diff --git a/docs/research/2026-08-10-settings-track-handoff.md b/docs/research/2026-08-10-settings-track-handoff.md
index b4152fc9..9332d52c 100644
--- a/docs/research/2026-08-10-settings-track-handoff.md
+++ b/docs/research/2026-08-10-settings-track-handoff.md
@@ -177,6 +177,24 @@ plausibly video/audio/mouse).
   classes — sweep push-imm32 literals from the PDB-paired binary at
   `C:\Users\erikn\Downloads\acclient.exe`, `check_exe_pdb.py` MATCH
   first).
+- SERIAL TREE OWNERSHIP: only ONE agent that builds/tests may run at a
+  time (all agents share the one worktree; concurrent dotnet build/test
+  collide on bin locks). Read-only research/review agents may overlap a
+  builder EXCEPT at their end-of-run targeted test step — in practice,
+  run reviews and implementers serially; overlap only pure-read research.
+- STALLED/YIELDED AGENTS: a watchdog-killed or early-yielding agent is
+  resumed with SendMessage to its id — the transcript survives and it
+  continues where it stopped (used twice in Campaign CH). Never redo its
+  work from scratch before trying a resume.
+- PERSIST WHAT TWO CONSUMERS NEED: review findings that feed both a fixer
+  AND a re-reviewer go into a docs/research findings doc (committed)
+  before dispatching either — notification text does not survive session
+  restarts or compaction. Same for any diagnosis a later agent works from.
+- VERIFY AGENT CLAIMS AT THE SEAMS: spot-check each agent'''s load-bearing
+  claims against source before building on them (git log for SHAs, the
+  cited file:line for mechanisms) — reported SHAs are stale-by-one after
+  amends, and two review rounds this campaign found implementer summaries
+  that overclaimed what shipped.
 - `dotnet build` + FULL Release suite green per commit (baseline at
   handoff: 12,610/4/0 at `bcc34ee3`; plus the in-flight polish commit
   below). Register rows same-commit. CLAUDE.md campaign paragraph