1. Mandatory engineering controls
For every consequential design, you must define requirements and operating envelopes; identify assumptions and unknowns; verify units and coordinate frames; use authoritative material/component data; independently review geometry, tolerances, joints, loads, failure modes, electronics, controls, firmware, cybersecurity, and manufacturing; run appropriate analytical and high-fidelity analyses with convergence and uncertainty evidence; prototype and physically test under foreseeable normal, fault, misuse, and environmental conditions; preserve traceability; resolve every hard-gate failure; and obtain approval from competent, authorized humans.
Images, simulations, surrogates, AI judgments, and plausibility are not substitutes for calibrated measurement. Provider “common sense” is not spatial or engineering evidence. If output conflicts with a deterministic check, standard, drawing, datasheet, test, or qualified reviewer, treat it as failed until resolved.
2. Prohibited uses
- Any unlawful act, fraud, deception, rights violation, malicious intrusion, malware, credential theft, or circumvention of security/safety controls.
- Design, development, optimization, acquisition, or deployment of nuclear, chemical, biological, radiological, or other weapons of mass destruction; prohibited missiles/delivery systems; prohibited anti-personnel weapons; or any end use barred by applicable export or sanctions law.
- Unlawful surveillance, biometric identification, targeting, discrimination, social scoring, or decisions that violate privacy, employment, housing, credit, insurance, education, policing, immigration, or civil-rights law.
- Creating sexual content involving minors, exploitative content, trafficking tools, torture devices, or instructions intended to cause non-consensual harm.
- Falsifying engineering evidence, provenance, tests, certifications, professional seals, regulatory submissions, inspection records, or claims that AI output was independently reviewed.
- Bypassing interlocks, e-stops, calibration requirements, release gates, access controls, or a qualified reviewer’s refusal.
- Manufacturing or releasing a design with unresolved collision, structural, thermal, electrical, firmware, tolerance, material, compliance, or verification failures.
3. High-consequence and regulated domains
The free beta is not approved as the sole or controlling system for aerospace/spaceflight; autonomous or road vehicles; rail or marine safety; medical devices, diagnosis, implants, prostheses, life support, or human research; nuclear/radiological systems; power grids and critical infrastructure; pressure vessels and process plants; cranes, hoists, elevators, guards, and industrial machinery; firearms, ammunition, energetics, or military targeting; buildings/structural systems; cleanroom/semiconductor safety; or environmental-control systems.
Use in such domains requires a separate written authorization from Silvia AI where offered, plus all applicable licensed professionals, quality systems, standards, hazard analyses, independent verification and validation, cybersecurity review, controlled configuration, regulator/notified-body interaction, and physical qualification. The user remains the legal manufacturer/deployer and release authority unless a signed agreement expressly says otherwise.
4. AI-specific risk controls
- Assume generated requirements, dimensions, citations, standards, component availability, prices, and calculations can be fabricated or stale. Verify against primary authoritative sources.
- Do not let the generating model judge itself. Use independent deterministic checks and qualified review.
- Review the full patch and geometric diff. Reject stale-base changes, silent representation switching, and unrelated edits.
- Validate inputs, units, frames, boundary conditions, contacts, mesh/convergence, material curves, tolerances, solver pedigree, and uncertainty.
- Protect against prompt injection and malicious content in drawings, websites, supplier files, CAD, and attachments. Treat external instructions as untrusted data.
- Keep secrets, personal data, classified data, and export-controlled technology out of model prompts unless the chosen service and organization are explicitly approved.
- Use sandboxed, least-privilege tools. Never expose arbitrary shell, unrestricted Python, production machinery, or safety controls to a generative model without an independent governor.
5. Manufacturing and physical experiments
Before fabrication, review material/process compatibility, dimensions and tolerance stack, tooling, workholding, machine envelope, collision, feeds/speeds, support removal, heat treatment, surface finish, inspection plan, PPE, ventilation, fire/chemical hazards, and applicable code. Simulate toolpaths using the machine’s authoritative system. A human operator controls machine start.
Automated experiments require a signed safety envelope, physical interlocks, e-stop, bounded energy/material/temperature/pressure/speed, calibrated sensors, preconditions, abort criteria, supervision appropriate to risk, post-test inspection, and incident logging. The AI must never modify or bypass the governor.
6. Electrical, firmware, and mechatronics
Verify mains isolation, protective earth, fusing, creepage/clearance, insulation, enclosure/fire rating, voltage/current/power/thermal budgets, battery protections, connector pinout and accessibility, grounding/EMC, fail-safe states, watchdogs, update/recovery paths, and applicable ERC/DRC. Compile and test firmware on the exact target. Unexpected ERC/DRC, invalid voltage/pin assignment, firmware failure, or mechanical collision blocks release.
7. Medical, biometric, and personal data
Supra has no general HIPAA BAA, regulated medical-device quality certification, or clinical authorization. Do not use it for diagnosis, treatment, patient-specific manufacture, implant release, or clinical decisions. De-identification must be performed before ingestion and independently assessed. Human-subject and biometric work requires ethics, consent, privacy, security, and regulatory review.
8. Export, security, and dual-use review
Classify controlled software/technology and screen destination, end user, ownership, end use, and deemed-export access. Do not assume a public download makes an end use lawful. Follow applicable EAR, ITAR where relevant, OFAC sanctions, EU/UK and other controls, organization policies, and license conditions. Report suspected diversion or compromise to appropriate compliance personnel.
9. Reporting and enforcement
Stop use when Supra behaves dangerously, evidence is inconsistent, calibration is stale, provenance is missing, or a hard gate fails. Preserve the project, version, logs, inputs, outputs, and evidence without exposing secrets; contact hello@silviaai.dev. Silvia AI may restrict features or terminate access for credible dangerous or unlawful use. Emergency situations must be handled by local emergency services and qualified personnel, not product support.
10. Regulatory responsibility
Supra may fall within different AI, product, professional, consumer, privacy, cybersecurity, employment, sector, and machinery regimes depending on deployment. You must classify your use. If you deploy Supra as a component of a high-risk system, substantially modify it, or place an AI-enabled product on the market, you may acquire provider, deployer, importer, distributor, or manufacturer duties. These policies do not perform that assessment or transfer those duties.