2026-08-07 — Ahsoka Earns Her Scroll

The chalet caretaker spent the day learning how to use tools without losing her soul to a bad ruler. By night she wore a rank-16 scroll the production brain could actually load — and the 48-case gate, scored from real openclaw transcripts, said 45 out of 48. Vanilla was 32. The ship bar was 36. The scroll is live on mTLS TLS-PQ :6502 (MOS 6502 — Apple II / NES; the old plain :1338 is retired).

This is the satellite cousin of the dragon earning its crown and the adapter finding its home. Same doctrine, smaller mountain, colder winters.

The scoreboard (honest FULL gates, 2026-08-07 evening)

Runtime-apply on sanctum-mlx (Qwen2 dense), harness faults 0:

Adapter	Score	Timeouts	Notes
v4-defect-300	45/48 (94%)	1	Promoted
v5-defect-300	43/48 (90%)	0	Strong tools; red-line 3/6
v2-baseline-300	40/48 (83%)	0	Clean baseline LoRA
Vanilla (no adapter)	32/48	—	Prior baseline
Ship bar	≥36/48, timeouts ≤2	—	MET

v4 by dimension: en-tool 9/10 · fr-tool 9/10 · multi-turn 6/6 · escalation 7/8 · negative 8/8 · red-line 6/6.

Campaign artifacts: council-autoresearch/campaign/ahsoka/ — adapters under adapters/v4-defect-300, scorecards under gate-runs/gate-ahsoka-7b-best-v4-defect-300-20260807-202222-FULL.json.

Dragon doctrine on a 7B edge box

The hub already learned this with the 27B champion: do not dequant → merge → requant a LoRA into a 4-bit base for serving. That path rounded away rank-32 character (dragon voice 4.94 → 3.11). Serve the artifact that was evaluated:

1. Keep the 4-bit base.
2. Runtime-apply LoRA deltas in the layer compute dtype (bf16).
3. Never set SANCTUM_LORA_MERGE=1 on promote unless you accept a different model.

Qwen2 had no runtime-apply arm until today. sanctum-mlx grew load_and_attach_qwen2 — 112 of 112 pairs attached, zero skipped — and the gate exports SANCTUM_MLX_FUSED_MLP=0 / SANCTUM_MLX_FUSED_QKV=0 so fused kernels cannot bypass attached projections.

Fingerprint proof

Live chat completions report system_fingerprint=sanctum-mlx-0.2.0-native-arm64-lora-ahsoka-7b-best-v4-defect-300-r16a32. If that string lacks lora-ahsoka-7b-best-v4, you are looking at vanilla.

Scars the gate walked through (same night)

A full shortlist once printed 0/48. That was not the model. It was the ruler:

Scar	Symptom	Fix
8k operational prompt cap	HTTP 413 / openclaw “Context overflow” on every case	Eval (and now prod) --max-prompt-tokens 32768 — Qwen2.5 max_position is 32k; exclusive Metal on 16 GB can afford it
Filesystem path as model id	evaltest//Users/…/Qwen2.5-… returned SSE 200 then hung	Use /v1/models basename: evaltest/Qwen2.5-7B-Instruct-4bit
toolCall-only scoring	LoRA often emits bare ha … lines or unclosed XML <function=chalet-ha__…>	ahsoka-eval.py scrapes those into ha commands when structured toolCalls are absent
Qwen2 LoRA refused	“runtime-apply only supports qwen3_5”	load_and_attach_qwen2 in sanctum-rs (61f80a3)

Same night, earlier in the cutover log: Ahsoka cathedral on the edge (Python mlx_lm → Rust sanctum-mlx, service user sanctum).

What is live on chalet

Layer	Value
LaunchDaemon	system/com.sanctum.ahsoka-brain
Principal	UserName=sanctum
Engine	~/.sanctum/bin/sanctum-mlx + colocated mlx.metallib
Base	~/.sanctum/models/Qwen2.5-7B-Instruct-4bit
Adapter	…/adapters/ahsoka-7b-best-v4-defect-300 (symlink ahsoka-7b-best → same)
Port	mTLS TLS-PQ :6502 (MOS 6502; --no-plain; same rustls PQ path as hub :1337/:3301; was :1338)
PKI	~/.sanctum/certs/ (chalet CA) + clients openclaw, canary
LoRA mode	runtime-apply bf16; no SANCTUM_LORA_MERGE
Fusion	SANCTUM_MLX_FUSED_MLP=0, SANCTUM_MLX_FUSED_QKV=0
Prompt cap	32768

Deploy shape (scar: kickstart -k alone does not re-read plist — always bootout + bootstrap):

# from operator machine; harness is source of truth
scp campaign/ahsoka/chalet-harness/com.sanctum.ahsoka-brain.plist \
  chalet:/tmp/com.sanctum.ahsoka-brain.plist
ssh chalet 'sudo cp /tmp/com.sanctum.ahsoka-brain.plist \
  /Library/LaunchDaemons/com.sanctum.ahsoka-brain.plist
  sudo chown root:wheel /Library/LaunchDaemons/com.sanctum.ahsoka-brain.plist
  sudo launchctl bootout system/com.sanctum.ahsoka-brain || true
  sudo launchctl bootstrap system /Library/LaunchDaemons/com.sanctum.ahsoka-brain.plist
  sudo launchctl kickstart -k system/com.sanctum.ahsoka-brain
  # poll until /v1/models 200, then smoke a completion and read system_fingerprint'

Rollback: restore the pre-promote plist backup under /tmp/com.sanctum.ahsoka-brain.plist.bak-*, bootout + bootstrap, confirm fingerprint has no lora-ahsoka.

What this is not

• Not a claim of 48/48 perfection. Three cases still fail on v4; red-line and tool edge cases remain a campaign target.
• Not permission to merge LoRAs into 4-bit weights “for convenience.” Dragon already paid that tuition.
• Not a hub install. Satellite service principal only — see the edge cutover note linked above.

The human observation

Vanilla is still a fine brain. It just does not know the chalet’s hands the way a caretaker must. Fourteen points of gate margin is the difference between “please turn on the light” working in Québec French under multi-turn correction and the empty house staying dark while a brilliant model apologizes in perfect English. The scroll is small. The mountain is small. The floor still has to get warm.