Method
For each of the 120 exact models in catalog v0.4.15, RCI collected the set of distinct source URLs cited by that model’s manufacturer-published (M) claims, counting both primary and additional sources. Claims in other evidence states—independently verified, derived, conflicting or unknown—are excluded, because the question under audit is specifically how deep the manufacturer-published layer runs. The computation is reproducible from the public component catalog JSON.
Results
- 84 of 120 models (70%) trace every manufacturer-published claim to a single manufacturer URL.
- 30 models draw on two distinct URLs, and 6 on three.
- Across the whole catalog, 151 distinct manufacturer URLs stand behind all 1,080 manufacturer-published claims—an average of nine claims per source URL.
| Hardware layer | Models | Single-URL models | Share |
|---|---|---|---|
| Motion & Actuation | 43 | 34 | 79% |
| Perception & Sensing | 28 | 17 | 61% |
| Compute & Control | 24 | 18 | 75% |
| Power & Energy | 14 | 10 | 71% |
| Interfaces & Integration | 11 | 5 | 45% |
The GO-M8010-6 example, confirmed
The reviewer picked the Unitree GO-M8010-6 record as the illustration: twelve published claims, all citing the same Unitree page. Recomputed: the record carries 14 claims—12 manufacturer-published and two explicit unknowns—and all 12 M claims cite one URL. The record is not wrong; the torque, speed, mass and interface values preserve their conditions and the two unknowns (continuous torque, independent robot validation) are honestly held open. But nothing on that page corroborates Unitree. The structure is RCI’s work; the numbers are one vendor speaking once.
What the metric does and does not say
A single-URL model is not a defective record. For many components the manufacturer datasheet is genuinely the only public specification, and preserving its exact values, units and conditions with a check date is the audit trail RCI exists to provide. What a single-URL model cannot support is any suggestion of triangulation: no second document constrains the first. The evidence-state system already said this claim by claim—M means manufacturer-claimed—but state labels on twelve rows from one source can read as more diligence than they are. The per-model source count removes that ambiguity.
The metric also does not measure truthfulness. A claim corroborated across three manufacturer URLs is still the manufacturer talking three times. Independent verification remains a separate, harder standard, currently met by six catalog claims with 33 models carrying any direct external measurement.
What changes
- Every component page now displays its distinct-manufacturer-URL count next to the claim total, with an explicit note when the count is one.
- The homepage evidence-composition section states the 84-of-120 figure alongside the M/V split.
- Raising source diversity is now tracked work: datasheet-PDF-plus-product-page pairs, manual revisions, certification filings and declaration documents each count as distinct URLs only when they are actually cited claim by claim, not by listing them once at the page level.
Corrections
No corrections yet. Material changes will be logged with timestamps on the Changes page.