Codeuteronomy
The source is authoritative; the agent's guess is not
A generated value carries no evidence. Where a specification, configuration, or measurement exists, it governs. A plausible substitute offered in its place is a defect.
An agent is fluent, and at a glance fluency is indistinguishable from knowledge. Ask for a threshold, a color, a version, an API's behavior, and it returns a confident, reasonable value whether or not it looked. The guessed value and the read value are formatted identically; you can't tell which one you got by looking. The guess is usually close, too. Close where you needed exact is a bug in the costume of an answer.
Route the agent to the authoritative source and make it derive from that. If a spec states the value, the spec wins. If a config holds it, read the config. If nothing records it, measure or ask; never split the difference, never reuse a number from somewhere else. When the answer matters, “that sounds right” is the cue to demand the citation. The agent won't volunteer that it never checked.
How to practice it
Point at the path.
“Read the config and use that value” instead of hoping it did.
Make it quote the line it used.
Task a subagent to extract the literal values with their sources. If it can't paste the source, it didn't read it.
Send back plausible values offered without a source.
Looking right is not a citation.
Diagnose from sources too.
“Read the actual definition” finds the bug; “nudge it until it looks right” hides it.
Encode “cite, don't guess”
for the domains where it keeps guessing.