Audiences AIRecoverable is a verdict
The downtime sheet that opens the morning production meeting reads like a settled matter. Equipment failure sits at the top of the loss column, large enough that nobody argues with it. In the documented case recorded in our Triage concept-validation work, it carried 70 per cent of all logged downtime. Budgets follow that column. The maintenance headcount and the capex case for a replacement press both get argued from what the sheet says.
Then automated capture went onto the same lines and the sheet met a measurement. In that documented case, automated capture run against the plant's own manual logs, equipment failure came to 18 per cent of lost time. Minor stops, the two-minute jams nobody walks to a terminal to record, measured 34 per cent. Speed loss took another 22 per cent, and it had never appeared on the sheet at all, because a line running slow does not look like downtime to anyone standing beside it. The category holding 70 per cent of the blame held less than a fifth of the hours. The concept document labels the case exactly as it should be read: a documented comparison of automated capture against manual logs, cited as directional context, no claim about any particular plant. The improvement budget, though, had been aimed with the left-hand column.
What the manual log actually records
The gap looks damning and is mostly innocent. An operator closes out a stoppage at the end of a shift, from memory, against a fixed code list. The machine is the easiest thing to blame and the hardest to cross-examine. A two-minute jam is beneath the effort of logging, so forty of them a week do not exist on paper. Running at 60 per cent of demonstrated rate still feels like production, so speed loss files itself under nothing at all. The sheet is a faithful record of what was easiest to name in the moment. Where the hours actually went is a different question, and the sheet was never designed to answer it.
This is why a diagnosis has to be computed from events and timestamps, with reason codes demoted to witnesses. In Triage's design, every code is treated as generic until the data corroborates it. When the codes agree with the computed picture, the agreement raises confidence. When they disagree, the computation wins.
The same code, twice
A corrected distribution feels like an answer. It still decides nothing, because the question that moves money sits one level deeper: of these losses, which is recoverable? Recoverability is invisible at the level of a single event. It lives in the context around the event.
An illustrative pair from the concept document shows the logic. The codes, durations and figures here are constructed on a generic plant to demonstrate the reasoning. No client sits behind them. Two idle windows on the same machine. Same code, NO MATERIAL. Same duration, three hours. In the first window the order book was loaded and work was due to run; the idle sits against a late receipt from a vendor with a record of late receipts. That is a readiness failure, and it is recoverable. The material path can be fixed, starting with the second source the plant already buys the same material from. In the second window the order book was empty and demand was soft. The missing material is a consequence of that, and the machine is correctly idle. Fixing it would mean making stock nobody ordered.
Same code, same three hours, opposite verdicts, and only the computed picture separates them. This is also why some loss categories are deliberately never routed to a fix: demand-driven loss, structural loss from a high-mix catalogue, capacity questions that belong in a portfolio review, not in a floor project. A system that routes everything to a solution has stopped judging.
What the verdict has to carry
A verdict that arrives unpriced loses to whichever project has the loudest sponsor, so each verdict has to carry its own economics. The concept document's illustrative recommendation card, constructed the same way, shows the shape. Avoidable changeover on one line: 41 A-to-B transitions in a month, against a demonstrated baseline of 34 minutes for that transition, while the line averaged 95. Operators had coded 38 of the 41 as SETUP, which agrees with the computation, so the codes get their corroborating vote. The card sizes the loss at about 68 machine-hours a month, and it prices in hours because no margin data was supplied. A currency figure would have been an invention, and one invented number is enough to cost a loss map its audience. The sizing baseline is what that line has already demonstrated; industry averages stay out of it. The card prints its confidence, 0.81, and waits for a human to accept or reject it.
One verdict sits underneath all of these: sometimes the honest output is no diagnosis. In Triage's design, when less than 40 per cent of coded downtime carries a non-generic reason, the system refuses to diagnose and ships a data-gap report with a logging prescription: capture these fields, resubmit, get re-scored. That refusal is the product working as designed, not a failure state. The cycle then re-runs monthly, because distributions drift. Changeovers creep and bottlenecks migrate, and last quarter's verdict slowly turns into a historical document.
Between the monitors that measure and the suites that act sits this judgment, and in most plants it is an analyst's spreadsheet or a consultant's one-time engagement, stale by the next planning cycle. We built Triage to hold the seat permanently. It reads the production and cost data a plant already produces and turns the loss column into verdicts, one card at a time, each waiting for a human to accept or reject it. Nothing runs until someone does.
Triage is that judgment layer, run as a monthly product inside your own network.