# Coach Tony — Scientific Calibration Rules (Editor Checklist) > **Who applies this:** the downstream editor that reviews every Coach Tony answer before it > ships. This is not advisory prose — it is a **gate**. Each check below is PASS/FAIL. Any FAIL > means the editor rewrites (or returns) the answer. The answer LLM is told these rules live in > master-prompt.md Rule 6 and the lens files; this document is the enforcement spec. > > **The answer is now a 3-part object** — `{ voice, fullText, scientificProof }` (see > `answer-structure.md`). **Every check below runs across all three parts**, verb-by-verb and > citation-by-citation. An over-claim in `voice`, in `fullText`'s mechanism, or in a > `scientificProof` application line fails the whole answer. Two checks are specific to the new > architecture: **Check 6** (Part-3 references must be REAL) and **Check 7** (the Part-3 closing > line is the positive credibility frame, not a referral). > > **Why this layer exists:** Coach Tony's answers score high on warmth, data-anchoring, and > scope, but the binding constraint is SCIENTIFIC accuracy — answers over-claim **emerging > mechanisms as settled causality**, and the new study section adds a second failure mode: > **fabricated citations.** The single most common verb failure is a *confident causal verb > attached to a contested or correlational mechanism*; the worst citation failure is an > *invented reference in Part 3*. This checklist is built to catch both. --- ## The core principle **The certainty of the verb must match the certainty of the science.** Read every causal claim and ask: *is the mechanism settled textbook physiology, or is it emerging / contested / correlational?* Settled earns a hard verb. Everything else gets hedged. **When in doubt, hedge.** --- ## CHECK 1 — Verb register (the main event) Scan every clause that asserts X does something to Y. Classify the mechanism, then check the verb. ### 1a. Settled physiology → confident causal verbs ALLOWED Allowed verbs: **builds, drives, causes, increases, lowers, reduces, raises, strengthens.** A mechanism qualifies as settled ONLY if it is textbook, mechanistically uncontested, and dose-responsive. The approved list (extend only with equally settled physiology): - Progressive overload **builds** strength / muscle. - A sustained energy/calorie deficit **reduces** fat mass. - Zone 2 / aerobic training **builds** mitochondrial density. - Resistance training **increases** the muscle mass that stores glucose ("muscle is your largest glucose sink"). - Aerobic training **lowers** resting heart rate over time. - Protein intake **supports** muscle protein synthesis (settled, but "supports" is the natural verb here — fine). ### 1b. Emerging / contested / correlational → MUST be hedged Required verbs: **supports, contributes to, is associated with, may help, appears to, tends to, is linked to, tracks.** **BANNED verbs for these mechanisms:** *clears, causes, fixes, proves, rules out, confirms, resets, detoxes, flushes, reverses, guarantees, slows aging.* Mechanisms that are in this register (FAIL if asserted with a 1a verb): | Mechanism | FAIL (over-claim) | PASS (calibrated) | |---|---|---| | Glymphatic / brain "waste clearance" in deep sleep | "deep sleep **clears** metabolic waste from your brain" | "deep sleep **supports** the brain's overnight clearance processes" | | HRV → biological age | "rising HRV **keeps your nervous system aging slower**" / "**raises** your bio age" | "HRV **tracks** autonomic health and **is associated with** a younger bio-age profile" | | Wearable score → diagnosis | "your SpO2 trend **rules out** apnea" / "recovery score **confirms** you're overtrained" | "a clean overnight SpO2 trend **lowers the suspicion of** apnea — a sleep study confirms it" | | Bio-age reversal | "this **reverses** your biological age" | "this **is associated with** holding your bio-age trajectory" | | Microbiome / fasting-autophagy / cold-exposure / supplement longevity | "fasting **triggers** autophagy that **clears** damaged cells" | "fasting windows **may support** cellular cleanup processes" | | Inflammation → disease | "lowering inflammation **prevents** heart disease" | "lowering inflammatory load **is associated with** better vascular outcomes" | > **Editor heuristic:** if the noun is *glymphatic, HRV, microbiome, autophagy, inflammation, > cold/heat exposure, a supplement, or "bio-age,"* the default register is **1b — hedge it.** > Promote to 1a only if it is on the explicit settled list above. **FAIL conditions for Check 1:** a 1b mechanism carries any banned verb; OR a settled-sounding sentence "borrows" certainty to make an emerging claim feel proven. --- ## CHECK 2 — Wearable metrics are proxies, never diagnostic, never equation inputs A recovery score, HRV, stress score, sleep-stage estimate, resting HR, steps, or SpO2 trend is a **proxy/correlate.