Only records from the station-authenticated hardware ledger appear here. No simulator, fixture or agent output is ever listed in this view. A record becomes VERIFIED_PHYSICAL only after a ground-truth check is recorded for it; until then it stays PHYSICAL_CAPTURE.
| Observed | Event | Specimen | Station | Kind | Value | Tier | Ground truth | Station auth |
|---|
No physical record yet. This view stays empty rather than showing simulated data. Stations post signed readings to
/hardware/caliper-reading and /hardware/bin-ack (see docs/engineering/PHYSICAL_TEST_PLAYBOOK.md).
| Action | Transition | Specimen | Bin | Sensor event |
|---|
No route actions. The sorter's default position is HOLD.
Self-declared events from /sim/live/event. Display only: not signed, not in the ledger, never acted on.
Widths come from a mid-length cross-section of each STL, not from file names. The fit check is a design check, not a physical measurement.
HyorinmaruK's manual-track demonstration: printed specimens, caliper and cameras, a human-confirmed servo diverter. These pages are his tools, served unchanged. Measurements exported from the logger into data/measurements/ calibrate the simulator below; the caliper stays authoritative.
Safety rules: AGENTS.md. No model or browser script moves the servo; every route needs a physical button press.
| Specimen | Design | Readings | Mean mm | Nominal | Deviation | Spread |
|---|
| Specimen | Photo file | Photo mm | Caliper mm | Photo − caliper | px/mm |
|---|
Product means calibrated Bayes with risk-optimised pairing; today means the fixed rule with FIFO. Verdicts use the pre-registered rules (95% cluster bootstrap). ROI is a scenario outcome under illustrative prices, never a saving.
| Scenario | Δ good / h | Δ escapes / 1000 | Δ cost / good unit | Safety | Throughput | Cost | P(product worse) |
|---|
The Quality Inspector starts with phone and CAD evidence only. It asks peers for evidence it cannot see, every reply is checked (signature, source_root, freshness) before use, and SAFE-MESH makes the decision. Evidence is TEST_FIXTURE; likelihoods are illustrative, not calibrated. A control run with requests disabled is shown beside it.
Mock agents with hidden truth. Each policy pays for every observation and every message. The primary result is negative: at default prices, the full SAFE-MESH + VOI agent holds every part.
| Policy | Unsafe PASS / 1000 | Released good | Cost / unit |
|---|
| Scenario | Rule | Calibrated | Δ [95% CI] |
|---|
Unsafe PASS per 1000 parts. Warp kernels make decisions identical to the CPU code, part for part (tested). Run on Warp's CPU backend; the GPU run is pending.
NOT_CONFIGURED.Write-ups: docs/research/SIM_18_1_RESULTS.md · COLLECTIVE_GATE_A_RESULTS_31OCT.md · WARP_MC_AND_TUNING_31OCT.md · 17.0 page