1 · Define the decision and comparison boundary
Each dataset begins with a fixed user question, included component class, excluded variants and minimum evidence fields. RCI does not expand the boundary mid-comparison to accommodate whichever products have cleaner marketing data.
2 · Establish a stable component identity
A record requires a manufacturer, model or orderable identifier, variant boundary and a traceable first-party source. Family-level claims are not silently assigned to every variant. Revisions, regions and discontinued models remain distinguishable.
3 · Preserve the source claim
RCI stores the reported value, original unit, wording, test condition, source URL, document title or revision, source location and last-verified date before any normalization. A screenshot or cached document may support an audit, but the public citation points to the lawful upstream source whenever possible.
4 · Normalize without deleting context
Units may be converted only through a documented, reversible transformation. Values with materially different test conditions remain separate. When a definition is unclear, the value stays unnormalized and the missing condition is published.
Peak and continuous ratings, precision-specific compute, range under different reflectivity, and cycle life under different depth-of-discharge assumptions are not collapsed into one field.
5 · Assign evidence at claim level
| Code | State | Operational meaning |
|---|---|---|
| V | Independently verified | Supported by a qualified external record independent of the manufacturer claim. This claim state does not by itself mean RCI physically tested the product; the separate EV0–EV6 work level states what was actually done. |
| M | Manufacturer claimed | Directly stated by the manufacturer; useful primary evidence, but not independent verification. |
| D | RCI derived | Calculated or normalized by RCI with the source values, units and method preserved. |
| C | Conflicting | Credible sources disagree. The conflict remains visible until evidence resolves it. |
| U | Unknown | No adequate public evidence. Unknown is a valid result, not a blank to fill with inference. |
The state describes the claim, not the prestige of the source. A value can change state as stronger evidence appears. Prior evidence remains in the change history.
6 · Declare what work was actually performed
Claim state and work level answer different questions. M/V/D/C/U describes the status of a claim. EV0–EV6 describes whether RCI only read a source, checked files, recomputed public data, handled a physical specimen or obtained an external test.
| Level | Work performed | What it supports | What it does not support |
|---|---|---|---|
| EV0 | Source-reported only Upstream author or manufacturer | What the responsible source states at a fixed location and date. | That RCI checked the artifact, recomputed the result or tested hardware. |
| EV1 | RCI source verification RCI Editorial Team | Identity, revision, source location, stated condition and rights boundary were checked. | File integrity, numerical reproduction or device performance. |
| EV2 | RCI artifact-integrity audit RCI Editorial Team | Downloaded files, hashes, manifests, source trees or release inventories match stated public artifacts. | That software works on a device or that hardware meets a performance claim. |
| EV3 | RCI public-data recomputation RCI Editorial Team | Specified aggregates or structural results can be recomputed from an identified public dataset. | Unreleased experiments, model training, hardware behavior or generalization beyond the dataset. |
| EV4 | RCI physical bench test RCI Editorial Team | The identified specimen produced the published result under RCI’s disclosed apparatus, protocol and conditions. | Population-wide performance, certification, safety or replication by another organization. |
| EV5 | External independent replication Named organization independent of RCI and the vendor conclusion | A second organization reproduced a defined result with disclosed deviations and raw records. | Accredited conformity or universal performance outside the replicated protocol. |
| EV6 | Accredited laboratory test Named laboratory with applicable accreditation and scope | A specified test or conformity conclusion was produced under the named standard, scope and accreditation. | Claims outside that test method, specimen, standard or accreditation scope. |
EV1 source verification, EV2 release-integrity work and EV3 public-data recomputation are desktop evidence work. They may be independent of vendor influence, but they do not mean RCI handled or measured a device. Only EV4 and above require a physical specimen; EV5 requires a named external replication, and EV6 requires an applicable accredited-laboratory scope.
Machine-readable definitions are published at /data/methodology-evidence-work-levels.json. Organizations interested in hosted testing, external replication or applicable accredited work can review the EV4–EV6 testing partner program.
7 · Pass a publication gate
Reference records require stable identity, upstream citation, last-verified date and visible evidence state. Comparative outputs additionally require compatible definitions and conditions. Research articles distinguish reported facts, RCI analysis and inference.
Every new research audit must disclose its highest completed EV level. Titles and summaries must not use “independent audit” as shorthand for physical testing. When no specimen was handled, the visible page must say so.
8 · Recheck, supersede and correct
High-volatility product pages and software support records are checked more often than stable standards metadata. Material corrections identify what changed, why, when and which downstream pages or files were affected. See the corrections process and public change log.