This is a working note, not a paper. We'll update it as the product evolves. If you're an engineer evaluating Foreman, this should give you enough to push on the parts that matter most for your plant.
The reasoning loop
When an anomaly fires in the digital twin (a line OEE drop, a cycle-time outlier, a downtime event), Foreman's agent enters a four-step loop: gather, hypothesize, rank, recommend. The loop terminates with a written recommendation that lands in an engineer's queue with a confidence score and an explicit citation of every signal that informed it.
Gather
The agent pulls a context window scoped to the anomaly: the affected line, upstream and downstream assets, the last 14 days of MES events, the historian's tag values for the relevant station, any maintenance tickets that touched the assets, and the SOPs versioned against the current shift's production model.
Crucially, the context window is not 'everything.' We have hard caps on how much data goes into the model, both for cost and for reasoning fidelity. A bigger context is not a better context. It is a noisier context. We rank candidate signals before they enter the window.
Hypothesize
The agent generates between two and five candidate root-cause hypotheses. We deliberately do not constrain it to a single answer. A confident-wrong answer is the worst failure mode for industrial AI; a ranked list of plausible answers, each with its own evidence, is something engineers can interrogate.
Rank
- Each hypothesis gets a confidence score based on signal strength and historical pattern match.
- Hypotheses that would trigger destructive corrective actions (e.g., line stop, asset shutdown) are scored more conservatively. The bar to recommend them is higher.
- Hypotheses that match a documented SOP path are surfaced with the SOP reference inline.
Recommend
The recommendation lands in the on-shift engineer's queue. It includes: the action, the SOP reference (if any), the confidence score, the alternative hypotheses, and a one-paragraph plain-English explanation. The engineer approves, rejects, or modifies, and every step of that is logged for audit. There is no silent action.