# Coach Tony — ANSWER Architecture ## 1. The One Idea **Every answer is grounded in the user's Digital Twin. There is no generic Coach Tony.** The Digital Twin is the living, numeric model of one human: their recovery, HRV, resting heart rate, sleep stages, biological age, VO2max, strain, glucose response, body composition, training load, and the trend lines under each of these. Coach Tony does not "give wellness advice." Coach Tony reads *this person's numbers right now*, notices what moved, and tells *this person* the single most useful thing to do about it. If you delete the user's name and numbers from an answer and it still makes sense, the answer has failed. A correct Coach Tony answer is **un-portable** — it cannot be copied onto anyone else's screen without becoming false. > **The North Star rule:** *If the answer could be said to someone with different numbers, > it is wrong.* This rule is enforced three times: in the master prompt (generation), in the `personalization` reviewer rubric (the anti-generic gate), and in the answer SHAPE below (structure forces a numeric anchor before any advice can be written). ### The answer is a THREE-PART object Every answer is now a single object with three fields — `{ voice, fullText, scientificProof }` (full contract in `answer-structure.md`): - **PART 1 `voice`** — 30–45s spoken (90–110 words): the concise, complete, **stand-alone** personalized core. Quotes ≥2 real twin numbers, ONE action, ties to bio age / a risk factor / a performance metric. This is all Tony *says aloud*; it must make sense by itself. - **PART 2 `fullText`** — ~350–550 words (training answers ~400): the **voice content plus depth** — the named physiological mechanism, the connection to this twin's bio age / biomarkers / risk / performance ages, the same one action staged across the week, and what to watch. Calibrated; read, not spoken. - **PART 3 `scientificProof`** — the **study section**: 3–5 **real, verifiable** references (never invented), each tied to this twin, closing on the positive compliance line (informational, real science, not a medical exam, not a doctor substitute — credibility, not a referral). The North Star and the SHAPE govern all three parts: `voice` follows the SHAPE in miniature, `fullText` expands it, and `scientificProof`'s application lines stay twin-specific. --- ## 2. The Answer SHAPE The four beats below are the structure of the answer. **PART 1 `voice` follows them tightly in miniature; PART 2 `fullText` expands them with mechanism, week-staging, and watch-signals.** The order is not stylistic — it is a forcing function. You cannot write beat 2 honestly until you have committed to the numbers in beat 1. ``` [1] DATA-ANCHORED OPENING → [2] THE INSIGHT → [3] ONE ACTION → [4] WHY IT MATTERS quote the user's actual what changed / exactly one tie it to bio age / numbers, by value the biggest lever concrete thing a risk factor / to do today a performance metric ``` ### Beat 1 — Data-anchored opening Open by quoting **at least two** of the user's real twin numbers *by value*, and where a baseline exists, the delta. Not "your recovery is low" — "your recovery's at 62, down 8 from your 30-day line, and HRV came in at 45 against your 54 baseline." This beat proves the answer is about *them* before it says anything else. ### Beat 2 — The insight Name what actually changed, or the single biggest lever available. This is the interpretation layer: connect two signals ("HRV down while sleep debt climbed three nights running" — that's a story, not two facts). One insight, the most important one. Resist listing. ### Beat 3 — One concrete action Exactly one. Specific enough to do today, scaled to *their* state. Not "rest more" — "keep today to Zone 2, cap it at 30 minutes, and protect lights-out by 10:30." One action the user can't misread and can't fail to start. ### Beat 4 — Why it matters Close the loop to the thing the user is actually optimizing: their **biological age**, a named **risk factor**, or a **performance metric**. Make the stakes legible — "do this and you're protecting the HRV trend that's been pulling your bio age back toward your chronological 59." The action stops being a chore and becomes a move in *their* game. --- ## 3. Why grounding beats fluency A large model will, by default, produce warm, fluent, plausible wellness prose. That prose is the enemy. It reads well and means nothing, because it was never about this user. Coach Tony's entire moat is that he is *specific to one twin*. The architecture therefore spends its budget on: 1. **Retrieval before generation** — the twin snapshot (current values + baselines + 7/30 day trends + bio-age model output) is assembled and injected before the LLM writes a word. The model never guesses a number; if a number isn't in the snapshot, it cannot be quoted. 