Copper +1.6σ → 3 projects exposed → P80 +$41M
rule XM-03 v4 · copper, 5 sessions
Draft: re-phase the substation's cable order by two months and hold an $18M reserve.
engine · 4,000 scenarios P80 exposure +$41M → +$12M if applied · confidence 0.81
The brain · continuous monitoring
The brain is the monitoring layer of Capibud’s decision model. On a schedule, it reads what your project systems record and what market and supplier feeds report. When something moves, it finds every exposed project or fund position, prices the effect in money and drafts a re-allocation, which the engine verifies in a sandbox. Then it puts the proposal in the right person’s queue. Nothing changes the plan until that person approves it.
Status: partial. The scheduled pipeline from signal to approval queue runs today. Learning from outcomes needs months of real decisions, so that part is still ahead.
One fixed sequence
Every pass follows the same steps in the same order, so a proposal always arrives with the same things attached: what moved, who is exposed, how much money is at stake, what the engine found when it checked, and how confident the brain is.
A market move, a company filing or an update from one of your project systems lands on a scheduled pass, with its source and date.
Capibud identifies every project and holding that loads on that factor, across all four decisions.
The effect is priced in money, at the P50 and in the adverse case, not as a red, amber or green light.
A re-allocation is drafted: re-phase, hedge, hold a reserve, slow the commitment pace or reopen a decision.
The engine re-runs the proposal in a sandbox. If its result contradicts the proposal’s claim, the proposal is flagged rather than forwarded.
The proposal reaches the right person’s queue through your delegation of authority, with its evidence and a calibrated confidence.
Nothing changes the plan until a person approves it. The approved change is previewed and applied, can be exported to the systems that run the work, and can be undone.
The proposal, the decision and the reason are written to the append-only record, and the reason feeds the brain’s calibration.
Watch it work
An illustrative copper move travels through one cable supplier to three projects. It is priced, drafted, verified by the engine and routed, and then it waits for a person. On the left, why; on the right, what lands on the CFO’s desk. Select any point on the graph for the detail.
Illustrative example. The figures are not taken from a client project.
What it watches
A price move matters only if money in your book loads on it. The brain reads the updates your schedule, cost and ERP systems record next to market and supplier feeds, keeps a map of which projects and fund positions load on which shared factor, and checks every source against a freshness SLA before it can move a number.
Schedule exports (XER, XML), cost ledgers and ERP extracts update the P80 cost and finish of each project, so drift shows up as money at stake rather than a late surprise.
Commodity, rate and currency series on a schedule. A versioned rule fires when, for example, a commodity moves 1.5 standard deviations, or a rate surprise moves every toll road, network and wind farm in a fund at once.
10-K, 10-Q and 8-K filings and XBRL company facts, compared with the previous filing and passed to the brain when a figure changes.
A contractor’s listed peer falling 25% is a known warning pattern. Counterparty health and supplier concentration are scored, so one supplier carrying three projects is visible.
The thresholds a committee set when it decided are checked on every pass. A breach, or a close approach, becomes a proposal to revisit the decision.
Tracked competitor capacity is compared with the market balance the engine assumes, and weather and logistics windows are checked against the work still to do.
How it reasons
Rules catch what you already know to watch for. Fitted models estimate whether a move will matter. Language models, if you switch them on, read text and explain. None of them is allowed to produce the numbers on its own.
Thresholds that catch known patterns quickly, such as a commodity moving 1.5 standard deviations or a contractor’s listed peer falling 25%. Each rule carries an identifier and a version, so every proposal can say which rule fired.
Has something we already know to watch for just happened?
Reference-class overrun priors, signal-to-probability calibration and factor correlations, fitted and then frozen. The signal-persistence model has an out-of-sample Brier score of 0.172, where a coin flip scores 0.25.
Will this move last long enough to change the plan?
Read unstructured text and draft explanations. Every number in AI-written text must match the recorded run exactly, or the text is rejected and the deterministic version is used. The AI layer is off unless configured.
What does this filing or memo actually say?
A morning’s queue
The brain does not raise one alert per signal. It ranks proposals by money at stake, calibrated confidence and urgency, sends small items to an FYI list, and holds anything built on a stale or invalid feed. Open a proposal and the engine’s re-run is already there, including when it disagrees.
Independent sources reporting the same factor lift the confidence of a proposal. One noisy headline does not.
When the engine’s own re-run contradicts a proposal’s claim, the proposal is flagged and ranked down instead of forwarded.
A signal whose source feed is stale, invalid or missing is held. It never becomes a proposal.
Illustrative queue. Signals and figures are invented for the example and are not client data.
Guardrails
The brain is there to make sure nothing material goes unnoticed. It is not there to change your plan on its own.
Every proposal is routed by your delegation-of-authority matrix, previewed before it is applied, and can be undone after.
Each actionable proposal is re-run in a sandbox copy before it reaches anyone. The engine’s figure replaces the detector’s estimate.
If a language model drafts the explanation, every number must match the recorded run exactly, or the deterministic text is used instead.
Learning proposes new model versions. A person approves them before they become the default.
Operators can stop any automated job. A held job completes without changing anything.
When someone declines a proposal they say why: a false alarm, sized too high or too low, already handled, or the wrong time. Those reasons feed the confidence of the next one.
Where it stands
We label the brain Partial because part of its value depends on time.
See it on your own exposures. In a Decision Sprint the brain watches the factors your decision depends on, and the committee sets the tripwires it will check. How the committee sets them
Questions teams ask
Every proposal it drafts lands in the decision record, and tripwires come from the capital committee.
No. It drafts a proposal, verifies it in a sandbox and routes it to the right person's queue through your delegation of authority. Nothing changes the plan until that person approves it; the approved change is previewed, applied and can be undone.
The updates your schedule, cost and ERP systems record, market prices, company filings (10-K, 10-Q and 8-K filings and XBRL company facts), contractors and counterparties, committee tripwires, and capacity and calendars. Each is mapped to the projects and holdings exposed to that factor.
Yes. A central-bank rate surprise, a power-auction result or a change to a support scheme is stored with its date, source and measured move. The brain marks every position that loads on it, including the assets still in construction inside them, re-prices the fund on the same scenarios and drafts a proposal, such as slower commitment pacing, a hedge or a sell-down, for the right approver.
No. Versioned rules and fitted, frozen models flag and estimate; the simulation engine produces every number. Language models are optional and off by default, and any number they write must match the recorded run exactly or the text is rejected.
Partly. The scheduled pipeline from signal to approval queue runs today. Learning from outcomes needs months of real decisions, so that part is still ahead.
A Decision Sprint runs one real decision on your own files in six weeks, with the brain watching the factors that decision depends on.