How ANPC Precision Health works

One orchestrator owns the coaching judgement. Narrow specialist agents acquire evidence, perform bounded writes, inspect the result, and render the report. They do not independently decide what the athlete should do next.

One lead coach makes the judgement. Several helpers do small jobs, but none of them gets to invent the answer.

The quality gates

• Edit pass: checks identity, units, expected prior state, and whether a proposed write is backed by accepted intent.
• Review pass: checks evidence links, lifecycle completeness, source coverage, and whether any action is stranded between “applied” and “verified”.
• Publication pass: renders from the durable review. The HTML is evidence of the review, never its source.

The safety checks

• Before a change: make sure it is the right workout, the right number, and a change you accepted.
• After the review: make sure every conclusion points back to real evidence and every accepted action was actually checked.
• Before the report: build the page from the saved review. The pretty page is a copy, not the master record.

The abstract model

The simple version

1. The applied Intermediate — Chest plan preserves 18 movements, 39 steps, and 28 loaded sets before execution.
2. The five-session window records 74,491 lb across 422 minutes and 154 working sets.
3. The opened Aug 12 session keeps its 25 working sets and 12,856 lb of movement-level volume distinct from the plan.
4. Athlete-authored notes and Coach Review conclusions remain separate records. The reports on this demo carry the conclusions in full, and the athlete's own words only where a conclusion quotes them.
5. Only an explicitly accepted Carry-Forward Action can change future coaching state.

1. Before the workout, the Intermediate — Chest plan saved 18 movements, 39 steps, and 28 loaded sets.
2. Across five sessions, the device recorded 74,491 lb, 422 minutes, and 154 working sets.
3. For Aug 12, the system keeps the 25 completed working sets and 12,856 lb separate from what was planned.
4. Your own notes and the coach’s conclusions are different records. The reports here show the coach’s conclusions in full, and your words only where the coach quotes them back to you.
5. A suggestion for next time changes nothing until you clearly accept it.

The training figures above are the bounded owner-approved snapshot used on the demo screens. Internal identifiers, artifact hashes, file paths, live decision transport, and body photographs and measurements are absent from every surface. The four embedded reports are otherwise unredacted: their recovery evidence, heart-rate series, nutrition entries, and coaching narrative are real and published by choice.

Those workout numbers are real figures approved for this demo. Nothing here carries a private ID, a file fingerprint, a file path, a working link back to the live system, or a body photograph or measurement. The four reports you can open are otherwise the real thing: their recovery readings, heart-rate traces, nutrition entries, and the coach’s own conclusions are shown on purpose.

The lifecycle works because “planned,” “performed,” “said,” “concluded,” and “accepted” never collapse into the same record.

The system uses a different notebook for each kind of truth. A plan cannot pretend to be a completed workout, and a coach’s opinion cannot pretend to be your own words.

A number is not automatically a coaching conclusion. It must survive five distinct questions.

A number does not explain itself. Before the coach uses it, the system asks five simple questions.

Example: from completed sets to a qualification

The provider record shows the prescribed reps and load were completed. The metric catalogue fixes the working-load definition. The active progression strategy requires completion plus the movement’s quality threshold across the required number of exposures. Only when those rows agree can the review state “ready to progress.”

What the chain refuses

• A heart-rate mean does not rank which strength exercise was “hardest.” It mostly reflects continuous work and is reported descriptively.
• Device-reported velocity or power does not become a precise proximity-to-failure estimate without validated measurement and an applicable rule.
• An external paper can constrain a claim; it cannot override the athlete’s recorded plan, provider evidence, or explicit decision.

Example: deciding whether to add weight

First, the device must show that the planned reps and weight were completed. Then the system checks the agreed quality rule over enough workouts. Only when both are true may the coach say “ready to progress.”

Things the system will not pretend to know

• Average heart rate cannot tell us which strength exercise was the hardest. It is shown as background information.
• A machine’s speed or power number cannot tell us exactly how close you were to failure unless that measurement and rule have been proven useful.
• A research paper can set sensible boundaries. It cannot overrule your recorded plan, completed workout, or decision.

