CALPHAD databases now ship with their decisions: ORCHID-SAGE record lets a Cr–Ni assessment be rebuilt on a second machine
Synopsis
Building a phase-diagram database has long depended on expert judgment that is rarely recorded, so published CALPHAD assessments are hard to regenerate or update as new data arrive. ORCHID-SAGE splits an assessment into a general protocol, a system-specific protocol and run cards for each execution, and distributes them with the database. For Cr–Ni, that record alone let a script on a second machine and an AI agent working only through an MCP server regenerate the database, with predictions inside the tolerances declared in advance.
Interpretation
A phase-diagram database can now be published together with the record of the judgments that built it, not only the fitted parameters. Published CALPHAD assessments usually report final parameters alone; which observations were used, what uncertainty each carried, how alternative models were chosen and which code version ran are largely lost once the assessment is complete. The Cr–Ni database was regenerated by a script on a machine differing in processor, operating system and numerical libraries; five deterministic checks of the 13 gave identical results, eight numerical checks agreed within tolerance, the objective differed slightly, and the eutectic and CrNi2-top temperatures showed no difference.
The record is organized into three machine-readable levels, so a rerun needs no contact with the original authors. A general protocol states rules for any system, a specific protocol records this system's evidence, departures and each decision with its responsible party, and run cards hash each step's inputs, code versions and outputs. An AI agent with no prior information beyond the specific protocol identifier executed it step by step through the ORCHID-SAGE server, finishing in 46 turns and 2 h 38 min at US$1.64 in model usage, and passed all 13 checks, eleven with no difference.
What the evidence cannot determine is reported as undetermined rather than printed as a settled value. A conventional assessment publishes one parameter set, and parameters along flat directions of the objective can shift widely without changing the phase diagram, so this uncertainty usually goes unnoticed. The final Cr–Ni fit showed two modes differing by only 0.0045 log units in objective but by up to 5.9 times the parameter tolerance in the first-order Redlich–Kister terms of bcc and fcc, while giving identical phase diagrams, invariant temperatures and thermochemical properties.
Perspective
The demonstrated scope is the Cr–Ni binary system, with regeneration starting from PhaseForge one-shot energetics; machine-learned-potential calculations were checked only for a small sample of structures on a second GPU, and a full regeneration of those calculations is possible but far more costly than the rest of the assessment.
Some data were extracted by a single reader and are not yet in a released ORCHID version, so they stay flagged until a second reading confirms them; the corrected estimator used for the liquid's excess entropy has not been independently reviewed and is treated as provisional throughout.