# NEW_PATIENT_QUEUE — surgery patients status (2026-05-02)

## Doctrine

A patient enters the FrankenLLM runtime only after:
1. **Baseline** measurement on the target task set.
2. **Surgery target** described (excision / LoRA / distillation / repack).
3. **Post-surgery delta** measured on the same task set.
4. **Verifier evidence** (anchor 19/19 or task-specific gate).
5. **Pack hash** frozen and recorded.

No exception. No "speculatively wire it in" path.

## Patients on the table

| patient                          | role candidate         | status        |
|----------------------------------|------------------------|---------------|
| Qwen 2.5 0.5B Physarum-05B       | base for current organs| in production |
| Qwen 2.5 7B Physarium-7B Q4      | top brain              | in production |
| Qwen3 small (0.6B / 1.8B)        | newer 0.5-class organs | scoping       |
| Gemma 2 2B / Gemma 3 small       | alt 0.5-class organs   | scoping       |
| DeepSeek-R1-Distill-Qwen-1.5B    | reasoning organ        | scoping       |
| DeepSeek-V4-Flash 284B/13B       | autopsy archive        | benched, kept |

## Work needed per scoping patient

Same template as Physarum-05B — recorded as `PATIENT_<name>_PASS_<n>.md`:

1. **acquire** — HF safetensors, tokenizer, license check.
2. **donor pack** — repack to native `.planck` BF16 (one shot, no surgery).
3. **donor baseline** — anchor + organ-only on the role's task set
   (TRIZ 100, ARIZ 100, MBPP/HE 264, …) using the donor untouched.
   First number gives the bar.
4. **surgery candidate** — pick a target organ
   (e.g. "Gemma 2B → phys02_renderer"). Define what this organ does
   that current organs don't.
5. **QLoRA / excision / distill** — produce candidate pack.
6. **anchor + organ-only** — gate against current production.
7. **wire into separate `PHYS0?_<NAME>_PACK`** — never overwrite a
   working pack of a different organ.
8. **report** — `PATIENT_<name>_PASS_<n>.md` with all 4 artefacts.

## Why this queue is paused

* Phys05 organ farm needs to be solid first (BD6 frozen, BD7 in
  flight, BD8/BD9/BD10 = critic / wound / json_repair / cache_matcher /
  test_writer / renderer still queued).
* Each new patient costs disk + WSL RAM. Adding three new bases
  before the current organ farm is gated would explode the storage
  and confuse the conductance router.
* The router (TRACK 2) needs to learn first on a stable organ set.
  Adding bases before then poisons the BD store.

## Trigger to start

NEW_PATIENT queue starts firing **after**:
* TRACK 1 done (all 0.5B organs gated or explicitly frozen).
* TRACK 2 done (Black-Dog conductance routing live).
* TRACK 4 done (critic + wound).

At that point the organ slate is stable and a new patient can be
benched against an unmoving target.

## Honest status

Current FrankenLLM uses **two patients in production** (Qwen 0.5B base
+ Qwen 7B Q4 top-brain). Everything else is research notebooks until
the queue fires.