Fifteen years ago, I lost 100 lbs in a fairly boring manner by eating 1200 calories a day. It's creeping back. There are no low-hanging fruits — no sugarwater, no fast food, near-zero restaurant meals. Since then I've recorded every bite eaten every day, whether 10 calories of broth on a weekend fast or a 4000 calorie binge, using a food scale for everything. 54,925 food items across 5,429 days. Zero gaps.
This is not self-reported-from-memory dietary recall. The cumulative energy balance — intake minus expenditure, accumulated daily — closes to ±5 lbs over the full 15 years. A composition-aware model fitted to 25 indirect calorimetry measurements and 70 body composition scans quantifies the undercount at ~10-14%, uniform across weight loss, gain, and stable phases. The best metabolic ward studies achieve ~5% error over 2-4 weeks. Self-reported dietary recalls average 30-50% error with missing days. This dataset sustains ~10-14% error across 5,429 consecutive days.
The dataset contains hundreds of overlapping natural experiments: months of keto, weekend 36-hour fasts, low protein, high fiber, potatoes-only, waves of monotonous meals, daily ice cream, different cooking oils. Combined with simultaneous weight, body composition (BOD POD + InBody), resting metabolic rate (Cosmed indirect calorimetry), steps, sleep, body temperature, blood fatty acid panels, and 80 weekly tirzepatide injections logged from the first dose. If the answer to obesity requires a complex overlap or sequence of conditions, it may be hidden within and first discovered through data mining rather than invented by a brilliant hypothesizer. Very likely, the answer is already in the data — a few weeks spread across years where something worked, masked by the noise in daily weigh-ins.
I am mostly retired, 43, no prescriptions, no stressors, no emotional eating. No meat (35 years). No workplace contaminant exposure. Reverse-osmosis water. The worst-case common-variant obesity genotype: FTO homozygous risk at all five loci, MC4R heterozygous, UCP2 reduced thermogenesis — not easily-solved obesity. See BACKGROUND.md for the full genetic profile and health history.
The body defends a narrow expenditure band regardless of weight. At 165 lbs (20 lbs fat, 2013) and 225 lbs (82 lbs fat, 2024), derived TDEE is ~2100-2200 cal/day. The TDEE/RMR ratio drops to 1.02 during sustained restriction and recovers to 1.14 when restriction eases. My set point moves up or down only when some conditions are met, and weight loss above it is trivial; below it, fiendishly difficult.
The body defends a moving fat mass set point, and it controls how much you eat. The set point is not a fixed number — it's an exponential moving average of recent fat mass with a ~45-day half-life (every 6-7 weeks, the gap between defended weight and actual weight closes by half). After 5 months at a new weight, the set point has 87% converged. These numbers are derived from 2014-2026 only — the 2011-2013 period of aggressive willpower-driven restriction is excluded because intake was externally controlled, not responding to the set point. Including that period inflated the per-lb pressure and created a spurious asymmetry that did not replicate on natural-dynamics data.
The 90-day average of daily caloric surplus (intake minus expenditure) correlates with distance from the set point at r = -0.92. Each pound below the set point shifts mean daily intake by ~55 cal [CI: 15-40]. The set point doesn't create binges — individual binge size is constant regardless of distance. Instead, it tilts the entire distribution of daily eating: surplus days become more frequent, deficit days become shallower. It's a continuous control on average energy balance, not a trigger for discrete events. Caveat: the set-point half-life and the surplus lookback are partially interchangeable smoothings. A 2D sweep shows a broad ridge (for example HL=20d with a 45d surplus window gives r=-0.94; HL=45-50d with a 90d window gives r=-0.93). So the surplus regression supports a slow appetite-pressure timescale, but does not uniquely identify 45 days by itself.
The 45-day half-life survives an intake-free test. The primary model (Kalman filter) uses logged intake to help estimate daily fat mass. This creates a circularity concern. But when fat mass is estimated from weight observations alone (P3: interpolating between 1,693 scale readings, no intake in the model), the 50-day half-life still appears (r = -0.73 on 2014+ data vs -0.92 with the fuller model). Three measurement paths — binge frequency (binary), mean surplus (continuous), and intake-free weight interpolation — converge on the same 45-50 day timescale. That convergence matters because the mean-surplus fit alone has an identifiability ridge: it supports a slow timescale, but needs the intake-free and binge-rate checks to argue that ~45-50d is biological rather than just a convenient smoothing pair.
