The rmForge Risk Analysis platform: questions and answers
← Return to FAQWhat language model does rmForge use?
rmForge uses state-of-the-art commercial large language models from leading providers, with backup models, so a failure of one model moves the step to the next. Specific model selection is managed by the platform and tuned for the risk-analysis task.
How long does rmForge keep my files?
Uploaded source documents and generated outputs are kept for the life of your account. You can ask for them to be deleted at any time by emailing support@rmforge.io; we confirm the request with you before anything is deleted.
Can I download my data?
Yes. Every completed analysis produces an Excel workbook that you can download from your dashboard: the Hazard Analysis workbook, or a Design or Process Failure Mode and Effects Analysis (FMEA) workbook. Your acknowledgment is recorded on the Disclaimer tab inside each workbook. There is no lock-in.
Does rmForge train its models on my data?
No. Customer data is never used for model training. Our AI model providers contractually exclude customer data from training under their API terms. rmForge's invocation pattern is API-only. No model traffic touches the chat product. Provider API terms contractually exclude API traffic from training.
How is rmForge different from a generic LLM?
Three things: (a) hard-coded task-specific sections and a controlled prompt designed around ISO 14971 and our internal specifications; (b) literature based risk techniques, purpose built code, deterministic grammar, and logic rules plus risk math (Po and Risk Level via table lookup, not AI judgment); (c) validation rules that check every row against the specification and flag any row that cannot be validated; three kinds of row are left out instead of flagged.
What if the LLM produces a row that doesn't match the spec?
rmForge runs an extensive validation rule set on every row. A row that names a hazard your setup answers do not allow, words a library harm differently, or puts a fault under misuse without saying why that is foreseeable is automatically rewritten, up to three times. A row that still cannot pass, or that still breaks a grammar rule after one rewrite, is kept and flagged in the output for human review, just like a normal risk analysis process, but with significant time savings. Three kinds of row are left out instead of flagged: a row with a required field missing or invalid, a row the software reads as being about workers who make, sterilize, pack or ship the device, and a row written under the wrong hazard category.
How does rmForge handle hallucinations?
Risk math is deterministic, not AI-generated. Po and Risk Level are looked up from controlled tables. Sequence of Events, Hazardous Situation, and Harms follow programmed rules and logic significantly narrowing LLM decision making to a finite rule set in the similar manner that a qualified human would evaluate risk.
Can I customize the HA prompt?
No. The prompt is controlled under our internal change-management process. Customization can be contracted through support@rmforge.io.
Does rmForge support FMEA?
Yes. rmForge builds a Design Failure Mode and Effects Analysis (FMEA) and a Process FMEA. Today each one is built on a finished Hazard Analysis: the rmForge operator starts it from that Hazard Analysis's status page, with Generate Design FMEA or Generate Process FMEA. Starting one also records, with the date, that you accepted the Hazard Analysis as its baseline. That record is not a signature; your approval stays on the Cover Page sign lines. A Use FMEA is not built yet.
Can rmForge handle non-English IFUs?
Today, English only.
Is rmForge a medical device?
No. rmForge is not a medical device and not Software as a Medical Device (SaMD). It is a *user review-required* documentation-generation tool. The user remains the legal manufacturer with all responsibility for risk decisions.
Who is responsible for the final risk analysis?
You are. rmForge provides a structured first draft, but the manufacturer is responsible for the analysis under ISO 14971, ISO 13485, and FDA QMSR. The customer acknowledgment form captures this.
How often does rmForge update its rules?
Major rule changes, such as to the severity scale, the probability scale, or the validation rules, go through an internal change control process with a revision history. Minor wording updates ship continuously.
Where is rmForge's infrastructure hosted?
Hosted on industry-standard cloud infrastructure. The application, database, and storage run on tier-1 US-region cloud providers. All traffic is HTTPS end-to-end.
Which parts of an rmForge Hazard Analysis are fixed rules, which are AI, and what does my team decide?
Fixed rules do the risk math: Po and Risk Level by table lookup, nothing multiplied, and P1 set to 5 for a software cause at Severity 4 or 5. They also decide which hazards your setup answers allow, check harm wording and grammar, and lay out the workbook. AI reads the text of your documents, answers the 48-question Hazard Questionnaire and drafts each row. A row that breaks one of the hazard or harm rules is rewritten up to three times and, if it still breaks the rule, kept with a note for your reviewer. AI also decides Essential Performance for medical electrical equipment, picks a harm code when the harm's name is not in the published list, and, when you bring them, matches risk controls to your own requirements and your harms to the library. Design and Process FMEA rows are drafted by AI too. Your team decides whether each row is right and what risk is acceptable, approves the Hazard Analysis on its Cover Page, and accepts it as the baseline before an FMEA is built on it. The How rmForge works page lists each part.