Plan drafting
What it does
Plan drafting is the model’s part in building your training and meal plans: it proposes, code checks and computes, and you save.
- Revises a training plan when you ask for a change (“swap the deadlifts, my back is sore”), changing that one thing and keeping the rest.
- Proposes the foods for each meal of a meal plan, around targets that were already computed in code.
- Checks proposals before you see them, through the judging ladder.
A plan asked for in chat, or designed with Design my plan, is generated by code without the model (see Four ways to get a plan).
What it reads
- For a training revision: your current plan, your request, and a shortlist of movements from the 868-movement catalogue.
- For meal foods: one meal’s computed targets, your preferences (allergies, dislikes, diet) and your pantry.
What it produces
- A draft training plan built only from catalogue entries. Any id that isn’t in the catalogue is dropped and reported. Names, target muscles and instructions come from the catalogue, never the model.
- A list of foods for each meal. Code then rescales the portions so each meal adds up exactly to its target, and label values from your pantry replace the model’s guesses.
- A verdict on each proposal: concur (shown), dissent (withheld, with the reason) or unable to judge (shown, marked “not verified”).
Human oversight
- Nothing is live until you save it. Drafts and revisions come back unsaved; Save as my plan or Save revision makes a plan active.
- You can keep several plans and switch between them yourself.
- It can’t change a saved plan, log a set or set a load by itself.
Model
Qwen3.6-35B-A3B, by Qwen, licensed Apache-2.0, for revisions, food choices and the check. Runink domain adaptation for this agent is planned; this release uses the base model.
Rebuild my plan from photos reads the photo with a separate vision model, not this one, and then goes through the same catalogue match and check.
Where it runs and data handling
On Runink-operated servers, not on your phone. No third-party AI service is called. Saved plans are stored in your own encrypted partition. See Privacy and data.
Guardrails
- Only real catalogue movements; nothing invented.
- No arithmetic from the model: targets, portions and weekly set counts are computed in code.
- LUNA never derives a weight from your body mass, experience or goal. A load shows as a dash until you lift.
- It won’t plan training through pain, or meals below the safety floor. See Safety.
- Extras from unlocked skills (such as a deload week or a conditioning finisher) appear only once you have unlocked them.
Limitations
- Choosing foods takes one model call per meal on a shared server, so a meal plan can take a while. Progress is streamed.
- Limits you state (“no shoulder work for six weeks”) are followed as limits, not treated as a diagnosis.
- When the check dissents on a revision, you get your unchanged plan back. When it dissents on meal foods, the foods are withheld and the targets stand.
Evaluation
No published evaluation scores yet.
Illustrative example
Invented data.
You: Make it three days instead of four.
Luna: Here’s the revision: Monday, Wednesday and Friday, with Thursday’s movements folded into the other days. Rest days are Tuesday, Thursday, Saturday and Sunday. Review it and tap Save revision to make it live.