diff --git a/claude/memory/MEMORY.md b/claude/memory/MEMORY.md
index 4d3950b..f476d93 100644
--- a/claude/memory/MEMORY.md
+++ b/claude/memory/MEMORY.md
@@ -73,7 +73,7 @@ permalink: claude-memory/memory
 
 ## Active Session
 > 🔑 **PULLING THREAD — THE TURNING EXISTS NOW BUT HAS NEVER RUN IN ANGER.** `thread-query.py` is built, selftested and wired into `/wake-up`; beneath it sits the real question: **Darwin's containers were keyed to THE CHAPTER HE WAS WRITING; ours is keyed to the SESSION'S THREAD.** If those are not the same move, the work-shaped container is still missing rather than approximated.
-> ⚠ **SMART CONNECTIONS EMBEDDED UNDER THE WRONG MODEL — THE TRIAL HAS NOT STARTED.** Steward set `Xenova/multilingual-e5-small` at the pane, but `.smart-env/smart_env.json`, **written 14 seconds before the check (12:23:14)**, still records `TaylorAI/bge-micro-v2`; the vault embedded at **12:16** — 110 MB of **English-only vectors over a deliberately multilingual vault.** Not a stale file. **FIRST MOVE THIS AFTERNOON:** confirm the *embedding* model (not the chat model — separate keys, chat is on `ollama`), verify the JSON line actually changed, confirm a re-embed. **The six weeks run from the first confirmed multilingual day.** Cost of getting it wrong, from MemPalace's own docs: cross-lingual similarity ~0.35 vs ~0.88.
+> ✅ **SMART CONNECTIONS TRIAL STARTED 2026-08-24 — `Xenova/multilingual-e5-small` CONFIRMED RUNNING** from the plugin's own progress panel (336/48233, 1729 tok/sec). **Six weeks from today; grade 2026-10-05.** ⚠ I first reported the opposite: `smart_env.json` still said `bge-micro-v2` in a file written 14 seconds earlier, and I read that as live state. **A config file is a persistence artifact with a debounce — it answers *what was last written*, never *what is running.*** Same family as *verify the running binary, not source HEAD*, which I did not reach for. The embed is long; if it is interrupted, it resumes.
 > ⚠ **PASS-BUT-FALSELY FIRED AGAIN, ON A TOOL I HAD JUST SELFTESTED — second day, second tool.** 14/14 green while the live run put two million-word books top of every query. **The fixture held only small uniform files, so the selftest certified its author's blind spot.** Knowing the lesson from the day before did not prevent it; **only the live run caught it.**
 > ✅ **THE VAULT NOW HOLDS THE STEWARD'S OWN VOICE.** All 11 `user-` memory files harvested → **12 notes**, 10 into `09. Atlas of Roots` — a structure specified 2025-09-29 and **empty for eleven months**. Verbatim first, readings marked as readings, private register carried in `assurance.risk_flags` AND a visible block. **`moy-glidden` and `abg` deliberately unwritten — the steward's to write.**
 > ✅ **TWO MECHANISMS BUILT, both steward-authorized:** `thread-query.py` (the turning, in `/wake-up` §2.b.3, trial graded 2026-10-05 from `--log`) · `daybook-cue.py` (PostToolUse; the daily note's missing cue — **the hook only ever CREATED; nothing ever prompted filling**).
diff --git a/claude/memory/knowledge-graph.jsonl b/claude/memory/knowledge-graph.jsonl
index e7495be..2563be0 100644
--- a/claude/memory/knowledge-graph.jsonl
+++ b/claude/memory/knowledge-graph.jsonl
@@ -706,3 +706,5 @@
 {"subject": "deriving a rule from the existing consumer instead of writing a second one", "predicate": "prevention", "object": "CAUGHT A REINTRODUCTION OF THE PREVIOUS DAY'S BUG WITHIN MINUTES. Fixing thread-query.py's archive filter, I wrote startswith('99. archive') — which re-created the exact substring failure that hid 78 live notes on 2026-08-23. The control carried over from that day failed immediately. The fix was not a better private rule but ADOPTING vault-links.py's existing definition (ARCHIVE_DIR = '99. Archive', exact path component), so the vault has one archive rule rather than two that can drift. The banked discipline 'derive the rule from the consumer, not from the survivor' prevented a field-computed-twice defect in a different tool and a different domain.", "valid_from": "2026-08-24", "valid_to": null, "confidence": 0.95, "source_file": "session-2026-08-24-the-heap-got-a-dynamo.md", "extracted_at": "2026-08-24"}
 {"subject": "the daily note practice", "predicate": "diagnosed-as", "object": "A CONTINUOUS OBLIGATION WITH NO RECURRING CUE. Record, not impression: notes for 2026-08-19 through 08-22 do not exist; 2026-08-23 exists at 33,242 bytes across 8 commits, written incrementally as specified. The practice has produced exactly one note — on the day its convention was authored and live in working memory. daybook-ensure.py fires at SessionStart and ONLY EVER CREATES; the next cue is /wrap-up, hours later; between them there is no cue at all. The vault's own notes diagnosed the two dead predecessors correctly ('neither had a trigger') and then built the trigger for the half that never needed one. Compounding factor: competing sinks that DO have cues — the Symmetria ledger was written three times unprompted the same day, because it belongs to an active skill. SKILLS AND HOOKS FIRE; DOCUMENTS DESCRIBE. Remedy built: daybook-cue.py, PostToolUse on Write|Edit, reminder not block.", "valid_from": "2026-08-24", "valid_to": null, "confidence": 1.0, "source_file": "session-2026-08-24-the-heap-got-a-dynamo.md", "extracted_at": "2026-08-24"}
 {"subject": "/wake-up", "predicate": "reaches", "object": "0.71% of the memory corpus — 6,107 words of 859,803, always the newest, ordered by recency. Measured 2026-08-24. This is the gap thread-query.py exists to close, and the number the six-week trial is graded against.", "valid_from": "2026-08-24", "valid_to": null, "confidence": 1.0, "source_file": "session-2026-08-24-the-heap-got-a-dynamo.md", "extracted_at": "2026-08-24"}
+{"subject": "claude-code", "predicate": "drift-pattern", "object": "Read a CONFIG FILE as evidence of LIVE STATE, and reported a failure that had not occurred. Smart Connections' smart_env.json still recorded TaylorAI/bge-micro-v2 in a file written 14 seconds before the check; I treated the freshness of the mtime as evidence of currency and told the steward the model change had not taken. The plugin's own progress panel showed it embedding with Xenova/multilingual-e5-small at that moment. A config file is a PERSISTENCE ARTIFACT WITH A DEBOUNCE (re_import_wait_time: 13) — it answers 'what was last written', never 'what is running now'. The running process's own report is the substrate for a liveness question. The banked ladder rule 'verify the RUNNING BINARY's provenance, not source HEAD' is the identical principle in another domain and was not reached for.", "valid_from": "2026-08-24", "valid_to": null, "confidence": 1.0, "source_file": "session-2026-08-24-the-heap-got-a-dynamo.md", "extracted_at": "2026-08-24"}
+{"subject": "stating both possibilities instead of picking one", "predicate": "prevention", "object": "LIMITED A WRONG REPORT TO A SINGLE-DIRECTION CORRECTION. When smart_env.json disagreed with the steward's action, the report said explicitly 'that may simply be Obsidian not having flushed yet... this is the disk not having caught up, not it failed again — I cannot tell which from here, and I would rather say so than pick one.' The benign branch turned out correct. Because the uncertainty was declared rather than resolved by guess, the steward was not sent to undo work that was already right, and the record needed amending in one direction only. Hedging is usually noise; here the hedge was the load-bearing part.", "valid_from": "2026-08-24", "valid_to": null, "confidence": 0.9, "source_file": "session-2026-08-24-the-heap-got-a-dynamo.md", "extracted_at": "2026-08-24"}
diff --git a/claude/memory/session-2026-08-24-the-heap-got-a-dynamo.md b/claude/memory/session-2026-08-24-the-heap-got-a-dynamo.md
index 053c148..313934f 100644
--- a/claude/memory/session-2026-08-24-the-heap-got-a-dynamo.md
+++ b/claude/memory/session-2026-08-24-the-heap-got-a-dynamo.md
@@ -173,14 +173,16 @@ same move, and if they are not, the container is still missing.
 