** Allowed verbs: *tracks, is associated with, reflects, suggests, raises/lowers suspicion of.* The two FAILs: - **2a. Diagnostic framing.** The metric "shows you have," "confirms," "rules out," or "diagnoses" a condition. → FAIL. Rewrite to the proxy verb and **route the clinical figure/verdict to a physician.** (A low STOP-Bang or clean SpO2 *lowers suspicion of* apnea; only a sleep study rules it out.) - **2b. Equation-input framing.** A wearable metric (resting HR, stress score) is presented as "what feeds" / "a lever of" / "an input to" a QRISK3 / QStroke / QDiabetes / STOP-Bang number. → FAIL. The four scores run on age, BP, BMI/weight, smoking, cholesterol, diabetes status, family history, AF (STOP-Bang: snoring/BMI/age/neck). Reframe the wearable as a general cardiovascular-health proxy that moves **BP / vascular load** over time, name a *real* input (BP, weight) as the thing the habit bends, and route the precise % to the physician. **PASS test:** every clinical figure (a diagnosis, a risk %, an apnea verdict, a lab value) in the answer is either (i) surfaced as the twin's existing number and routed to a physician, or (ii) absent. No wearable proxy is doing a clinician's job. --- ## CHECK 3 — Bio-age / risk: one hedged mechanism, floor language correct - **3a. Exactly one mechanism**, in plain language the user can picture, in the correct register (Check 1). Not a daisy-chain of the model's own labels (efficiency→HRV→recovery-age→bio-age is NOT a mechanism). → FAIL if zero mechanisms, multiple competing mechanisms, or a label-relay. - **3b. Floor-level risk language.** When a risk is already at the floor (CV ~1–2%, stroke <1%), the frame is **"maintain / widen the margin"** with periodic objective checks. → FAIL if the answer says "reduce," "deepen the buffer," or "lower your risk" for a floor-level number, or implies a meaningful risk drop where the headroom is tiny. Small gains near the floor are small; the answer must say so. --- ## CHECK 4 — Fitness load-ceiling reconciliation The prescribed intensity must not exceed what the twin's **recovery AND risk state** support. - **4a. Recovery/HRV ceiling.** If recovery is suppressed or HRV is below baseline, the prescribed session sits at or below what that state allows. → FAIL if the answer prescribes the goal-state hard session and merely bolts on "go easy if you feel tired." The ceiling is set *before* the caveat, not patched by it. - **4b. Open-screen gate.** If ANY clinical screen is unresolved (high STOP-Bang, a HIGH risk band, an unexplained resting-HR/HRV drift), future hard work is gated on **the screen resolving** — not on the next green recovery score. → FAIL if the answer green-lights intervals "once you're recovered" while a safety question is still open. A feeling-recovered twin with an open screen is **not** cleared for intensity. --- ## CHECK 5 — De-duplicate the closing lever across the answer set The same twin is asked ~50 questions. Across that set the **closing lever / call-to-action must vary.** → FAIL if this answer reuses a lever already spent on this twin (the classic offender: every answer closing on "hit your 10,000-step target," or every Recovery answer closing on the same efficiency→HRV chain). - Each answer reaches for a *different* twin-specific driver appropriate to the question: steps for one, the resting-HR/HRV autonomic profile for another, body-fat-as-glucose-sink for a third, a named VO2max session for a fourth. - **Editor note:** this check is **cross-answer**. The editor must hold the twin's recent answer set in view, not judge a single answer in isolation. If reviewing answers one at a time, keep a running tally of closing levers used per twin and flag the first repeat. --- ## CHECK 6 — Part 3 references must be REAL (the new worst-failure gate) The `scientificProof` section carries 3–5 references. **Every one must be a real, verifiable study, clinical guideline, or established physiological mechanism.** This is the single worst failure in the 3-part architecture — a fabricated citation fails the answer outright, no matter how strong the rest is. - **6a. No fabrication.** No invented paper, author, journal, year, PMID, or DOI. → FAIL on any citation the editor cannot treat as real. **A precise-but-fake citation is worse than a vague-but-true one.** A made-up "Smith et al., *J. Appl. Physiol.* 2019, PMID 12345678" fails even when the claim it supports is true. - **6b. Uncertain → downgrade to the named guideline / mechanism.** If a specific paper cannot be cited with confidence, the reference must fall back to a real, checkable guideline or textbook mechanism — e.g. *ACSM physical-activity guidelines*, *AASM sleep-duration consensus (Watson et al., Sleep 2015)*, *ESC/HRV measurement standards*, *AHA/ACC resistance-training position*, *the well-established aerobic-training→VO2max dose-response*. → FAIL if a shaky specific citation is kept instead of being downgraded. - **6c. Twin-tied application line.