A mill's genealogy is not a compliance chore — it is the same record that computes cost per tonne, issues the test certificate, proves melted-and-poured origin and declares embedded emissions. SteelOS Mill keeps all of it on one graph.
At the bottom, automation vendors own the furnace and the caster. At the top, an ERP owns the ledger. In between is everything that makes it a mill.
Primetals, SMS, Danieli, TMEIC. They regulate furnace power and caster speed. We consume their signals and never assume their job — closing a control loop on a 1600 °C furnace is not our liability to take.
Material genealogy, heat logs, quality certificates, rolling campaigns, yield, metal balance, process cost per tonne. Filled — when it is filled at all — by a separate MES bought and integrated at enormous cost.
SAP, Oracle, Ramco — or, far more often, Tally and a patchwork of spreadsheets. It knows what you invoiced. It does not know which heat it came from.
The genealogy that proves a rebar bundle came from heat H-4471 is the same genealogy that computes its cost per tonne, generates its EN 10204 certificate, establishes its melted-and-poured origin, and supplies its embedded emissions for a CBAM declaration. There is no reason for that to live in two systems.
So there is one MaterialLot node and one GenealogyEdge DAG, and routings are data rather than code. Adding a mill archetype — induction, electric-arc, a galvanizing line — means authoring routings, not shipping a fork.
The discipline that makes it hold: edge fractions carry cost and metal balance; measured product analysis carries grade and certification. The two are never conflated. An estimated blend fraction may not masquerade as certified chemistry, and a transition billet from two heats is born HELD.
WRITTEN ONCE. READ FIVE WAYS.
Beachhead is the induction-furnace and re-rolling chain: scrap and sponge iron in, TMT rebar out. Electric-arc mini-mills, flat products and coating lines follow — on the same core, by configuration.
The heat → cast → roll → bundle lineage, with conformance checked against the grade window before you tap.
An induction furnace does not refine — whatever you charge is roughly what you tap. Least-cost blend, hard residual limits.
Mill test certificates generated from genealogy. A 3.1 backed only by a ladle analysis is refused.
A mill has no price — it has a base price and an extras matrix. Grade, size, width, packing, coating, certificate type.
The coffin schedule, melt and cast plans, finite-capacity scheduling and capable-to-promise.
Index-linked pricing, assay-based settlement, and a radiation check that blocks receipt rather than warning about it.
Cost per tonne per stage, standard-versus-actual variance, and a metal balance that reconciles inventory to the ledger.
Electricity is 20–40% of conversion cost. Per heat, per grade, against time-of-day tariff, with peak-load alerts.
Embedded emissions per tonne traceable to metered energy and charge mix. E-invoice IRN, e-way bill, BIS licence validity.
Everything below runs from the grade master and the purchase ledger — data a mill has before it has us. Capabilities that genuinely need years of your own process data are not sold as though they don't.
The cheapest blend of scrap, sponge and ferroalloys that lands inside the grade window. A bounded mixed-integer program does the arithmetic — the model only explains the recipe and flags tramp-element risk. The operator confirms before charging.
A supplier's MTC or assay, as a PDF or a photograph, becomes structured chemistry attached to a lot — with a confidence score per field. Low confidence goes to a human, because a misread corrupts the graph permanently.
Does this heat meet the grade it was ordered for? Deterministic rules give the verdict and cannot be overridden by the model. If it fails, the model ranks the best alternate grade against your open order book.
“Which open heats are off-spec on manganese?” — answered against tenant-scoped data, in words, without anyone building a report.
There is no product to sign into today. The API answers exactly one endpoint — a health check — and it says so in its own response body. Deploying it proved the pipeline, not a product.
What does exist is the part that is expensive to retrofit: multi-tenancy that fails closed, a gap-free audited number series, a general ledger whose every entry must balance, a metal-balance engine on dry-metallic equivalent, and a radiation check that blocks a receipt instead of logging it.
The first release you could run a mill on is M4 — melt, cast, roll, quality and certification. Early-access partners shape what lands in it.
We are choosing a small number of induction and electric-arc mills to build M3 and M4 against. If your chemistry, your grades and your paperwork are in the room, they end up in the product.