"Correction" is now "Ideal response", and the box invites the actual answer

CLAUDE.md

@@ -1053,7 +1053,7 @@ mean two different things, stop and ask which. Do not pick the more likely readi
 proceed.
 
 This costs a round-trip and saves a wrong change to a live agent. The asymmetry is the whole
-argument: a question costs minutes, while a misread correction ships wrong content to
+argument: a question costs minutes, while a misread ideal response ships wrong content to
 customers, looks resolved on the dashboard, and is only caught if someone re-reads the
 transcript later.
 
@@ -1075,7 +1075,7 @@ blocked on, and what you would do under each reading.
 
 **THE PROSE IS THE FEEDBACK. THE FIELDS ARE A HINT, AND OFTEN A WRONG ONE.**
 
-Reviewers write the correction under the bad answer, and/or a **Proposed fix**, and click
+Reviewers write the ideal response under the bad answer, and/or a **Proposed fix**, and click
 *Mark reviewed*. They frequently do not touch the dropdowns — and they are not expected to.
 Writing "this is wrong, it should have said X" is the valuable part; turning that into
 `diagnosis` and `fix_target` is clerical work, and it is **your** job.
@@ -1261,7 +1261,7 @@ Full conventions are in `README.md`. The essentials:
   `GitHub-`, `Misc-`).
 - **`FAQ-<Domain>.md` is the exception**: one per agent corpus, authored here with no upstream
   source. It is the home of record for answers that exist nowhere else — verbal SME guidance,
-  observed behaviour, corrections an upstream owner has not yet made. Put such content
+  observed behaviour, ideal responses an upstream owner has not yet made. Put such content
   **there, not in a `Docusaurus-` or `Conf-` file**, because those are re-derived from their
   sources and the next reconciliation would silently delete it. Every entry carries
   `Source` / `Added` / `Confidence` / `Promote when`. Never add an unconfirmed claim: the

DAILY-WORKFLOW.md

@@ -81,7 +81,7 @@ everything alone rather than tidying up first.
 | Save without submitting | **Save & Publish → Save progress** |
 | Submit for review | **Save & Publish → Send my reviews in** |
 
-**You do not have to fill in the dropdowns.** Writing the correction in your own words is the
+**You do not have to fill in the dropdowns.** Writing the ideal response in your own words is the
 valuable part. The classification fields are clerical and get filled in later.
 
 ---

contributor-prompting-guide.md

@@ -146,7 +146,7 @@ specifically.
 You do this part in your **browser**, not by typing at the AI. Read the question, read the
 answer, and if the answer was wrong, write what it *should* have said.
 
-**You do not have to fill in the dropdown boxes.** Writing the correction in your own words is
+**You do not have to fill in the dropdown boxes.** Writing the ideal response in your own words is
 the valuable bit. The AI works out the rest from what you wrote.
 
 If you are unsure whether an answer is right, say so in your note rather than guessing. A

scripts/eval_batch.py

@@ -213,15 +213,15 @@ def wants_change(fm, body):
     to be approved before the batch can be sent, so it blocks the work that WAS asked for.
 
     Keyed on the prose as well as the fields, deliberately. The review form opens pre-filled as
-    "nothing wrong", and CLAUDE.md is explicit that reviewers frequently write the correction and
+    "nothing wrong", and CLAUDE.md is explicit that reviewers frequently write the ideal response and
     never touch the dropdowns - so `kb_action: none` beside a paragraph of "it should have said
     X" is a common and legitimate state. Reading only the field would drop exactly the reviews
     that matter most.
     """
     for k, v in (fm or {}).items():
         if field_asks_for_change(k, v):
             return True
-    # Written feedback: a non-empty correction under any exchange, or a proposed fix.
+    # Written feedback: a non-empty ideal response under any exchange, or a proposed fix.
     for m in re.finditer(r"<!-- review:\d+ -->\n(.*?)<!-- /review:\d+ -->", body or "", re.S):
         t = re.sub(r"\*\*Review\s*[—-]\*\*.*?(?:\n|$)", "", m.group(1), count=1)
         if t.strip():

scripts/fetch_transcripts.py

@@ -394,8 +394,13 @@ def render(slug, meta, data, deleg=None):
             L += ["**Tools called:** _not recorded for team conversations — see Delegation above_", ""]
         L += ["**Q:**", "", "> " + scrub(unescape_question(e.get("question"))).replace("\n", "\n> "), ""]
         L += ["**A:**", "", "```markdown", scrub(e.get("response") or ""), "```", ""]
-        L += [f"<!-- review:{i} -->",
-              "**Review —** _verdict:_ · _should have said:_", "",
+        # EMPTY. This block used to be seeded with
+        #   **Review -** _verdict:_ - _should have said:_
+        # which put an annotation prompt where the answer belongs, and every downstream tool
+        # then needed a rule to ignore it (see reviewtext.PLACEHOLDERS, kept for the
+        # transcripts already carrying it). The field wants the answer the agent should have
+        # given, so it starts blank and the form offers "Copy Foundry response" to seed it.
+        L += [f"<!-- review:{i} -->", "",
               f"<!-- /review:{i} -->", ""]
     L += ["---", "", "## Proposed fix", "",
           "_What should change so this answer is right next time? For an instructions or"