Inside Learn
Research and evidence you can inspect
Useful setup advice shows where it came from and what was actually checked. Use this page to read the evidence labels on our guides and to plan your own reproducible records.
Our existing lab guides are source-verified, not lab-tested. Source dates appear on individual articles. A source check does not establish compatibility, performance, or a supported PebbleRack configuration.
Read the evidence label.
- Published source: an attributed statement from a provider, documentation page, or paper. Its methods and limits belong with the claim.
- Source-verified: instructions checked against the linked documentation on the date shown. The full procedure has not been run on PebbleRack hardware.
- Estimated: a calculation with stated assumptions. It is not an observed speed, memory, cost, or compatibility result.
- Tested: an observation tied to a named configuration, method, date, and evidence. We do not apply this label to hardware based on documentation alone.
Read critically
Connect a report to your task.
For each report, identify the original question, method, configuration, and limitations before applying its conclusion to your equipment.
- 01Read the model cardTrace evaluations to their original source and distinguish author-reported results from independent reproduction.
- 02Identify the repository and revisionKeep the original model and a quantized derivative separate in your records.
- 03Read the route comparisonsConsider practical alternatives alongside local AI before making a purchase.
- 04Browse source-checked setup guidesIndividual articles show their source-check date and link to original documentation.
Reproducible records
What a useful test record needs.
Capture exact hardware, operating system, drivers, runtime version, model file and revision, context length, task inputs, concurrency, and date. Use non-sensitive inputs and record repeated runs, errors, output checks, and recovery steps.
If you measure latency, memory, power, or generation rate, name the measurement method and conditions. Keep first-run setup and warm runs distinct. Comparing different configurations requires an explanation of what changed.
Assistants can help organize records, but generated tables and suggested commands are not observations.