← analyzer · receipts index

Blind screen vs. NIST's measured bridge, generation two — AM-Bench 2022 (AMB2022-01)

Run 2026-08-10 · blind protocol · files here: NIST's STL + its published sha256, the unedited blind report JSON, the measured-deflection table as machine-readable JSON, the springback figure.

The artifact

The 2022 successor to the 2018 bridge: IN718, 75 × 5 × 12.5 mm, twelve 45°-overhang legs — with a twist. One leg (L10) is secretly hollow: 0.5 mm outer walls, single- and double-track-width internal walls, sealed at the top with powder inside forever. Built on NIST's AMMT testbed; ridge tops skim-ground; CMM-measured before and after all twelve legs were EDM-cut, so residual stress lifts the part.

Protocol

Geometry downloaded and sha256-verified; screen run at the closest windows (IN718 → inconel, AMMT → eos_m290 envelope); raw report committed before the measurement files were fetched — commit c0ddce8, 2026-08-10 (see the README's provenance note). Disclosure, stronger than usual: the reconnaissance that selected this target had already surfaced NIST's headline numbers, so the operator was not ignorant of them. The blindness claim attaches entirely to the deterministic code and the auditable commit order.

What the blind screen said vs. what NIST measured

Blind screen NIST's measurement / build fact Verdict
overhang pass — 0.39% of surface needs support at the 45° limit, risk faces at the span underside The part builds support-free by design, and did Agrees at the process edge
thin_wall flag — min sampled thickness 0.109 mm, p05 0.499 mm L10's internal walls are deliberately single/double track width (~0.1/0.2 mm); the outer walls are 0.5 mm by design Both designed-in thin features found to the third decimal
stress_concentrators watch — 338 mm of severe concave edges across the leg-junction band (z 1–12 mm) On EDM release the part lifts monotonically toward the free end: −0.003 → 1.809 mm Localizes where the measured springback releases from
Build-time warp screens: pass The build completed; deflection appears only after the cut Agrees, boundary stated
trapped_powder at the default 0.62 mm scan grid: "no enclosed cavities resolved at this pitch" L10 is a sealed, powder-filled cavity A miss — honestly labeled by the tool itself (the sealing walls are thinner than the grid), see below

The miss, and the catch

The sealed cavity leaked through the default voxel screen because its 0.5 mm walls are thinner than the 0.62 mm scan grid — and the report's own language anticipated exactly this ("passages narrower than the grid are invisible here"). The tool's built-in refinement pass, which re-screens doubtful volumetric checks at twice the resolution, caught it immediately: enclosed cavity ≈ 0.03 cm³, no drainage path — exactly where NIST sealed it. That verification now runs as standard: cavity findings are checked stable under grid refinement before they're claimed.

We publish the miss at the same prominence as the hits because a screening tool that states its resolution limits can be calibrated — and one that hides them cannot.

The boundary

Same as 2018: screening localizes, it does not predict millimeters. The measured deflection table in this folder is our standing validation target for the physics tier under development — published here so that when that tier makes claims, the exam it must pass is already public. NIST's 2022 challenge drew 138 blind modeling submissions from 19 groups (2018: 46, none meeting criteria) — the field is improving, and honest screening still beats dishonest simulation.

Figure

comparison figureAudit this receipt on GitHub → raw reports, data files, commit history