Framework
Why Doesn’t a Robot’s Own Page Count as Evidence for Its Classification?
Before one rule changed, 473 links pointing back at this site were counted as outside proof across 237 robot records. Excluding them moved 81 records and cut the verified count from 328 to 258. No robot changed rung. What changed was who is allowed to testify about a machine, and a larger gap the rule leaves open.
Photo by Pavel Danilyuk on Pexels
Every robot in the Robovations database carries an evidence label, and the label answers one question: how much material from outside this publication backs the record. For 237 records, part of that material was never outside anything. It was us.
An automatic process had been closing a channel called longitudinal with two links: the robot’s own page on this site, and its entry on the Tracker. Both were marked complete, the same status an independent teardown earns.
The rule that removed them is short. What it excludes, and why the line sits where it does, is the subject here.
A channel that cited itselfOur Own Pages Were Counted 473 Times
The evidence harvester fills six channels for each robot: manufacturer disclosure, third-party testing, owner reports, teardowns, incident or recall records, and long-run tracking of the same machine. The last channel exists to reward a robot that has been watched for months, not reviewed once at launch.
The harvester built that channel from material it already held. Neither link involved anyone outside this publication looking at the machine. Across the database the pattern repeated 473 times on 237 records.
The fix is one test applied before any link counts: is the publisher Robovations, or does the link point at this site? If either is true, the link never reaches the tally.
Six channels, one publisher excluded
What feeds an evidence label, and what does not
- Kept and indexeda failure enters a record that the next buyer, regulator and manufacturer can all read
- Loosethe evidence exists, and finding it depends on who happened to post
Independence is the whole pointA Page Cannot Corroborate Itself
An evidence label stands for a specific kind of confidence: that more than one party, working from more than one vantage point, arrived at compatible facts. A teardown and a spec sheet can disagree. When they agree, the agreement is what carries weight.
A robot’s own page cannot disagree with itself, and neither can its Tracker entry. Both were written by the same publication applying the same rules to the same record.
A mirror does not become a witness by reflecting twice.
The distinction
The maker still counts. The classifier does not.
A manufacturer describing its own robot is a second party to the classification, so its material still counts. Firmware notes on a maker’s support portal are recorded as ongoing rather than complete, and ongoing rows never counted toward a verified label. What the rule removes is narrower: this publication’s own page, cited as proof of its own claim.
Excluding the self-citations moved 81 records between evidence tiers. The number carrying the verified label fell from 328 to 258.
The Autonomy Ladder rung is graded on a separate axis. Of 496 classified robots, 286 sit at Level III, Conditional Autonomy, 166 at Level II, 39 at Level IV and 5 at Level I. No robot changed rung when its channels were recounted.
A rung records what a machine does. The evidence label records how well that is documented, and a record that still reads verified after the fix earned the word from someone other than us.
What moving the line would doEither Direction Breaks the Label
Move the line to admit self-citation again and the label stops measuring evidence. Every new robot gets its own page and its own Tracker entry the day it publishes, so the verified count would climb with the size of the database, not with anything anyone outside learned.
Move it the other way, to exclude manufacturer material, and nearly every record loses its base layer. The maker’s specification is where most records begin, and it is a real second party even when it is not a neutral one.
The line sits at the one source that can never be a second party: the classifier itself.
What the rule does not reachA Link Count Is Still Not a Channel Count
The longitudinal channel answers a question the other five cannot: what has happened to a machine since launch. Few outlets follow one consumer robot for a year, so for most of the 237 records nothing else was there. The fix removed the substitute. It did not supply a replacement.
A second gap stays open. The harvester grades a robot by how many links it holds, while the published Evidence Standards describe a count of independent channels. One obliging forum thread can supply a dozen links to a single channel.
Only 86 of roughly 490 robots hold five of the six channels the methodology names.
Closing that gap means changing what the label counts, not just who may contribute to it.
A page this publication writes about its own record is not a second opinion, however often a channel repeats it. Eighty-one records changed label once the voice behind part of it turned out to be ours.
