Every factory recipe has steps nobody can explain anymore. In Primo Levi's "Chromium" (The Periodic Table, 1975), a paint plant kept adding ammonium chloride to every batch twenty years after the problem it fixed had vanished: a fossil. Other steps look pointless but are load-bearing, and changing them without a full review can be deadly. In 2013 at the Williams olefins plant in Geismar, Louisiana, a poorly reviewed change let a reboiler be cut off from its relief valve; it ruptured and killed two workers (US Chemical Safety Board, 2016). Now AI optimisers propose "remove this, shorten that". CROMO is an investigator for process engineers, quality/EHS and plant managers. Pick any step and it digs through decades of the plant's own records (scanned change requests, faxes, incident reports, handwritten logbooks, supplier records) to find why the step exists and whether that reason still holds. Each verdict is FOSSIL, FENCE, ACTIVE or UNKNOWN, with an evidence trail, a strength grade and quotes verified verbatim in the source. Models read the evidence; auditable rules decide: fences are never removed, fossils only through a controlled trial, unknowns go to a person. It also flags steps affected by supplier or equipment changes (Tripwires), checks proposed changes before they happen, prices each fossil, and sends trial plans for engineer approval through an Evolus workflow. All model calls run on the AMD Token Factory with Google's Gemma 4 31B-IT, which reads scans and handwriting (2.2% character error rate) and does the reasoning. On a synthetic 178-document archive with planted answers, CROMO reached 89% fence recall and a 0% false-fossil rate (no-LLM baseline: 56% and 11%) and found €305k/yr of fossil steps. A full audit cost $0.04 in metered tokens. MIT licensed.
Category tags: