advance that would state exactly what would prove us wrong; and every negative result or discarded hypothesis is published, not discarded. In round R84, we repeated the experiment with eight pre-registered seeds — different initial data chosen before collection, with no room for selection. The repetition showed that a theoretical formula holds as an upper bound for what is observed in practice — and showed, honestly, that the first attempt's result was a rare tail case, not the rule.
Where to check
Everything is verifiable on the platform that originated the facts: the current release badge, the 35 published releases with what changed in each one, and the updates page are at https://camileai.com/cwm. It is the house rule: nothing here claims more than the platforms show. If you want to follow the next change, that is where it appears first.
How we tested
The window adjustment is optional and documented in the public API contract; without it, behaviour is identical to the previous version, bit for bit. The R84 round experiment is recorded with eight distinct seeds and an independence criterion between them. Anyone who wants to can redo it: the published numbers can be reconstructed.