Day 1 — WITSML lands.
The pilot agreement is one Wolfcamp lateral; the WITSML store from the existing well-management system is the source of truth.
On day one the agent connects to the WITSML store and transfer the operator already runs. Surface and downhole pressure, flow rate, ESP load, SIBR — every channel sampled at the cadence the store already returns them in. The first cohere vector the agent builds binds each of those curves to a single time index so the detection sweep on day five has a stable clock to run against. No re-instrumenting the well, no polling the historian at a different cadence — the WITSML stream is treated as the truth and the agent reads from it.
Day 2 to 3 — the threshold envelopes tune.
Per-curve rolling baselines built from the offset-well window, so the envelopes are tighter than the SCADA alarm and looser than gut feel.
The agent watches the first forty-eight hours to establish a per-curve rolling baseline. Casing pressure on the A-annulus gets its own envelope, built off the offset-well baseline window the operator has on file. The basin noise floor on casing pressure sits around forty psi on this pad; the super-major's existing SCADA alarm fires at two hundred and fifty psi. The envelope we tune toward is somewhere in between — tighter than the SCADA trip, looser than the operator's gut feel. Real envelopes, not magic numbers. Each curve gets the same treatment.
Day 5 — the first anomaly.
Casing pressure on the A-annulus trends eighty psi above a ninety-minute rolling baseline — well below the SCADA alarm, but a consistent monotonic drift.
Five days in, the casing-pressure channel on the A-annulus trends eighty psi above its rolling baseline across ninety minutes of telemetry. The signal is well below the two hundred and fifty psi SCADA alarm — the operator's existing trip still has margin — but the agent flags it anyway because the drift is monotonically rising, not the spike-and-relax pattern the basin noise floor produces. The signature matches a slow pack-off or a packing-element seating issue, not a barrier breach. The agent writes the “pull off bottom and circulate two bottoms up” imperative copy automatically and pushes the alert with the raw WITSML trace window attached, no human in the loop on the trip.
Week 2 — the alert triage review.
A standing thirty-minute review with the night foreman walks the false-positive budget and re-tunes the threshold envelopes by tens of psi — the operator signs off on every change.
At the end of week two the running review is a thirty-minute sit-down with the night foreman. We walk the false-positive budget: which alerts got suppressed, which got escalated, and which envelopes need re-tuning. The night foreman reads the raw cohere vector the agent attached and the raw WITSML trace window behind every alert. Threshold envelopes move by tens of psi, not hundreds — and the operator signs off on each change before the new envelope ships. The point of the review is to keep the agent honest about the basin noise floor, not to tune it into a numb system.
Week 4 — the first weekly PDF.
The fleet's first auto-written weekly reservoir report lands in the operations lead's inbox as a one-page PDF, with the raw WITSML trace windows attached.
Week four is the first weekly hand-off. The operations lead opens the inbox on Monday morning and the fleet has emitted a one-page PDF summary of the well's last seven days — production rate trend, water-cut delta, casing pressure, and the AI-written commentary the agent authored from the trace windows. Every flagged event from the period is triaged into a single screen-ready read-out so the engineering team walks into the Monday standup with the same brief the night foreman has already seen. Detection to briefed queue, in seconds instead of weeks.