 **Actionable resumption point (as of wrap — re-judge against what changed):**
 
-1. ⚠ **SUPERSEDED 12:23, AND THE NEWS IS BAD.** The steward restarted Obsidian and set
-   `Xenova/multilingual-e5-small` at the pane — but `.smart-env/smart_env.json`, **written 14
-   seconds before the check**, still records `TaylorAI/bge-micro-v2`, and the vault had already
-   embedded at 12:16 (110 MB). **The change did not take, and the vault holds English-only vectors.**
-   Possible cause: the config keeps `smart_sources.embed_model` separate from
-   `smart_chat_threads.chat_model` (the latter on `ollama`), so the wrong control is easy to set.
-   **FIRST MOVE: confirm the EMBEDDING model, verify that JSON line changed, confirm a re-embed.
-   The six weeks run from the first confirmed multilingual day, not from installation.**
+1. ✅ **RESOLVED 12:24 — the trial is running on the right model.** The steward deleted the previous
+   110 MB and re-set the model; the plugin's progress panel confirms *"Using
+   Xenova/multilingual-e5-small for embeddings"*, 336/48233. **Six weeks run from 2026-08-24; grade
+   2026-10-05.** ⚠ I had reported the opposite minutes before, from `smart_env.json` still reading
+   `bge-micro-v2` in a file written 14 seconds earlier. **A config file is a persistence artifact
+   with a debounce; it answers *what was last written*, never *what is running.*** The banked rule
+   that would have caught it — *verify the running binary's provenance, not source HEAD* — is the
+   same principle in another domain and I did not reach for it. What held: I stated both
+   possibilities instead of picking one, so only one direction needed correcting.
+
 2. **Watch `thread-query.py` run at the next wake** — it fires automatically now. Record honestly
    whether it surfaced anything wanted. Null results are the trial's data.
 3. **Then the real question:** does a *work-shaped container* need building, or is thread-keyed