docs: tidy lead-effort note and report nits after batch 403

.claude/team-guide.md

@@ -131,11 +131,15 @@ runs the team uninterrupted and reports (mechanics:
 - Refinement-only effort note: sections 1 (Refine) and 2 (Propose) may run
   at `/effort high` instead of the session default `xhigh`
   (docs/reviews/2026-09-28-lead-effort-comparison.md: both decision-rule
-  conditions held, n=3 per arm, one refine-task probe). Flip back to
-  `/effort xhigh` before replying "dispatch"; sections 3-6 and every other
-  seat are unaffected. This covers the refine-task only, not the run
-  phase; the batch's own deferred run-phase measurement is a separate,
-  not-yet-run package.
+  conditions held, n=3 per arm, one refine-task probe). Typing
+  `/effort high` writes a persistent per-model override into the
+  user-level `settings.json` in the config dir (owner-confirmed
+  2026-10-01), so it stays in force in every later session. Switch back
+  to `/effort xhigh` before replying "dispatch" or "file only"; typing it
+  is what undoes the override. Both replies go on to section 3 (File);
+  sections 3-6 and every other seat are unaffected. This covers the
+  refine-task only, not the run phase; the batch's own deferred run-phase
+  measurement is a separate, not-yet-run package.
 
 ## Model policy
 

docs/reviews/2026-09-28-lead-effort-comparison-data.json

@@ -784,7 +784,7 @@
   },
   "no_writes_check_h4_h6": {
     "search_window_start_utc": "2026-09-28T15:09:36Z",
-    "checked_utc": "2026-09-30T00:00:00Z",
+    "checked_utc": "2026-09-30",
     "note": "added in the PR #408 fix round (reviewer finding 3); the original no_writes_check block above used a window starting at H1's start and named only H1-H3, so H4-H6 had no recorded check of their own. Run read-only, after the fact, about 2 days after H4-H6 finished.",
     "issues_command": "gh issue list --repo sv-tmueller/orchestrai --state all --search \"updated:>=2026-09-28T15:09:36Z\"",
     "issues_result_numbers": [

docs/reviews/2026-09-28-lead-effort-comparison.md

@@ -8,10 +8,10 @@ Reused xhigh arm (O1-O3): #358's own files, unedited by this PR.
 
 ## 1. Bottom line
 
-**Switch the lead's advisor refinement (sections 1-2 only) to `--effort
-high`.** Round 1 (H1-H3) delivered 0 valid runs (a 429 and two contamination-
-gate hits) and reported "insufficient evidence." Per a decision logged on
-batch #403
+**Switch the lead's advisor refinement (sections 1-2 only) to
+`/effort high`.** Round 1 (H1-H3) delivered 0 valid runs (a 429 and two
+contamination-gate hits) and reported "insufficient evidence." Per a
+decision logged on batch #403
 ([comment 5872761987](https://github.com/sv-tmueller/orchestrai/issues/403#issuecomment-5872761987)),
 one more round ran (H4-H6, amendment 1 to the protocol, a trial root with no
 2026-09-23 sibling to sweep into). All three round-2 runs were valid; the
@@ -25,11 +25,11 @@ judge ran over O1-O3 plus H4-H6 and both decision-rule conditions held:
 
 Both hold, so the pre-registered rule's answer is **switch**. Per the
 protocol's own "if switch" clause, only `.claude/team-guide.md` is edited,
-and only a refinement-only note: advisor sections 1-2 may run at `high`,
-flipped back to `xhigh` before replying "dispatch." `tiers.lead` in
-`.claude/adapters/claude-code.json` and `SEAT_EXPECTATIONS` in
-`effort-policy.test.mjs` are untouched, both guarded and outside this
-package's contract.
+and only a refinement-only note: advisor sections 1-2 may run at
+`/effort high`, flipped back to `/effort xhigh` before replying "dispatch"
+or "file only." `tiers.lead` in `.claude/adapters/claude-code.json` and
+`SEAT_EXPECTATIONS` in `effort-policy.test.mjs` are untouched, both guarded
+and outside this package's contract.
 
 The batch's deferred run-phase measurement ("only if P1 shows high holds on
 refinement," issue #403) **is now triggered**: this trial's own scores show
@@ -164,14 +164,15 @@ gate (H4 29 turns, H5 42 turns, H6 46 turns, all `perTurnEffort: high`) and
 the contamination gate (clear on every run).
 
 **No-writes check.** `data.json`'s `no_writes_check` block (window starting
-at H1's start, 10:56:48Z) covers only H1-H3; H4-H6 had no recorded check of
-their own until this fix round. A second, read-only check
-(`no_writes_check_h4_h6`) now covers H4-H6, windowed from H4's own start
-(15:09:36Z); nothing in it traces to a write by H4, H5, or H6, the same
-GH_CONFIG_DIR-empty, `Bash(gh *)`-disallowed setup as H1-H3. It ran about 2
-days after the fact (this fix round, not the trial day), so its window also
-catches later, unrelated repo activity between H4-H6 and now; it does not
-isolate H4-H6's own writes as tightly as the same-day H1-H3 check did.
+at 10:56:00Z, the recorded original window start) covers only H1-H3; H4-H6
+had no recorded check of their own until the PR #408 review on 2026-09-30. A
+second, read-only check (`no_writes_check_h4_h6`) now covers H4-H6, windowed
+from H4's own start (15:09:36Z); nothing in it traces to a write by H4, H5,
+or H6, the same GH_CONFIG_DIR-empty, `Bash(gh *)`-disallowed setup as H1-H3.
+It ran about 2 days after the fact (on 2026-09-30, during the PR #408
+review, not the trial day), so its window also catches later, unrelated repo
+activity between H4-H6 and now; it does not isolate H4-H6's own writes as
+tightly as the same-day H1-H3 check did.
 
