Research desk
Research pages explain evidence boundaries and disagreements. Publication volume is not a proxy for independent verification; each page must state what work was actually performed.
34 of 61 current notes carry 10–11 August publication dates, and 28 titles use “Zero” or “0”. Shared audit structures made this possible, but the cadence and repeated headline formula reasonably raised questions about review depth.
A catalog built on published-source availability systematically over-samples manufacturers that already document well and under-samples the opaque vendors an independent index exists to scrutinize. This note names that bias, quantifies where it shows, and commits to documentation-gap records as the counterweight.
Eight of the fourteen power-energy records are removable smart battery packs from three portable-power vendors, three are bare 21700 cells, and none is a 48 V system pack, charger, power-distribution unit or hot-swap dock of the kind humanoid and mobile-robot power chains are actually built around. The gap is published as explicit debt with named first targets.
The interfaces-and-integration layer promises mechanical, electrical, data, timing and thermal edges. What it currently contains is ten CAN/CAN FD interface devices from four vendors plus one unmanaged Ethernet switch. This audit states the mismatch plainly and lists what an honest integration layer requires before the category name is earned.
Data Brief 005 measures how many distinct manufacturer source URLs stand behind each catalog record: 84 models rest on a single URL, 30 on two and 6 on three, with 151 distinct URLs behind all 1,080 manufacturer-published claims. The metric now appears on every component page, because a well-structured single-source record is still a single-source record.
An external reviewer called the catalog rehosted manufacturer specs with decorative evidence states and performative numbers. RCI verified each factual claim against its own frozen data—most held—and shipped changes: direct-measurement counts now lead every evidence headline, the self-set 90-model milestone is replaced, source concentration is disclosed per model, and the pre-registration file is no longer presented as data.
A practical workflow for checking exact model identity, claim meaning, operating conditions, evidence state, source date and unresolved conflicts before treating two robot-component specifications as comparable.
RCI publishes the uncomfortable baseline plainly: 1,080 of 1,205 catalog claims are manufacturer-published, only six carry independent verification, and RCI has completed zero physical component tests. The note also addresses the August publication burst, repeated headline patterns and AI-assisted workflow.
Data Brief 004 compares frozen catalog v0.4.14 and v0.4.15 — motion-actuation grows 26→43 exact records across joint modules, reducers and frameless kits, with seven new manufacturers — while independently verified claims stay at 6, comparisons stay at 6 and zero claim IDs are removed.
A field note states the single derivation rule behind the five pinned ToddlerBot configurations — deep-merged default and configuration YAML filtered against fixed XML at one commit — producing family-level actuator assignments, not order codes, as-built inventory or a purchased J4012/Orin NX.
A field note walks the nine-node, ten-relation Berkeley Humanoid Lite system map — NUC, USB hubs, four USB-CAN transports, two actuator banks, alternative IMU paths and protected power — while live documentation stays an observation layer, not the frozen v1.1.0 BOM.
A field note shows how 42 pinned AlohaMini2 BOM rows become a Host/Client system map while the unexplained four-cable/two-controller count, battery-branch allocation and camera transport remain open unknowns with zero RCI tests.
A field note maps the 46 official shared-table rows to the Waffle Pi configuration column and its two exact-model crosswalks while keeping the Raspberry Pi, camera and LDS revision questions unresolved rather than answered.
The pinned SO-101 build documentation links a 5 V supply for the 7.4 V ST-3215-C001 follower path while FEETECH publishes a 6–7.4 V servo range and Waveshare publishes 9–12.6 V for the SKU 25514 controller; RCI records the three statements side by side and holds the conflict open with zero physical tests.
A recomputed baseline-to-current comparison shows exact-model linkage rising from 26 to 45 profiles and all fifteen sector dossiers gaining a linked exact record, while preserving every screening, linkage and zero-physical-test boundary.
Data Brief 003 compares frozen catalog v0.4.7 and v0.4.14 — 76→103 records and 599→945 claims — while independently verified claims stay at 6, comparisons stay at 6 and zero claim IDs are removed.
A launch audit recomputes the v0.8.21 terminal data layer from the published dataset—five morphologies, 168 topology nodes, 159 interface relations, 21 lifecycle relations—and preserves the boundary that no relation is physically validated and no complete rebuild package exists.
A reproducible 96→101 source-directory comparison identifies five official engineering release channels, recomputes five-layer coverage and preserves item-level rights and bounded network-observation limits.
A claim-provenance audit separates the named Kria AI SOM and developer-platform descriptions from modeled workflows, a controller result and a GMKtec proxy benchmark that did not test the named module.
A fixed-paper review separates a shared actuator-level policy and one reported physical transfer case from arbitrary-configuration validation, repeatable hardware evidence, exact component identity and independent RCI execution.
A facts-first 6-K review separates broad power, sensor, embedded-processing and SiC manufacturing signals from robot revenue, exact product demand, fab capacity, qualified yield and independent validation.
A facts-first SEC filing review separates one combined quarterly shipment signal and company-level supply context from exact lidar counts, model adoption, customer deployments and independent performance validation.
