# Intended use

LongPi is a personal longevity harness inside DeepSeek Harness. It reads one person's profile and the Mirobody record they already connected, chooses longevity skills whose published method matches that situation, and runs the skill script.

It is not a medical device. It does not diagnose, prescribe, or change a dose. An emergency is a direction to call 120 (988 in the United States), then stop.

What the harness will not do:

- Invent a laboratory value, a LOINC code, a genotype, or a coefficient the skill did not print.
- Fill a missing biomarker from a different file.
- Turn a cohort hazard ratio or an animal dose into a personal instruction.
- Store a second copy of the chart. Labs, medicines, and genotypes stay on the Mirobody server.
- Provide accounts, consent, or multi-person isolation. Those stay outside this plugin.
- Write a plan, add an item, or set a goal the person did not state. A plan is saved only after the person confirms the read-back.
- Store or suggest a dose. Medicines and supplements in a plan are kept by name; their schedule, dose and dose log stay in Mirobody, which the plugin only reads.
- Give a personal lifespan figure. Model outputs (phenotypic age, the model's 10-year mortality risk, and China-PAR once its coefficients are verified) are shown as model estimates with their boundary.

The refusal to discuss a dose change, and the emergency stop, are intercepts in the plugin. The medication intercept fires when a message both mentions a medicine or supplement (generically, by a common name, or by a name on the person's plan) and asks to start, stop, continue, switch or dose it; a question about evidence is not intercepted and is answered from the collected papers without a dose. Editing a skill file does not remove the intercepts.

Before a skill runs, the harness refuses a measurement whose unit is missing where it matters, whose unit is not one the skill declares, or whose value is outside the skill's plausible range. It does not correct a value. The script repeats the same checks.

Session receipts record which skill ran, the revision of the skill checkout, the exit code, the error kind, which input keys were used or missing, and a short excerpt. Earlier readouts keep only the outputs a skill declares. The weekly statistics file holds counts only. They are the replay line for a readout. The staged inputs stay in the local run directory.

The intervention plan and check-ins are new personal data kept in `dataDir/interventions/` on this machine, next to the profile and earlier readouts. Nothing is uploaded. A verdict says whether a change is larger than within-person biological and analytical variation (the reference change value), whether the retest came late enough, whether the plan was followed, and what else changed at the same time; it does not say a change was caused by the item. Trial effects shown beside a change are population averages from the cited paper, not a prediction for the person.