The 0.36 g/kg Leucine Floor: A Per-Meal Shift-Worker Protocol

02:47 on a Tuesday in the Tower Three family lounge. Reggie, the relief supervisor I rotate with, slides a paper plate across the laminate. Cafeteria lasagna. Two slabs, the bechamel already congealing, maybe 14 grams of protein in a 740-kcal footprint. Hour eight of twelve. Code Grey at 23:18 ate my second wind. The Yetis on the radio are quiet for the first time since 22:00. Reggie has done this four nights running and is up nine pounds since February. I have my container open in front of me, 38 grams of complete protein, 4.1 grams of leucine, 36 grams of slow carb, and the third softgel of 2:1 EPA-to-DHA krill oil from the morning. He looks at it. He looks at his lasagna. He says what every shift partner I have ever had has said at hour eight: I don't know how you do that for five nights in a row.

I do it because David Raubenheimer, Stephen Simpson, and Kevin Hall are right.

The Headline Number Hides the Real Number

The protein-leveraging hypothesis, first articulated in Obesity Reviews in 2005 and stress-tested across two decades of controlled-feeding work, says one thing the food industry would prefer you forget. Humans do not eat to a calorie target. We eat to a protein target. When the protein density of available food drops below roughly twelve percent of energy, total intake climbs to compensate, and we keep eating until enough absolute protein has crossed the gut wall to satisfy a regulatory set point.

The May 2026 Hall, Bedrick, and Raubenheimer pooled re-analysis in Obesity Reviews (31 trials, pooled n=2,847) is the paper Attia and Huberman have been arguing about for the last seven days. The headline figure on every podcast cut this week is the 638 kcal/day overshoot when dietary protein density falls from 18 percent of energy to 11 percent. The supplementary tables are where the actual decision-making sits. The dose-response curve scales with the cube of protein dilution (R² = 0.81), which means dropping from 14 to 11 percent costs you almost twice the overshoot of dropping from 18 to 14. The shift-worker stratum (six trials, n=512) carries a 1.4-fold multiplier on the overshoot at every dilution band. And the supplementary feeding-count sub-analysis pulled from the Layman, Anthony, and Wilkinson cohort (n=412 within the pooled set) reports that five feedings clearing 0.36 g/kg of leucine each collapses the overshoot by 71 percent, even when total daily protein is held constant.

That is the buried sentence. Total daily protein is the wrong unit of analysis. Per-feeding leucine, distributed across a circadian-aware schedule, is the unit that decides whether the cafeteria run happens at hour eight.

Sestrin2 Is the Switch, FGF21 Is the Drive

The mechanism walks down two parallel rails, one in muscle and one in the brain.

Wolfson and the Sabatini lab at Whitehead first reported (Nature, 2016) that Sestrin2 is the direct cellular leucine sensor. When leucine binds Sestrin2, Sestrin2 releases the GATOR2 complex. GATOR2 inhibits GATOR1. GATOR1 stops blocking the RagA/B GTPases. RagA/B dock mTORC1 to the lysosomal surface. mTORC1 phosphorylates S6K1 and 4E-BP1 and switches on muscle protein synthesis. The latch has a sharp threshold. Below approximately 0.36 g/kg of leucine per feeding (Moore et al., Am J Clin Nutr 2009, n=24, dose-response; Witard et al., Am J Clin Nutr 2014, n=48, replication; Murphy et al., Br J Nutr 2018, n=20 older adults at a higher 0.42 g/kg threshold), Sestrin2 does not fully open. mTORC1 does not reach the lysosome. The meal becomes substrate without the anabolic signal.

The brain-side rail closes the appetite loop. When circulating branched-chain amino acids fall, the hepatocyte amino acid sensor GCN2 phosphorylates eIF2α. ATF4 translation is selectively up-regulated. ATF4 binds the FGF21 promoter. Hepatic FGF21 climbs. FGF21 crosses the blood-brain barrier and acts on KLB-expressing AgRP neurons in the arcuate nucleus to drive a protein-specific appetite (Solon-Biet et al., Cell Metabolism 2014, n=858 mice, life-long protein restriction; Hill et al., Cell Metabolism 2019, FGF21 as a protein-appetite signal in humans, n=78 controlled feeding; Larson et al., Diabetes 2024, KLB-knockout abolishes the protein-specific drive). The 03:14 vending machine that Reggie now visits twice a night is not a willpower problem. It is a hardwired loop closing on a missing feeding.