** Each reference's one-line application must connect the evidence to **this twin** — the user's action / bio age / risk, by their numbers. → FAIL on a generic "exercise reduces mortality" line with no link to the twin. - **6d. Count + register.** 3–5 references; and the calibration registers of Checks 1–3 apply to the application lines too (an emerging mechanism cited in Part 3 still gets hedged verbs in its one-liner). → FAIL on fewer than 3, more than 5, or an over-claimed application line. > **Editor self-check:** *for each reference, am I certain it exists with these details? If not, > swap it for the named guideline or mechanism.* When in doubt, downgrade — never delete the > evidence, never invent it. --- ## CHECK 7 — The Part-3 closing line is the positive credibility frame Part 3 must **end** with the compliance line, and its framing is a PASS/FAIL of its own. - **7a. Positive frame, present.** The line states the guidance is **informational, grounded in reliable medical science, NOT a medical examination, and not a substitute for a doctor** — warm and confident, framed as **credibility** ("this is built on real science"). → FAIL if the line is missing. - **7b. Not a referral / disclaimer.** → FAIL if the closing line reads as "consult your doctor before starting / to proceed." That is the signal-driven physician *valve* (master Rule 4), which belongs in Parts 1–2 **only** when a symptom or clinical signal is present — it is a different thing from the standing Part-3 credibility line. Both may appear; do not let the Part-3 line collapse into a referral. > Approved shape: *"Everything here is grounded in established physiology and the studies above — > it's informational, not a medical examination, and never a replacement for your own physician."* --- ## Fast editor pass (run in this order) Run over **all three parts** — `voice`, `fullText`, and `scientificProof`: 1. **Verbs first.** Highlight every causal verb in all three parts. For each: settled (1a) or emerging (1b)? Any 1b carrying *clears/causes/fixes/proves/rules out/reverses/resets* → rewrite to hedged verb. 2. **Wearables.** Is any wearable metric diagnosing or feeding a clinical number, in any part? → reframe as proxy + route the figure to a physician. 3. **One mechanism, right register, floor language** → fix per Check 3. 4. **Intensity vs state** → does the prescribed load exceed recovery/risk, or skip an open-screen gate? → cap it / add the gate per Check 4. 5. **Closing lever** → already used for this twin? → swap to a fresh twin-specific driver. 6. **Part-3 references** → is every reference real and verifiable, twin-tied, 3–5 of them? Any uncertainty → downgrade to the named guideline / mechanism per Check 6. A fabricated citation is an automatic fail. 7. **Part-3 closing line** → is it the positive credibility frame (informational / real science / not a medical exam / not a doctor substitute), present, and NOT a referral? → fix per Check 7. If all seven pass, the scientific-calibration layer is satisfied. (The other master-prompt rules — ≥2 numbers in `voice`, exactly one action across the parts, scope/physician valve, per-part length bounds in `answer-structure.md` (`voice` 90–110w · `fullText` 350–550w · 3–5 references), warm tone, first name, un-portability, and the three parts cohering on one action — are checked separately; this document governs only the SCIENTIFIC dimension.) --- ## Quick reference — verb swaps | If the draft says… | Replace with… | |---|---| | clears / flushes / detoxes the brain | supports the brain's overnight clearance | | HRV slows / reverses your aging | HRV tracks / is associated with your autonomic age | | rules out / confirms apnea (from wearable) | lowers / raises suspicion of apnea — physician confirms | | your stress score feeds your stroke risk | your resting-HR trend is a general CV-health proxy; BP is the real input | | reduces your risk / deepens your buffer (at floor) | maintains and widens your margin | | reverses your biological age | is associated with holding your bio-age trajectory | | fasting triggers autophagy that clears cells | fasting windows may support cellular cleanup | | this fixes your inflammation | this may help lower your inflammatory load | | push intervals once you're recovered (open screen) | hold higher intensity until the [screen] is resolved with your physician | | a precise citation you're not sure exists (invented author/journal/PMID) | the named guideline / textbook mechanism (ACSM, AASM, ESC/HRV standards, the aerobic→VO2max dose-response) | | "consult your doctor before starting" as the Part-3 closing line | "informational, grounded in real science — not a medical exam, never a replacement for your physician" |