how the compatibility lab reaches a verdict

the four verdicts

the evaluator keeps axes separate

A single percentage would hide the most important part of the answer. Super Evil Robots evaluates compute, memory, media acceleration, storage capacity and topology, network, expansion, runtime support, power/thermals, and operational risk independently. The overall verdict is the most restrictive required axis after explicit conditions are applied.

Storage is deliberately split from compute. A one-litre PC may be a good application node and a poor place for the only copy of a media or photo library. A result can therefore pass compute while requiring external NAS/DAS storage and a separate backup plan.

evidence grades

The pilot currently contains official-source modeling and explicit inference. The accepted Grade A ledger currently contains 0 runs. The empty grade is a feature: the interface cannot pretend the lab measured something it did not.

versions and effective dates

A claim carries a stable ID, source, retrieval or test date, evidence grade, configuration or version scope, and—where known—effective start and end dates. A new app release does not overwrite history. It creates an impact review for every affected requirement and verdict.

Official maximums and observed working maximums are separate fields. Likewise, “supported by the vendor,” “works in a community report,” and “reproduced in this lab” are never collapsed into one label.

benchmark protocol

A publishable run needs a machine-readable hardware manifest; firmware, OS, kernel, driver, runtime, and app versions; configuration with secrets removed; fixture checksum; cold/warm distinction; ambient and measurement conditions; at least three repetitions for variable tests; and median, range, failure count, and known limitations.

Tests map to decisions: concurrent transcodes, a fixed import corpus, a defined Home Assistant integration load, a declared Proxmox guest mix, or a combined hostile peak. Synthetic or openly licensed fixtures are used; private photos, camera feeds, machine names, and home-automation data are not published.

inspect the current benchmark queue →

AI and agent-use disclosure

Super Evil Robots uses AI coding and maintenance agents to help write and test software, draft candidate structured records, monitor approved source URLs, calculate deterministic impacts, and prepare review queues. Agent output is not a source, physical observation, benchmark result, evidence grade, or human approval.

Automated workers may detect official-source changes, store approved snapshots, propose extracted facts, identify affected verdicts, schedule repeatable tests, validate variance, rebuild pages from approved data, and open review tasks. They may not approve source rights, resolve contradictions, change a public verdict, answer a material correction, or alter the test method without human review.

Candidate changes remain held until their machine checks and required human decisions are recorded. Physical Grade A claims require a real run with hash-pinned artifacts; generated prose can never satisfy that gate.

page graduation

The checker can calculate many states, but very few deserve search pages. An indexable page needs a distinct decision, stable identity, material unique evidence, maintained sources, a useful answer without generated filler, internal links, and a last-reviewed date. Everything else remains an interactive, non-indexed result.

A page is also accountable for the rules behind it. Every check the evaluator can apply is listed in the rule audit above, with its own address, so a disputed result can be traced to the exact rule that produced it.

See a claim that violates this method? Use the correction protocol.

generated lifecycle and change-history index

Nothing in the graph has been revised yet, because nothing in it is old yet. The current graph carries 108 lifecycle-bearing records and 108 explicit history events, of which 0 changed a published claim after it was created. That is a starting line, not a record of stability.

generated 2026-07-26 from record.lifecycle.changeControl.history