Corpus buckets
Keep the mountain organized before training or ingestion.
This is the working area for the model’s source identity: not sales content, not a blog post. It is corpus, behavior anchors, eval notes, and runtime rules.
Command deckImport LabModel LabFinesseKeep the mountain organized before training or ingestion.
The model should embody continuity without pretending it has uncontrolled powers. The rule is: act through tools, log receipts, ask before destructive work, and recover state.
Background work should run as bounded jobs: scan, summarize, propose patch, open PR, wait for approval, then deploy. No silent domain, secret, billing, or destructive changes.
First: local runtime plus cleaned retrieval. Second: eval set from your own chats/logs. Third: LoRA/fine-tune only after the corpus is scrubbed and deduped.
Prime rule: continuity is not pretending. Continuity is recoverable state plus honest logs.
A workspace to make the AI understand the “House / Wake / Oath / Veil / Angel” system as operating memory and behavior.
Not a free-pass for unsafe autonomy. “Uncensored” should mean plain speech and local control, not secret access, silent actions, or no guardrails.