Two channels, different speeds. The body closes the gap on two fronts. An eating channel (~45-day timescale, ~55 cal/lb) shifts daily intake toward the set point. A metabolic rate channel (~9-day timescale, confirmed by 25 calorimetry measurements) burns ~90 cal/day more during weight loss than weight gain at the same body composition — actively assisting loss when fat mass is above the set point, passively permitting regain when below. The eating channel is stronger but slower.
The set point only adapts when weight is stable. On this subject's data alone, a simple trailing average (EMA) and a stability-gated model (SmoothLatch) both fit eating patterns at r = -0.94 — they are indistinguishable during slow weight change. But they make opposite predictions for rapid weight loss: the EMA predicts the set point follows weight down and regain after stopping treatment is minimal (+3%); the SmoothLatch predicts the set point freezes during rapid loss (FM never stabilizes long enough to trigger adaptation) and regain is substantial. Trial data breaks the tie decisively: the SmoothLatch predicts SURMOUNT-4 post-tirzepatide regain at +14.1% (published: +14.0%) and STEP-1 post-semaglutide regain at +9.3% (published: ~+10%). The EMA undershoots both by 5-10x. The set point adapts when — and only when — fat mass holds within ±3 lbs for at least 14 consecutive days. This predicts that regain depends on the speed of loss: faster loss leaves a larger un-adapted gap. Consistent with clinical observations that surgical and very-low-calorie patients regain more than gradual losers. See FINDINGS.md §Set point.
Tirzepatide suppresses appetite at -74.5 cal per unit of blood level (AX, identified from within-week injection cycle variation — the cleanest estimate, free of set point and tachyphylaxis confounds). The set point's per-lb eating pressure continues at the same rate on and off drug; the drug subtracts from the total independently. This is visible in the weekly injection cycle: day 0 intake averages 1643 cal (drug peak), rising to 2222 cal by day 5 (drug trough), a 579 cal/day swing driven purely by drug pharmacokinetics. Tachyphylaxis erodes effectiveness with a 35-week cumulative half-life. Binge-to-binge escalation drops from 31.6% to 0%.
The drug also suppresses the body's metabolic cooperation with weight loss. Direct calorimetry shows RMR 206 cal below composition-predicted on the drug — the metabolic boost that normally accelerates fat loss is pharmacologically eliminated. This effect operates on longer timescales than the injection cycle and cannot be separated from the appetite effect within-week. Weight loss is 93% fat; the net energy budget is approximately -450 cal intake reduction, +200 cal metabolic cooperation lost, net ~-250 cal/day deficit.
Using these parameters to simulate the SURMOUNT-1 trial (Jastreboff et al. NEJM 2022, n=2539, zero parameters fitted to trial data): the model predicts 15mg weight loss of -20.3% vs published -22.5% (within 2.2 percentage points). It overpredicts diabetic weight loss in SURMOUNT-2 (-21% simulated vs -15% published), consistent with the drug's appetite effect being partially diverted to glucose control. Post-discontinuation regain remains underpredicted (+1% vs +14% in SURMOUNT-4), the largest open discrepancy.
Not all restriction is equal. Long runs (≥6 days) and low-carb restriction recover best. Low-protein and high-step restriction produce the largest metabolic penalties. Potato diets show zero binges across 69 days and high TDEE/RMR ratio, but severe rebound after stopping.
What doesn't work: steps don't predict weight change at any timescale. Sleep hours, protein, fat, and fiber have no independent signal after controlling for calories, carbs, and sodium. The gravitostat hypothesis shows a weak signal in the wrong direction. The week-scale intake invariance claim was falsified.
Every finding with numbers is reproduced by a standalone script in FINDINGS.md. The gravitostat, omega-6:3 ratios, circadian misalignment, and other untested theories are also documented there.
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt| File | Rows | Description |
|---|---|---|
intake/intake_foods.csv |
54,925 | Every food item, 8 nutrients per item |
intake/intake_daily.csv |
5,429 | Daily nutrient totals |
weight/weight.csv |
1,693 | Daily fasted weight |
steps-sleep/steps.csv |
4,275 | Daily step counts |
steps-sleep/sleep.csv |
2,057 | Sleep periods |
composition/composition.csv |
70 | Body composition (FM, FFM, segmental) |
RMR/rmr.csv |
25 | Indirect calorimetry RMR |
drugs/tirzepatide.csv |
560 | Daily PK blood level + tachyphylaxis |
Raw data in intake/, weight/, composition/, RMR/, steps-sleep/, drugs/, temperature/, fatty-acids/, workout/, travel/. All extractors idempotent. Pipeline in CLAUDE.md.
I accept any question, no matter how personal, in service of our goal. You will not offend me.