How this was checked
Each factual assertion in the critique was tested against the current frozen component catalog (v0.4.15: 120 exact models, 1,205 claims), the evidence-status ledger, and the live page copy it quoted. Where the critique cited a number—the M-to-V ratio, category counts, the single-URL pattern on a named component page—the number was recomputed from the published data rather than argued about. Strategic judgments (whether the moat exists, whether the project should exist) are not factual claims and are not adjudicated here; they are acknowledged where the underlying facts support them.
Point-by-point results
| Critique claim | Verdict | What the data shows |
|---|---|---|
| About 90% of the catalog is rehosted manufacturer specification data. | Held | 1,080 of 1,205 claims (89.6%) are manufacturer-published. Six claims carry independent verification. RCI has run zero physical tests. |
| The power layer is too small and skewed to portable smart packs; the interface layer is mostly CAN adapters. | Held | Power-energy holds 14 records, eight of them removable smart packs from three portable-power vendors and none a 48 V robot system pack. Interfaces holds 11 records, ten of them CAN/CAN FD interface devices. |
| Inclusion reflects published-source availability, so the index samples manufacturers that already document well. | Held | The bias is real and structural. It is now named as a bias rather than a neutral rule, and documentation-gap records are committed as the counterweight (RCI 061). |
| Evidence states are close to decorative: one source copied twelve times with a schema around it. | Held in substance | 84 of 120 models trace every manufacturer-published claim to a single URL; the GO-M8010-6 example is accurate. The M/V/D/C/U states are real distinctions, but state labels cannot add source diversity that does not exist. Per-model source counts are now published (RCI 058). |
| Headline numbers perform: a 90-of-120 physical-evidence milestone that immediately retracts to 33 direct measurements, and a pre-registered design published as a data release. | Held | The 90/120 framing led with the biggest available number and required the reader to decode the retraction. Site release 0.13.0 inverts every such headline and relabels the pre-registration file as a protocol document, not data. |
One factual point in the critique is partially out of date: the absence of mainstream humanoid joint-module suppliers and Chinese supply-chain coverage. Since catalog v0.4.15, motion-actuation includes exact records for Unitree GO-M8010-6, MYACTUATOR RMD-X8-P9-25, FEETECH ST-3215 variants, Damiao DM-J4310-2EC, Leaderdrive LCS-20-100 and Laifual FSS-20-100. Coverage remains thin relative to the market, and the broader thrust—that scale is insufficient for real selection work—still holds.
What changed in site release 0.13.0
- Direct measurements lead every evidence headline. The homepage, evidence page and data page now open with 33 of 120 models having direct external component measurements and 14 of 1,205 claims independently supported. The broader 90-model bounded-coverage figure is stated second, as context, with its weaker evidence classes named.
- The self-fulfilling milestone is gone. The previous milestone set its target at 90 models after 90 were reached and reported 100% complete. The published milestone is now direct component measurements for all 120 catalog models: 33 done, 87 remaining, 27.5% complete.
- Source concentration is a published metric. Every component page now displays how many distinct manufacturer URLs stand behind its manufacturer-published claims, with an explicit warning when the answer is one. The full audit is RCI 058.
- The pre-registration file is labeled a protocol, not data. The data-releases page now lists it under “Protocol registration — not a dataset” and states that it contains zero observations.
- Category pages state what is actually on the shelf. Each catalog layer on the component index now shows its current subcategory composition, so ten CAN adapters cannot present themselves as an integration layer. The coverage audits are RCI 059 and RCI 060.
- The inclusion rule is named as a bias. The component index and RCI 061 state that published-source inclusion over-samples transparent manufacturers and under-samples the vendors most needing scrutiny.
What a copy change cannot fix
Honest framing does not create evidence. After this release the catalog still contains zero RCI physical tests, six independently verified claims, and 84 single-source models. The critique’s first point—that the structured middle layer sits on top of rehosted specifications, and that durable value requires physical testing or scale—is a work problem, not a wording problem. The open paths remain the D455 × J4012 bench protocol, the SO-101 rebuild package, external testing partners, and the six zero-record coverage-debt areas listed on the evidence page. Until one of those produces data, this site’s honest description is: a source-structured, condition-preserving index that is explicit about not being a verification body.
Corrections
No corrections yet. Material changes to this note will be logged with timestamps on the Changes page.