Coaching references are versioned source cards, not links pasted into prose after the fact. A report pins the catalogue revision it used, and retired keys remain visible so an old claim never silently points somewhere new.

The coach keeps a dated library of the research it used. Each report remembers the exact version, so an old explanation cannot quietly change when a newer paper is added.

None of this exists without the machine underneath it. Mission Control is a coaching layer over a Speediance Gym Monster 2, and it is only as good as what that device is willing to record and hand back. The GM2 keeps a real execution record — exercise identity, resistance per set, reps, timing, and left/right work on unilateral movements — and that record is what the coaching layer is built on.

Speediance’s own writing on connected ecosystems makes the point better than a testimonial would. In How Connected Ecosystems Integrate Third-Party Fitness Apps and Devices, Yang Chen argues that the value of an integration is contextual rather than cosmetic — that “integration is only as good as the weakest measurement and mapping step,” that normalising what two systems each call a “session” is the hard part, and that smart strength equipment should be judged “like training infrastructure, not just consumer electronics.” That is precisely the wager this system makes. The hub role the article describes is the one Precision Health plays here: it holds the device record, reconciles it with a wearable stream and hand-typed notes, and refuses to invent the fields nobody measured.

The article is also candid about what the data cannot prove — completion, resistance, reps, volume and heart-rate context are trustworthy; calorie estimates and automated movement-quality scores are not. That boundary is enforced in code here rather than left as advice, and it is a good deal easier to enforce when the manufacturer says the same thing in public. Credit where it is due: a device that records honestly, and an ecosystem open enough to read from, is what made a governed coaching system possible in a spare room.

None of this works without the gym machine itself. Everything here sits on top of a Speediance Gym Monster 2, and it can only be as good as what that machine records. It saves what you actually did — which exercise, how much weight on each set, how many reps, how long, and each side separately on one-arm or one-leg moves — and that record is what everything else here is built on.

Speediance writes about this themselves. In How Connected Ecosystems Integrate Third-Party Fitness Apps and Devices, Yang Chen makes the case that connecting gadgets together is only worth it when the shared information actually helps your training, and that the whole chain is only as strong as its weakest measurement. That is the same bet this project made: keep the machine’s record, add your own words and a watch, and do not make up the parts nobody measured.

The article is honest about the limits too — reps, weight and completion can be trusted, while calorie counts and automatic “form scores” cannot. This system enforces that rule instead of just agreeing with it. Credit where it is due: an honest machine with an open ecosystem is what made a home gym system like this possible at all.

The paradox is the point, and it does not resolve. Every plank in this page has been replaced at least once; the argument about whether that makes it a different page is two thousand years old and no closer to settled.

What the count does not establish

• It counts commits, not authorship. A commit is a unit of change, not a unit of human effort, and the two do not convert.
• A large share of those commits were themselves AI-authored. They say so: each carries a co-author trailer naming the model that wrote it.
• No rule exists here that turns a revision count into a proportion of human contribution, so none is asserted. Section 04 refuses claims for exactly this reason, and the refusal applies to the page as much as to the training data.
• The figure is a floor taken on the date stamped above, not a live tally. It is rounded down and refreshed occasionally, so it runs behind the real count and never ahead of it.

The smallest safe conclusion is the first sentence of this section: drafted by a machine, revised by a person, many times. This page does not claim the larger one.

There is an old puzzle about a ship. If you replace its planks one at a time until none of the first ones are left, is it still the same ship? People have argued about it for two thousand years and never agreed. So this page will not pretend to settle its own version of it.

What the number does not tell you

• It counts changes, not who made them. A change can be one word or one thousand.
• Plenty of those changes were made by the AI too, and each one is signed to say so.
• So the number cannot be turned into “this much was written by a person”. Nobody here has a rule for that, so nobody here is claiming it.
• It is a “more than” number, counted on the date above. It is rounded down on purpose, so the real figure is always a bit higher than the one printed.

The honest answer is the first sentence: a computer wrote the first draft, a person went back over it, and that happened many times. Anything more than that would be guessing.