Validating a trail on read: each cited section's stored hash against its current hash decides prefetch, hint or miss

Every read checks the fingerprints first

A trail with three cited sections. Pick how many changed since it was recorded. Hashes are shortened and illustrative.

None changed One changed All changed

cited section

stored hash

current hash

check

Authentication › Tokens [n33]

4be1c0

match

API › Sessions [n58]

91d7a2

match

Authentication › Refresh [n35]

c36f0e

7f41b8

changed

hintsome sections valid

Only the sections that still match are listed. The changed one is named and skipped. This is the recall result the agent reads:

ragtrail: similar question solved before (similarity 0.82, mode=hint)
  Useful nodes still valid:
    Authentication › Tokens [n33]
    API › Sessions [n58]
  Changed since: Authentication › Refresh [n35] (skipped)
Start with read("n33") before searching.

Prefetch also needs a high similarity score. Whatever the mode, the agent writes the answer from the current text, so a changed section is never quoted from the cache.