Cones are belief states that know how to update
A language model that compresses its own history is performing a belief update. It fails in exactly the way a forecast must not: information lost, and nobody charged for it.
A belief state is a bottleneck: everything the agent might need later must pass through it, or it is gone. Berkeley’s recent work on belief-graded summarization makes the failure vivid, and it is the same bottleneck a probability cone lives in — with one deliberate difference.
Berkeley’s AI group recently published a framework, ABBEL (arXiv:2512.20111), for making a model act through long horizons. The design has real roots: instead of carrying the entire interaction history in context, the model compresses what has happened into a short natural-language belief state, and updates that state as new evidence arrives. The authors are explicit that the formulation is inspired by recursive Bayesian estimation — posterior, observation, updated posterior — and they isolate the summarization task instead of leaving it inside the model. The idea is that a belief state is a usable model of the world-so-far, and that grading it makes the loss visible.
The compaction gap
The honest part of the paper is what it measures. In a word-guessing game where a policy gets up to sixteen attempts, a policy that reasoned from a compressed summary improved with training but never closed the gap to a policy with full context. More data did not close it. The summary model just got better while staying behind. That is the failure mode: a summary is a lossy compression, and nothing in the loop was paying for the loss. The information that would have changed a later decision was dropped before the decision existed, and no score complained.
A cone is a belief state with a body
The forecast cone is a belief state of the same shape: everything the decision can still learn must pass through it, and anything that did not make it in is not available later. The difference is what the bottleneck carries. The cone is a distribution over the way the future forks, and its update rule is mechanical. When a signal arrives it enters the ledger as a node version; the ensemble re-normalizes the distribution in exactly one place; the revision is scored against a shadow timeline by deino-eval, in Brier and log-loss, gated by Gate 1b before any forecast-facing number is surfaced. Where ABBEL grades a summary with another model, the calibration loop grades the belief state with a number a risk officer can read.
A natural-language summary is a belief state whose loss nobody prices. A cone is a belief state whose loss is the product.
What a priced belief buys
- Sealed evidence: an update happened because a specific observation entered the ledger, not because the model re-read its own prompt.
- A mechanical rule: the re-normalization is one function, so “what changed” is a diff between two sealed states, not a retelling.
- A score on every revision: if an update degrades Brier on the shadow timeline, it ships as a regression, not as a better feeling.
The line we will not cross
We are pre-revenue, building the proof-of-concept; the layers are built but inert until calibration fills them. So the honest claim is narrower than it sounds: we do not claim the cone is correct, we claim it updates in a way that can be replayed, and that when it is wrong the record shows what it saw, when it saw it, and how it moved. That is the difference between a belief and a loss — a belief that can be replayed is the only kind you can take in front of an examiner without flinching.
A belief update is only honest when someone else can replay it: what you saw, when you saw it, and how the room moved. Otherwise it is just a new opinion.
— field notes — the oracle