This repository is the public validation record for a free screening tool for metal laser powder bed fusion (LPBF): amprintanalyzer.com. The tool reads part geometry and reports build risk — distortion and recoater-strike zones, trapped powder, sub-process-limit features — with every claim labeled by its evidence class (measured / screening / estimate).
A screening tool asks to be trusted. This repository is how we earn that instead: predictions are committed before measurements are read, misses are published alongside hits, and every number here traces to a public data source you can check.
| Receipt | Artifact | Ground truth | Outcome (short) |
|---|---|---|---|
| ambench-2018 | NIST AM-Bench 2018 bridge (IN625) | NIST CMM deflection + neutron/synchrotron strain | Blind screen agreed with the build facts at the process edge; deflection prediction correctly refused (out of screening scope) |
| ambench-2022 | NIST AM-Bench 2022 bridge (IN718) | NIST CMM ridge deflections (±4 µm) + 2,248-point strain field | Blind screen: buildability, thin features to the third decimal, stress band localized — and one documented miss-then-catch (sealed powder cavity vs. scan-grid resolution) |
| easi-arch | EASI-STRESS benchmark arch (316L, two machines) | Neutron + synchrotron residual stress, four instruments, three facilities | Localization comparison (not blind, labeled so): flagged band sits directly under the measured near-yield ligament stress; includes one retracted claim, in writing |
| tpms-factorial | 81 printed IN718 TPMS lattice specimens (published study) | Measured compression properties per specimen | Pre-registered stiffness exponent confirmed (1.2 vs 1.22–1.27 fitted); constants off in known directions — documented as calibration candidates, deliberately not silently applied |
See PROTOCOL.md. The short version: for blind runs, the screening report is committed to version control before the measured results are downloaded, so the order is auditable; the screen is deterministic code with nothing to tune between download and report; disclosures state what the operator knew and when. Where a run is not blind (e.g. the arch, where measurement files were opened first to establish coordinates), the receipt says so in its first lines.
The receipts dated on or before 2026-08-11 were executed in a private
working repository; their blind-commit hashes and timestamps are quoted
in each writeup (e.g. 3129539 for the 2018 bridge, c0ddce8 for the
2022 bridge) and we will show the underlying history to any serious
auditor on request. From this repository's creation forward, blind
runs commit here first — the public chronology becomes the proof, not
an attestation of it.
Our writeups, figures, and reconstruction files: CC-BY-4.0. Third-party data retains its own terms, credited in place: NIST AM-Bench files are US-government public domain (17 USC §105); EASI-STRESS data is CC-BY-4.0 (Zenodo, EU project 953219); the TPMS study is CC-BY-4.0 (Materials 14:3962). The analyzer itself is a hosted tool; this repository contains no analyzer source code.
That's the point. If your measurements contradict a screening claim — here or on your own parts — tell us through the tool's feedback strip or open an issue here. Disagreements are the calibration data.
Full repository: raw reports, data files, protocol, commit history →