Docs/03 Attribution

03 Attribution

Crediting the lessons that actually bore on the work.

The question that drives everything downstream: which served lessons actually bore on the work?

Serving is a retrieval decision; using is an outcome. Conflating them breaks two things at once — an irrelevant lesson that happened to be injected accrues a record of usefulness it never earned, and the selector is never told that what it chose went unread.

Two sources, best first:

  • The in-session reflector. The fork holds the real conversation, so it can see a principle being applied and not merely a command being run. It reports with rose used --session <id> --used <ids> --unused <ids>.
  • The digest judge. assess is shown the served lessons and the session digest. Weaker: influence on reasoning is nearly invisible in a digest of commands, so it under-credits principles. The digest now includes the agent's own reasoning for exactly this reason.

Either way the prompt is strict and defaults to false: being on-topic is not being used; being read and found irrelevant is not being used; having been served a lesson and then done the opposite means it was not used.

A lesson served but unused is scored neither success nor failure. It has three possible causes and only the first is a retrieval problem:

CauseMeaning
irrelevantrecall over-served
redundantthe agent knew it anyway
relevant and ignoredthe lesson isn't landing — a salience failure

rose status reports precision (used ÷ served).

And whether the reflector is catching anything at all#

A reflector capturing a lesson is the system working. The live session capturing one means the evidence was sitting there and the reflector walked past it, so a person had to notice. ROSE_CHILD is set in every spawned reflector, so this is observed rather than inferred: captures record by=reflector or by=session, and rose status reports the miss rate.

The metric this replaced counted captures that followed no nudge, and read a high share as the agent has outgrown the scaffolding. A session where the user asks "why did you not learn that?" and the agent then adds the lesson by hand scored perfectly on it — the worst outcome the system can produce, reported as its best.