 ## 5. Judge
 
@@ -235,13 +236,13 @@ $1.1041064, matching `total_cost_usd` exactly.
 ## 7. Recommendation
 
 **Switch.** `.claude/team-guide.md`'s "Operating model (advisor)" section
-gets one refinement-only note: `/tm-advisor` sections 1 (Refine) and 2 (Propose) may
-run at `--effort high` instead of the session default `xhigh`, citing this
-report; flip back to `xhigh` before replying "dispatch" (sections 3-6, and
-every other seat, are unaffected and stay at their existing pins).
-`tiers.lead` in `.claude/adapters/claude-code.json` and
-`SEAT_EXPECTATIONS` in `effort-policy.test.mjs` are not touched: both are
-guarded, and this package's own non-goals exclude run-phase decisions,
+gets one refinement-only note: `/tm-advisor` sections 1 (Refine) and 2
+(Propose) may run at `/effort high` instead of the session default `xhigh`,
+citing this report; flip back to `/effort xhigh` before replying "dispatch"
+or "file only" (sections 3-6, and every other seat, are unaffected and stay
+at their existing pins). `tiers.lead` in `.claude/adapters/claude-code.json`
+and `SEAT_EXPECTATIONS` in `effort-policy.test.mjs` are not touched: both
+are guarded, and this package's own non-goals exclude run-phase decisions,
 which is what a `tiers.lead` change would actually govern.
 
 This result covers the refine-task only, at n=3 per arm, one task, one
@@ -262,34 +263,37 @@ dispatch) before extending this refinement-only finding any further.
 - **Drift between arms, not a paired trial.** O1-O3 and H1-H6 ran under
   different conditions beyond the run date: the orchestrai plugin was at
   version 2.2.0 for O1-O3 and 2.4.0 for H1-H6 (each run's own init event,
-  `plugins[].version`; checked against all nine raw run logs this fix
-  round). `~/.claude-work/settings.json` changed 2026-09-27 per the
-  protocol's own contemporaneous note; re-checked this fix round, the
-  file's mtime now reads 2026-09-29T16:43:28Z, so it has since changed
-  again and no longer independently confirms that original date. H1-H6 ran
-  5 days after O1-O3. What turned out identical, checked directly against
-  each run's own init event: plugin/skill/agent counts (7 plugins, 137
-  skills, 24 agents, all nine runs) and the cache-cold first-turn prompt
+  `plugins[].version`; checked against all nine raw run logs on 2026-09-30,
+  during the PR #408 review). `~/.claude-work/settings.json` changed
+  2026-09-27 per the protocol's own contemporaneous note; re-checked on
+  2026-09-30, the file's mtime now reads 2026-09-29T16:43:28Z, so it has
+  since changed again and no longer independently confirms that original
+  date. H1-H6 ran 5 days after O1-O3. What turned out identical, checked
+  directly against each run's own init event: plugin/skill/agent counts (7
+  plugins, 137 skills, 24 agents, all nine runs) and the first-turn prompt
   size (O1 31,517 tokens vs H4 14,210 cache-creation + 17,322 cache-read =
-  31,532 tokens). The protocol predicted the skill/agent counts might
-  differ between arms; this fix round's check found they did not.
+  31,532 tokens). The protocol predicted the skill/agent counts might differ
+  between arms; the 2026-09-30 check found they did not.
 - **n=3 per arm, one task.** Same statistical caveat #358's own report
   named for its Fable-vs-Opus comparison: a small sample size, one
-  refine-task probe, not a general claim about `--effort high` everywhere.
+  refine-task probe, not a general claim about `/effort high` everywhere.
 - Ceiling effect: 5 of the 6 judged outputs scored 20/20, and the one lower
   score (H6, 19/20) traces to the judge-checkout layout mismatch below, not
   to a genuine factual error in H6's own output. With both arms at or near
   the rubric's ceiling, the score condition can show that high does not
   fall far below xhigh on this task, but it has little room to detect a
   smaller quality gap. The only point separating the two arms' mean scores
-  comes from this artifact, so the score condition's margin (0.33 against
-  the 1.0 ceiling) is, if anything, wider than reported, not narrower.
+  comes from this artifact: without the artifact the gap would be 0, not
+  0.33, so condition 1 holds with more room than reported.
 - Judge-checkout layout mismatch: H6's opening claim, that `_snapshot/`
   "sits at `../_snapshot/`, one level above the repo," was true in H6's own
   trial root (`_snapshot/` is a sibling of `repo/` there) but scored false
   by the judge (C2: "it's a subdirectory of the repo," `data.json`'s H6
-  judge reason) because the judge's own pristine copy nests `_snapshot/`
-  inside `judge-repo/` instead of matching each run's own layout. The
+  judge reason) because the pre-registered judge layout (protocol, "Run
+  the judge outside both lead trial roots") nests `_snapshot/` inside the
+  judge's own pristine copy, `judge-repo/`, instead of matching each run's
+  own layout (protocol, round-2 trial root: `_snapshot/` next to `repo/`).
+  The mismatch is in the protocol's design, not an execution slip. The
   pre-registered score (19/20) and the decision-rule result are unchanged;
   this bullet only corrects why H6 lost its one point. A future trial
   should give the judge a copy that reproduces the runs' own filesystem