A fixed-paper review separates one RTX 4060 backend benchmark and five reported resolutions from excluded rendering overhead, absent physical-camera comparisons, untested robot workloads and unresolved public-code availability.
An independent byte-integrity and disclosure audit separates twelve hash-matched assets and a fixed source tree from absent public checks, untested hardware, unresolved model rights and conflicting extension-license wording.
A quantitative map of nine source groups, five hardware layers and five reuse classes explains where frontier signals begin, how claims resolve upstream and what RCI may cite, summarize, comment on or republish.
A common evidence frame joins robot vision, LiDAR, main compute and battery-management releases while separating source visibility, public CI and digest matches from physical-device and robot validation.
An independent release audit separates eight hash-matched Windows assets and 28 successful public checks from unresolved beta semantics, exact firmware bytes, retained result bundles and camera or robot validation.
An independent release audit separates a byte-matched source tree and explicit Mid-360S code surfaces from missing firmware matrices, public test results, device benchmarks and robot integration evidence.
An independent release audit separates a complete public source tree, extensive test structure and ten hardware links from unpublished HIL material, release results and robot battery-system validation.
An independent release audit separates fixed source identity, three EULA-gated runtime distributions and class-level support statements from missing exact-host protocols, result bundles and robot workloads.
An independent release audit separates one hash-matched PyPI sdist, 35 publisher-indexed wheels and a 138-bullet migration surface from missing exact-sensor benchmarks, run logs and test protocol.
An independent release-integrity audit separates three hash-matched SDK assets, five PyPI wheels and issuer compatibility statements from missing device benchmarks, raw logs and test protocol.
A source-level audit separates a tested BOPP/gold-electrode expansion actuator and four reported NRMSE values from an unbuilt three-stack robot, untested puncture and future MRI work.
A public-package audit separates six-unit architecture and three author-reported axis MAEs from the missing dataset, controller code, complete hardware identity and test definition.
An exact-model audit separates conditioned ST datasheet specifications from two manufacturer-authored humanoid-current plots with no public raw traces or complete test method.
An independent integrity, BOM, topology and version-drift audit separates a stable six-file design core from source-reported performance and missing raw test traces.
An independent parse reconciles all eight paper totals, preserves the reported cross-unit model gap and exposes an unresolved 4-unit versus 12-ms timing boundary.
An independent recomputation separates 10 Hz grid density, 0.1 A / 0.5 V reporting resolution, repeated current states and a bounded current–motion association.
An exact-model audit separates connectors, power, data, timing, software and validation across a D455, MID-360, J4012 and RRC2054-2 battery.
Data Brief 002 compares frozen catalog v0.4.0 and v0.4.5, then separates evidence depth, external validation and search discovery outcomes.
A source-level audit separates two public Rev7-to-Rev8 changes from seven unresolved differences across current official pages, manuals and datasheets.
A twelve-field audit separates public parts lists from versioned, as-built and independently replicated evidence across five robot platforms.
RCI-BLM-1 defines a complete quasi-static torque–angle hysteresis test, four explicit metrics and machine-readable raw and summary templates.
A hardware-boundary audit separates complete packs from bare cells, then tests ten voltage, current, connector, communication, charging, thermal, mechanical and transport evidence fields.
A hardware-boundary audit separates modules, development platforms and complete systems, then tests TOPS, power, software and lifecycle disclosures.
A source-level audit separates 25 range observations and five error-metric boundaries, then scores 13 fields required to reproduce each selected claim.
An 18-record audit separates numeric backlash, categorical no-backlash wording, accuracy and qualitative language—then measures seven missing reproducibility fields.
An 18-record audit separates continuous, rated, manufacturer-calculated, stall-only and peak-only torque evidence—and measures the missing thermal conditions.
A range audit separates reflectivity, ambient light, probability of detection, channel scope and maximum representable range across nine revision-bounded LiDAR records.
Exact-model experiments add bounded evidence for a servo, depth camera, LiDAR, edge-compute board and battery cell without turning one result into a product-wide endorsement.
The v0.4.0 audit measures evidence states, comparison-field coverage and three official-source conflicts across five robot-hardware layers.
A manufacturer datasheet can be the best primary source and still be a manufacturer claim. RCI keeps source quality and claim status separate.
A practical boundary for actuation, perception, compute, power and the interfaces that bind them into one machine.
A stable boundary for the motor, transmission, sensing, drive, bearings, brake, housing and interfaces sold as one joint module.
How to keep thermal duration, speed, current, voltage and test conditions attached to robot actuator torque claims.
A practical separation of lost motion, angle mapping error and repeated positioning behavior.
Separate the physical link, protocol, device profile, control semantics, clock behavior and conformance evidence.
Why reflectivity, texture, illumination, target size, confidence and detection criteria belong beside every range number.
A selection model for precision, sparsity, memory, sensor I/O, latency, power and software lifecycle.
The evidence required to connect cells, pack protection, communication, charging, thermal limits and certification.
A six-layer test for mechanical, electrical, data, timing, thermal and software compatibility.
The first cross-category data-availability audit finds that populated fields still hide incompatible conditions—and that independent verification remains the largest gap.