Foresight — the Business World Model

The model that knows what happens next.

Foresight is a business outcomes predictor. It holds your company’s past — every decision, what it predicted, and what actually happened — and turns that into what happens next: outcomes as ranges, with the drivers behind them and a version you can hold it to. It learns your loop inside your own boundary, and nobody else’s model learns from you.

Problem

Your company’s past keeps getting thrown away.

Decisions get made, outcomes arrive, and nothing connects the two. Six months later nobody can say what the pricing change actually did, which channel really worked, or how long that launch truly took — so the next decision gets made the way the last one was: from memory, and confidently.

Brainis records the pair every time: the state, the decision, the prediction it carried, and the outcome that landed. That record is what makes prediction possible at all — and Foresight is what the record becomes.

Mechanism

What a prediction is made of.

Four layers we own turn your company’s record into an outcome: what an action will do, which decisions tend to work, how far the number can be trusted, and what the futures look like before you commit. Take a layer to read what it does.

Ours

Prediction

Name an action and it tells you what that action does to your company — as a range with the drivers behind it, never a bare sentence. How likely this mission is to land, how likely this deal closes, how far the timeline slips, where the cash goes.

These four layers are what predicts your outcomes. In your workspace they are tuned on your company’s own loop, and they stay inside it.

Yours

Your company gets its own model.

One model per company, tuned on that company’s own loop. The longer you run it, the better it gets at your business — and at nobody else’s.

  • Tuned on your loop

    Your decisions, your outcomes, your calibration curve. The model adapts to how your company actually behaves — which is a different and far more useful thing than getting better at companies in general.

  • Inside your boundary

    That adaptation lives inside your own tenant boundary and is deleted with your organization. It never leaves — so there is nothing to leak, and nothing to hand to anybody else.

  • Never sold, never borrowed

    We do not sell customer data. API and enterprise workspaces are never part of shared training at any tier, and contributing to the shared model is an explicit choice — structure, never your content.

Anatomy

The prediction record.

Foresight never returns a bare number. Every prediction it makes is an object you can open:

Prediction record

objective
ten design partners by day 60
action
focused design-partner program
predicted range
8–14 qualified conversations
confidence
0.68 — calibrated to mean it
model_version
foresight-1
drivers
the features that moved it — an unexplainable prediction doesn't ship
calibration note
how much the range narrows as your outcomes land
horizon
day 60 — the date the outcome is set against it

Registered the moment it’s made — range, drivers and model version sealed before the answer is known, so what it claimed can still be read back once the outcome is in.

Worked example

Aster’s D-104, sealed before the answer.

Decision D-104 · qualified conversations

predicted range · sealed at decision time

Aster, week two. The model predicted 8–14 qualified conversations for the design-partner path, and shipped that prediction with its drivers and its version attached — because a number nobody can interrogate is not evidence.

The Decision Ledger froze it the moment the decision was sealed. Not summarised, not revised later when the answer was known: preserved.

Day 60: eleven. Inside the range — and readable as such only because the range was written down in week two and could not be edited afterwards.

Scoring a prediction against its outcome is a human step today. Automatic grading, and the calibration record it would feed, are not running yet.

Control

Confidence you can check.

Every prediction carries its range, its drivers and the version of the model that produced it, written down before the outcome exists. That is what makes a confidence claim checkable at all — the record is fixed, so it cannot be quietly rewritten to match how things turned out.

UnderstandDecideDeliverVerifyLearn

Company State feeds the model; the model's prediction rides the decision into the ledger, and stays there to be read against the outcome.

Hard question

Do you actually train models, or just prompt someone else's?

We train the layers that do the predicting — and on this question those are the ones that matter. Reading your company’s state, predicting what an action will do to it, judging which shapes of decision actually work out, setting how far that prediction can be trusted: all of that is ours, and it is trained on the one thing nobody can scrape — real sequences of state, decision and outcome, with the outcome attached. The text layer underneath drafts and summarises — a general-purpose language model, called through provider adapters — and it is not where the prediction comes from.

What data trains it?

Your model trains on your company, inside your boundary, and is deleted with your org. For the shared model there is exactly one door: a structure-only export — the shape of a decision, its evidence classes, its prediction, its outcome — with identifiers stripped and an anonymity gate in front of it, and only from organizations whose consent tier allows it. API and enterprise workspaces are never in shared training at any tier, and we do not sell customer data. Your documents, your messages and your customers' records are not training material.

Why should I believe the numbers?

Because they were made checkable before you asked. Every prediction is registered the moment it's made, with its range, its drivers and its version, and that record is sealed — it cannot be revised once the answer is known. Being checkable is not the same as having been checked: automatic grading of predictions against outcomes is not running yet, so we publish no accuracy figure and you should not infer one. When it does run, the record that grades a prediction will be written by the same independent Verify that grades the work, never by the model being graded, and BusinessBench publishes the method we measure with so it can be argued with.