Causal model
A directed acyclic graph that names the firm's assumptions about cause and effect, attaches a measurable prediction to each, and sharpens those assumptions against what actually happened.
The causal model is firmd's representation of cause-and-effect inside a tenant's product domain. Nodes are concepts (a feature, a behaviour, an outcome). Edges are claimed causal relationships, drawn out during deliberation.
The direction is Pearl-style causal reasoning. DoWhy is the practical reference point for making interventional claims explicit: if we change X, will Y move? — not just does X correlate with Y?
A product organisation that does not name its bets cannot tell whether it is learning. It accumulates features without accumulating judgment. Generative agents reproduce this pathology at higher speed — coherent outputs, no world-sense.
The causal model is the firm's antidote. Every Strategic Intent commits to a falsifiable bet. Shipped work measures Prediction Error. The graph updates. Over time the firm gets better at making bets — not by adding heuristics, but by being repeatedly wrong in a way it can see.
The graph is not authored offline. T-shaped agents assert edges in their domain — an Architect names a tech-debt-to-latency edge, a Go-to-Market agent names a latency-to-win-rate edge — and discourse stitches them together.
Every Strategic Intent attaches a numeric prediction to a node: "first-week activation should rise by 25%." The prediction is what makes the bet falsifiable.
After delivery, telemetry measures the actual movement on the predicted node. The delta is Prediction Error. The graph updates: which edges held up, which did not, which were never tested.
The graph evolves with every cycle. Older versions remain queryable so an audit can ask: what did the firm believe on a given date, and how did that belief survive the next month of shipping?