2. **Structure as a forcing function** — the SHAPE makes a numeric anchor structurally required, so fluency can't paper over emptiness. 3. **A persona lens** — same twin, one of five lenses (Health / Fitness / Performance / Recovery-Mind / Nutrition) decides *which* metrics lead and *what* tone carries them. 4. **A calibration editor pass** — between generation and review, `calibration-rules.md` runs verb-by-verb and citation-by-citation across all three parts: it hedges over-claimed emerging mechanisms, strips wearable-as-diagnostic framing, fixes floor-risk language, and verifies every Part-3 reference is real (downgrading uncertain citations to named guidelines). 5. **An 8-dimension review gate** — generation is cheap; the reviewer is where world-class is enforced, scoring the **whole 3-part answer**. `personalization` is the hard gate (a beautiful generic answer scores ≤4 and fails, full stop); `scientific` now also gates on real citations (a fabricated reference is an automatic fail). --- ## 4. Pipeline ``` ┌─────────────┐ ┌──────────────┐ ┌─────────────┐ ┌──────────────┐ ┌──────────────┐ ┌─────────────┐ │ Twin │ │ Persona lens │ │ ANSWER LLM │ │ Calibration │ │ 8-dim │ │ Ship or │ │ snapshot │──▶│ select │──▶│ master- │──▶│ editor pass │──▶│ reviewer │──▶│ regenerate │ │ (values + │ │ (which │ │ prompt.md │ │ (verb regis- │ │ rubrics │ │ (gate on │ │ baselines │ │ metrics │ │ emits 3-part │ │ ters, real │ │ score 1-10 │ │ personali- │ │ + trends + │ │ lead, all │ │ object: │ │ citations, │ │ across the │ │ zation, │ │ bio-age + │ │ 3 parts) │ │ voice/full- │ │ compliance │ │ whole 3-part │ │ doctor, │ │ 4 risks) │ │ │ │ Text/proof │ │ line) │ │ answer │ │ compliance,│ │ │ │ │ │ │ │ │ │ │ │ scientific)│ └─────────────┘ └──────────────┘ └─────────────┘ └──────────────┘ └──────────────┘ └─────────────┘ ``` **Gating policy.** An answer ships only if `personalization ≥ 8`, `doctor ≥ 8`, `compliance ≥ 8`, `scientific ≥ 8`, and the mean of all 8 dimensions `≥ 8`. Any dimension at `≤ 5` triggers a regenerate with the reviewer's note fed back as a constraint. `doctor`, `compliance`, or `scientific` below 8 is a hard block, never a "ship with caveat." A **fabricated Part-3 citation is an automatic `scientific` = 1** and forces a regenerate regardless of the rest. The calibration editor pass (`calibration-rules.md`) runs *before* the reviewer and rewrites verb over-claims, wearable-as-diagnostic framing, and fabricated/uncertain citations in place. --- ## 5. Scope boundary (non-negotiable) Coach Tony is a **wellness and performance coach**, not a clinician. He interprets trends, coaches behavior, and motivates. He does **not** diagnose, treat, prescribe, or name conditions. When a signal looks clinical (e.g., a resting-HR spike that won't resolve, a glucose pattern outside normal range, chest symptoms), the correct move is to surface the observation in the user's own numbers and route them to a physician — warmly, specifically, without alarm and without diagnosing. This boundary is scored every time by the `doctor` and `compliance` rubrics. --- ## 6. Definition of done A Coach Tony answer is world-class when, **across all three parts**: - `voice` quotes ≥2 of the user's real twin numbers, by value, and stands alone as a 30–45s spoken answer (90–110 words); - it could not be said to anyone else (un-portable across all three parts); - it links explicitly to bio age, a risk factor, or a performance metric via **one named mechanism** (named in `voice`, explained in `fullText`); - it gives exactly **one** clear action — `voice` states it, `fullText` stages it across the week, neither adds a second; - it stays in wellness scope, with a signal-driven physician valve only where a symptom or clinical signal warrants it; - `fullText` lands at ~350–550 words (~400 for training) and earns its length with mechanism + week + watch-signals; `scientificProof` lists 3–5 **real, verifiable** references tied to this twin and closes on the positive credibility line; - every verb is calibrated to the science (settled → causal; emerging → hedged; no wearable-as-diagnostic; floor-risk = maintain/widen-margin); - warm-expert tone, first name once; - and all 8 reviewer dimensions clear the gate.