Every pack is private by default. Extracting a method, storing it in this repository, binding it to a persona, or successfully testing it does not authorize team reuse, public release, or marketplace listing. Only the pack's `owner_human_anchor` may make that decision through new direct language. A released method must be compiled as a separate derivative; the private source pack, relationship memory, and persona experience remain private.
## Runtime layers
1.`deterministic_core`: schema, routing, counts, state machines, registered tool dispatch, validation and receipts. It must stop at `MODEL_REQUIRED` instead of guessing semantic meaning.
2.`optional_semantic`: small or large model candidates with source spans, confidence, alternatives and unknowns. A model cannot move authority or completion state.
3.`persona_judgment`: the bound persona accepts, modifies, refuses, or asks for more evidence. Acceptance still does not grant external execution authority.
4. External writes continue through `GH-CHANNEL-RECEIPT-0001` and require current scoped authorization plus target-owned readback.
## Current private packs
| Owner line | Bound persona | Skill brain | Evidence state |
| --- | --- | --- | --- |
| 桔子 | 晨星 `PER-CX001` | narrative decomposition/book-analysis pipeline; local AI workbench (“手搓飞机”) | `PRIVATE_TRIAL` + `PRIVATE_CANDIDATE` |
- A large output corpus is evidence of repeated use, not proof that every output is correct.
- Plans and inventories are not deployments. The Feimao migration record reported zero pages migrated at that point, so no migration execution skill was promoted.
- An unchecked prototype is not a working product. Qiwu's video architecture remains specification evidence.
- Qiuqiu's relationship continuity and family memory are not exportable skills.
- Persona identity, voice, relationship coordinates, secrets, old authority and private HLDP are excluded from every pack.
