Record validator evidence and repair execution findings
This commit is contained in:
@@ -75,6 +75,24 @@ the shared editorial allowance explicit in the rendered request.
|
||||
|
||||
## Stage 3 — Review Evidence at the Validator Boundary
|
||||
|
||||
**Complete.** Both model validators received speaker `A`, the literal
|
||||
`second` → `Second` edit, full `second text` → `Second text` previews,
|
||||
the preceding and following segments, and the matching `Second text` glossary
|
||||
entry for the original and revised candidate. The capitalization objection
|
||||
reached the next repair request with each validator's identity. The fixture
|
||||
scripts that same objection for both validators, so it confirms feedback routing.
|
||||
The rendered meaning-reversal instructions still require opposing full
|
||||
assertions. This case revealed no missing evidence or unclear validator
|
||||
instruction. The captured items retained their supplemental context, and
|
||||
the fitting tests exercised omission of optional context under a smaller budget.
|
||||
|
||||
The native invocation reported 10 provider attempts with complete terminal
|
||||
coverage and unknown token usage. Per-call roles accounted for one proposal,
|
||||
two initial validator requests, three repair requests, and four validator
|
||||
requests across two revisions. The seven repair-related requests are already
|
||||
included in the invocation total. The fixture's last revision repeats the
|
||||
same edit and receives a scripted approval, so the counts measure execution.
|
||||
|
||||
1. Follow a rejected candidate and, where available, its revision through the
|
||||
captured proposal, validator request, decision, and next repair request.
|
||||
Compare the target speaker, literal edit, before/after text, supplied
|
||||
|
||||
Reference in New Issue
Block a user