Quality & Safety tools
A collection of demonstration tools around standards compliance and information security. Each one really runs — nothing is simulated.
Six tools. Five without AI.
An ASIL, a risk value, an 8D that is refused closure: these are rules. AI is only essential to the QualityCrew audit.
QualityCrew
ASPICE / ISO 26262 audit written by four agents. Run the demo OnlineSentinelScan
Information leaks in public repositories. Run a scan OnlineSafetyScope
HARA and ASIL level, with no waiting. Flows on into ThreatScope Rate an item OnlineThreatScope
TARA all the way to cybersecurity goals. Carries over the severity from your HARA Analyse the threats OnlineRegWatch
Public signals around the standards. Never republishes the content of standards Run a watch OnlineCauseTrace
8D complaints and root-cause analysis. Requires the escape root cause, not just the occurrence one Open an 8DWhere we stand
Five of the six tools deliver their result without AI: SentinelScan uses none, it is optional in four others, and only QualityCrew depends on it.
tools deliver their result without AI
Example: a hazardous event rated S3, E4, C2 gives ASIL C. Only severity crosses into the TARA, where it becomes a severe impact; exposure and controllability do not cross. Combined with a high feasibility — a radio dongle, 6 points — it gives a risk of 5.
only severity crosses the HARA → TARA bridge
Five of these six tools produce their result with no artificial intelligence at all. Determining an ASIL or a risk value are decision tables: the answer is exact and instantaneous. Refusing to close an 8D whose escape root cause is missing is a set of ordered rules. Attaching a watch signal to a standard is a written rule, reproducible and tested offline. Looking for an exposure on public repositories is an API and criticality criteria. Dropping a language model in would add nothing but latency and uncertainty.
Only one really depends on it: QualityCrew, where four agents write the audit — that is its whole purpose. Everywhere else AI stays optional: it sweeps guide words to prime a line of thought, or sums up in one sentence why a signal deserves attention. Never does it rate, never does it decide what is kept. Knowing when not to use an LLM is part of the craft.
RegWatch never republishes the content of a standard. These documents are paid-for and protected: the tool reports only the title, the date and the link to the source — the body of the pages is not even downloaded. And because not all sources are equal, every signal shows its own tier: an ISO committee is not a consultancy blog.
In CauseTrace the AI does not fill in: it asks. An 8D report goes to the customer; an invented fact there would be far worse than a clumsy sentence. So it tightens what the engineer wrote, and anything it would like to add without knowing — a date, a number of parts — becomes a question, never a value. An Ishikawa lead taken up arrives unqualified: it is the engineer who says what it is.
SafetyScope and ThreatScope talk to each other. The severity of a hazardous event becomes the « safety of persons » impact of a damage scenario. Exposure and controllability, for their part, do not cross: an attacker picks their moment, and can neutralise the driver's recourse.