Fully decentralized by design · No factory deployment claimed
Assured Scientific Factory

Verified evidence.
Safer decisions.

Dual Vision, CAD geometry, verified metrology, 21 specialist roles, digital factory simulation, cost optimization, and human-governed physical sorting.

RESEARCH REFERENCE — no trained NVIDIA weights, live Nebius account, real camera data or factory-certified safety metrics are included.
Implementation snapshot
Specialist roles21
NVIDIA integration slots8
Vision stations2 planned
Scientific modeSeeded

Role count does not imply 21 separately trained or running LLMs.

Manufacturing evidence pipeline

Digital Factory6-stage flow-shop reference
→
Independent SensorsPhone / Pi / CAD / Caliper
→
SAFE-MESHLineage + HOLD + active inspection
→
Human ReviewPASS_REVIEW ≠ PASS
→
Sorter ACKOnly if physical sensor confirms

What is implemented?

HMAC evidence / ACL / event ledgerTESTED
Bayesian VOI & matchingTESTED
Factory flow shop simulationTESTED
Physical Pi camera + NVIDIA weightsNOT TESTED
Real servo sortingNOT CONNECTED

Decision policy

Example: phone clear, Pi suspect, metrology verified
HOLD

Disagreement requires human reinspection. Neither LLM majority voting nor synthetic CAD geometry can authorize a real release.

Local photo-to-reference comparator

Use your own images. They are processed **locally in this browser** by an intentionally simple pixel-difference baseline. This is NOT a calibrated defect model or NVIDIA inference.

Upload reference and live first; same angle, light and scale required.
Reference
Current
Absolute pixel-difference heatmap

A numerical difference is sensitive to lighting and pose. CAD files should first be rendered to a matched projection; a mere shape mask cannot certify 0.1 mm tolerances.

Try the user's original CAD sample

STL slot top 2D projection

Slot nominal STL top orthographic silhouette (2D raster)

STL tab top 2D projection

Tab nominal STL top orthographic silhouette

Seeded Digital Factory

4%5%

Requires `uvicorn qualityrelay17.api:app` on same host. Offline static example below remains readable.

Backend not queried

Registered comparisons

Phone onlyOne visual source
Metrology onlyClearance model
Conservative rulesStrong comparator
SAFE-MESH referenceSame safe decisions + inspection overhead

This reference deliberately does NOT demonstrate scientific superiority. Run physical experiments to test our hypotheses.

Example 30 × 1000 simulated specimens — not factory evidence

Quality / Metrology

Tools: check_clearance, check_visual, read_evidence; no direct authority over final PASS.

Design / Assembly

Review CAD revision, geometry fit, compatibility, check for independent risk evidence.

Inspection Planner

Rank next best inspection by expected decision loss, time and cost.

Agent Evidence Room — structured request trace

Deterministic roles over TEST_FIXTURE evidence, in one process. Shows the typed protocol (REQUEST → RESPONSE → VERIFY → UPDATE → PROPOSE → DECISION), the ACL, and whether a request actually changed anything versus the no-request control. Not an LLM, not a network, not a benchmark.

Requires the local API.

Nebius adapter status

Live model inferenceDISABLED BY DEFAULT
21 rolesREGISTERED
Server-enforced tool ACLREFERENCE TESTED

Select an entitled model from the live Nebius model catalog and explicitly opt in before making paid API calls.

Model inputs

Costs are assumed until independent factory invoices/logs verify them.

The backed economics module prevents duplicate terminal defect outcomes.

Observed economic conclusion

NO PROVEN SAVINGS

Current reference policy makes the same HOLD/PASS decisions as conservative rules, but may spend more on additional inspections. Any claim of savings needs actual physical or pre-registered repeatable evidence.

QualityRelay 17.0 research-reference website · Local photo baseline does not upload images to a server · Demonstration actions are non-actuating.