How a case gets here, and what it does not prove

One case is one real-world event involving a physical commercial robot: a delivery robot, a robotaxi, an industrial arm, a surgical system, a consumer or service machine. Software agents and chatbots are out of scope. So are driver-assistance systems such as Tesla Autopilot or FSD (Supervised): a vehicle is a robot here only when it drives with no one responsible inside, so Level 2 crashes, fatal ones included, are not cases. NHTSA publishes those separately in its Level 2 ADAS Standing General Order data.

Articles are evidence about a case, not cases themselves. A collision, the investigation that follows it, the recall, the claim and the settlement are one file with a timeline, because they are one event. This is the whole reason the register exists: that sequence cannot be reconstructed cheaply after the fact.

The timeline holds one entry per step. When several outlets cover the same step, one entry stays in the timeline and the rest stay in the case's source list, so the sequence reads as a sequence and the evidence is still all there.

Nothing publishes on a single source. A case must clear all of the following, and an item that fails any one of them waits rather than appears: That rule governs cases built from news reporting. Cases resting on a document from a regulator or a court - 69 of them, from NHTSA, OSHA, CPSC and the EU Safety Gate - publish from that single document on purpose: the document is the evidence rather than an account of it, and it does not go through the news gates. Separately, 140 news cases published before the corroborating set was recorded keep only the source they were written from. The second domain was required at the time and simply not retained, and each of those pages says so.

• At least two independent domains reporting the same facts. Syndicated copies of one text count once, and tier 3 does not count toward the two at all: a real outlet plus a forum thread is one source, not two.
• At least one source at tier 1 or tier 2. Blogs, forums and social posts are never sufficient on their own, in any number.
• Where a case names a company and involves serious injury or death, the bar rises: three independent sources minimum, at least one of them tier 1. This is where the risk of getting it wrong is highest, so this is where the standard is highest.
• No contradictions between sources on the key facts: company, model, date, place, severity.

tier 1

Wire services, national outlets, regulators, court documents.

tier 2

Regional and trade publications, and any outlet that does its own reporting.

tier 3

Blogs, forums and social posts; tabloids; sites that retell someone else's reporting; portals that reprint other outlets' articles. Supporting evidence only, never a basis.

The test for tier 3 is not prominence or tone. It is whether the outlet reports. A publication that retells a story first told elsewhere, or a portal that reprints it under its own domain, cannot independently confirm anything: it is the same account wearing a second name, and counting it twice would be counting one source as two.

An outlet we have never seen before is treated as tier 2, not as suspect. Much of what this register exists to record is first reported by one local paper near the warehouse or the intersection where it happened. An allowlist of famous names would quietly discard exactly those incidents.

Cases are assembled by an automated pipeline with AI verification, under the rules above. There is no human confirmation step before publication, and pretending otherwise would be the dishonest part. What replaces it is a set of hard gates: an item either satisfies every one of them or it stays unpublished.

The gates are not a description of code, they are the code. Since October 2026 the rules above, the source tiers, the court-record policy and the stripper that removes private individuals' names run from an open-source package, tmr-gates (MIT licence), which this pipeline imports unchanged. Its README documents every rule with its threshold, and its tests are the acceptance criteria of this register: anyone can run our own cases through the same gates and get the same decisions.

Two independent AI passes run on each candidate and do not share conclusions. The first asks whether the sources describe the same event and agree on the facts. The second is adversarial: it looks only for reasons not to publish, such as confused companies, an old event presented as new, satire, one source in several wrappers, or named private individuals. Publication requires the first to pass and the second to find nothing blocking.

Press in a language other than English is judged by the same two-source rule, the same tiers and the same three-source rule for deaths and serious injuries, with two allowances that the gates record on each case: a regional paper that names the injured worker or writes "yesterday" is not refused for that, because the name is stripped before publication and the event date is required anyway, and the risk ceiling for such a case sits between the press ceiling and the court-record ceiling. The thresholds are in the package README.

Published cases are re-verified automatically. If new sources contradict the record, a case is marked Disputed in public; if it no longer clears the gates, it is withdrawn without anyone having to notice.

Fleet sizes are not public. Nobody outside a company knows how many of its machines are deployed or how many hours they run, so incidents per thousand units cannot be computed and we do not publish a number that pretends to be one.

The index measures something that normalises itself: the share of a company's recorded incidents that reached a regulator, a recall or a court.

formula

cases with escalation stage at regulator or beyond, divided by all cases on record for that company

chain

report, follow-up, investigation, regulator, recall, fine, lawsuit, settlement, criminal charge

threshold

Companies with fewer than 8 cases are excluded: at that sample size one case moves the share by double digits.

This is a computed metric, not a safety rating. It says how often incidents attributed to a company had consequences, which is a fact about outcomes and press coverage together. It is not a judgement of how safe a product is.

• The denominator exists only for driverless cars in California. Rates per million miles are computed where an operator reports both its crashes to NHTSA and its miles to the CPUC for the same months. For recalls a rate per unit is possible where the notice disclosed the fleet size. For everything else we have a numerator and no base.
• Coverage is still mostly English-language. Until October 2026 an incident in Shenzhen or Osaka entered the register only if somebody wrote about it in English. The collector now also reads German, French, Italian, Spanish, Dutch, Polish, Japanese, Korean and Chinese press, with the same gates, and each source carries its language. That coverage is newer and thinner, so Asia and Europe remain under-represented relative to reality.
• Attention is uneven. A widely covered company accumulates more cases than a smaller manufacturer with the same incident. Case counts measure reporting as well as events.
• Routine faults are invisible. News covers deaths, fires and lawsuits. The everyday malfunctions that would make up real statistics never get written about at all.
• Dates and places come from article bodies, not headlines. That correction followed a real error: a fire at UC Berkeley was recorded as a fire in San Diego, because a San Diego paper had covered it.

An error is not deleted quietly. It is added to the case timeline as a Correction, with the previous version kept, which is the press standard and works in favour of trust rather than against it.

Any company named in a case may respond. A request makes the pipeline re-run the same cross-verification that published the case, on the same sources: the disputed mark follows the evidence rather than the request, because a form anybody can submit would otherwise discredit any case at will. The response itself is published in the case unedited. Target turnaround is five working days.

cases

257 published

escalated

44 reached a regulator, recall, court or settlement

sources

662 archived

companies

48 named

earliest

1979, the oldest documented event on file