How we calculate everything

• Deterministic first. Every number is computed by a published formula in your browser or on our server. The AI Analyst only interprets numbers that were calculated this way — it never generates them.
• Calculation vs estimation. Some outputs are exact arithmetic on your inputs (FFMI from a given body fat, plate loading). Most are estimates built on population research. We label the difference and show ranges wherever the uncertainty is material.
• No false precision. Results are rounded to what the method can support, and where error margins are known they are stated next to the result.
• For healthy adults, for education. Nothing here is medical advice or a diagnosis.

Current position. Lean mass = weight × (1 − body fat). Height-normalised FFMI = lean mass ÷ height² + 6.1 × (1.8 − height), after Kouri et al. (1995).

Ceiling. Kouri et al. found drug-free male lifters rarely exceeded a normalised FFMI of ~25, with trained men averaging ~22. We model a typical male ceiling of 24 (range 22.5–25.5). For women, where the research base is much smaller, we use 20 (range 18.5–21.5). If you are already above a bound, the ceiling is raised to meet you — the model never tells anyone their limit is below where they already are.

Rate. Each year of consistent training closes a fraction k of the remaining gap to the ceiling: lean(t) = ceiling − (ceiling − lean₀)·e^(−λt), with λ = −ln(1 − k). k starts at 0.60 for beginners and falls to 0.45 (1–3 years), 0.35 (3–6) and 0.28 (6+). It is multiplied by 0.7 for women and by an age factor (1.0 under 30, 0.9 in the 30s, 0.8 in the 40s, 0.7 in the 50s, 0.6 from 60). The ranges use 0.7× and 1.25× k with the low and high ceilings. This reproduces the widely used heuristics of Lyle McDonald and Alan Aragon (≈9–11 kg in year one for men, roughly halving each year).

Frame model (Pro). Casey Butt (2009) fitted maximum lean body mass and maximum circumferences of drug-free champions to height, wrist and ankle circumference. We apply it for men only, because it was derived from male athletes, and present it as an upper bound.

Limitations. Population estimates, not individual predictions. The largest source of error is the body-fat input (±3–5 points for home methods). Research on women and older lifters is limited.

• Body fat (U.S. Navy) — Hodgdon & Beckett (1984). Typical error ±3–4 points versus DEXA. Output clamped to 2–60%.
• BMI-based body fat — Deurenberg et al. (1991). Shown only as a comparison; it overestimates body fat in muscular people.
• Lean body mass — from body fat when provided; otherwise Boer (1984), with James (1976) and Hume (1966) for comparison.
• FFMI bands — descriptive bands derived from Kouri et al. and later FFMI surveys; they describe, they do not diagnose.
• Waist-to-height ratio — boundaries at 0.4, 0.5 and 0.6 following Ashwell's work on central adiposity.
• Target body fat — goal weight = lean mass ÷ (1 − target). A second figure assumes ~20% of weight lost is lean tissue.

• BMR — Mifflin–St Jeor (1990) by default; Katch–McArdle when body fat is supplied; revised Harris–Benedict (Roza & Shizgal, 1984) for comparison.
• TDEE — BMR × activity factor (1.2 / 1.375 / 1.55 / 1.725 / 1.9). Individual error is roughly ±10%; we show that range and explain how to correct it with scale data.
• Bulking — monthly gain targets of 1–1.5% (beginner), 0.5–1% (intermediate) and 0.25–0.5% (advanced) of bodyweight, after Alan Aragon. Converted with ~5,500 kcal per kg of mixed tissue.
• Cutting — 0.5–1.0% of bodyweight per week (Helms et al., 2014), slower when leaner; ~7,700 kcal per kg; floors of 1,500 kcal (men) and 1,200 kcal (women); time-to-target simulated week by week.
• Protein — 1.6 g/kg (point estimate) to 2.2 g/kg (upper CI) for building muscle (Morton et al., 2018); 2.3–3.1 g/kg of lean mass in a deficit for lean lifters (Helms et al., 2014); ~0.4 g/kg per meal (Schoenfeld & Aragon, 2018); adjusted bodyweight above ~25% (men) / ~32% (women) body fat.
• Macros — protein first, fat 15–45% of calories with a 0.5 g/kg floor, carbohydrate the remainder.

• One-rep max — mean of Epley (1985) and Brzycki (1993), Lombardi shown for comparison. Reliability labelled by rep count.
• RPE — reps-to-failure percentages from the chart popularised by Mike Tuchscherer; RPE is treated as reps in reserve, with linear interpolation for half-points.
• Plates — greedy loading from the largest selected plate; exact remainder reported.
• Training volume — zones informed by Schoenfeld et al. (2017) and Pelland et al. (2024): <4 maintenance, 4–9 minimum effective, ~10–20 productive (upper bound rises with experience), then diminishing returns. ~10 hard sets per muscle per session flagged as a practical ceiling.
• Progressive overload — double progression with optional deloads every 6 weeks.

• Trends — least-squares slope over 28 and 90-day windows; no rate is reported from less than a week of data. Weight is smoothed with an exponential moving average (α = 0.3).
• Composition — lean and fat mass from each check-in that includes body fat; the change between first and latest is classified (recomposition, lean gain, fat loss…). Because body-fat estimates carry several points of error, we read direction, not decimals.
• Strength — change in estimated 1RM per lift across check-ins.

The Analyst is a large language model (Google Gemini 2.5 Flash via the Vercel AI Gateway) that receives a factual brief built from your saved check-ins and results — every number in it computed by the methods above. It is instructed to answer only from that brief, to say when the data cannot answer a question, and to avoid medical, drug or supplement advice. Conversations are not stored. AI can still be wrong: treat its answers as a knowledgeable second opinion, not a prescription.