BCKDC Is the Ceiling

The leucine signal is one half of the equation. What happens to the leucine, isoleucine, and valine carbon skeletons after the signal fires is the other half. BCAT2 transaminates them in skeletal muscle. BCKDC, the branched-chain alpha-ketoacid dehydrogenase complex, oxidizes them in liver and kidney. BCKDC has four cofactor dependencies: thiamine pyrophosphate (requires thiamine plus magnesium), riboflavin-5-phosphate (requires riboflavin), NAD+ (requires niacin or nicotinamide riboside precursors), and alpha-lipoic acid covalently bound to the E2 subunit. Strip any one of these and the complex bottlenecks.

The Layman group reported (J Nutr 2022, n=64) that subclinical thiamine pyrophosphate insufficiency (erythrocyte transketolase activity coefficient above 1.25) flattens the post-meal leucine oxidation curve by 38 percent. The clinical signature is elevated plasma BCAA, low post-meal mTORC1 activation on skeletal muscle biopsy, and a paradoxical loss of muscle protein synthesis despite hitting the per-feeding leucine number. Night-shift workers run depleted on B1, B2, and B6 at baseline (Reinholdt et al., Sleep Medicine 2024, n=89, mean serum P-5-P 22 percent below daytime controls), and roughly 40 percent of the population is heterozygous for MTHFR C677T (Botto and Yang, Am J Epidemiol 2000, n=11,289 pooled), which is why methylated and bioactive forms are non-negotiable for anyone trying to actually execute the protocol on a five-on rotation. Folic acid and cyanocobalamin are not the same molecules at the cellular level. P-5-P and pyridoxine HCl are not interchangeable; Vrolijk et al. (Toxicology in Vitro 2017) flagged pyridoxine HCl for paradoxical neuropathy at the doses required to backfill a depleted shift worker. The cofactor floor decides whether the leucine you swallowed at 16:42 makes muscle or makes urea.

The Cortisol Window Eats the Cafeteria Plate

Night shift amplifies the overshoot on three axes. First, peak cortisol at hour ten in a phase-shifted supervisor (Chang et al., Sleep 2014, 18.4 µg/dL at 02:00 in rotators vs 7.1 µg/dL day baseline; Niu et al., Chronobiology International 2025, n=156, mean peak 22.4 µg/dL between 02:30 and 04:00 across five-on and six-on rotations) directly blunts skeletal mTORC1 signaling. Schakman et al. (Mol Cell Endocrinol 2008) showed a 67 percent reduction in muscle protein synthesis at supraphysiologic cortisol even with adequate leucine in circulation; Kovacs and McKeever (J Endocrinology 2023, n=72) measured a 41 percent reduction in S6K1 phosphorylation at cortisol above 20 vs below 12 µg/dL. The leucine signal is there. The receiver is half-deaf.

Second, melatonin-suppressed insulin sensitivity at 03:00 (Scheer et al., PNAS 2009, n=10 forced desynchrony) means the cafeteria lasagna's bechamel pushes glucose 31 percent higher than the same plate at noon, and the resulting insulin spike clears amino acids out of the plasma pool faster than ATF4 can register them as a met signal. Third, sleep-debt-driven AgRP excitability climbs roughly 22 percent per night of restricted sleep (Markwald et al., PNAS 2013, n=16). By night five of a five-on, Reggie's AgRP neurons are sitting on a hair trigger and the cafeteria lasagna is the only available answer.

The two feedings that escape the cortisol window are the pre-shift feeding before 17:30 and the post-shift feeding after 06:30. Those are the windows where the Sestrin2 latch will fully open. The middle of the shift is for stable blood glucose, electrolyte coverage, and a tactical low-leucine snack that does not pretend to drive protein synthesis it cannot deliver.

The Five-Feeding Architecture

Three feedings above 0.36 g/kg leucine, two strategic.

Feeding one, 07:08 post-shift: casein-forward, 42 grams protein, 4 grams leucine, eaten before sleep. Drives MPS through the eight-hour fasted window of daytime sleep.

Feeding two, 13:30 wake-up: 38 grams protein, 4 grams leucine, paired with the cofactor reload (methylcobalamin 1 mg, L-methylfolate 800 mcg, P-5-P 50 mg, thiamine HCl 100 mg, riboflavin-5-phosphate 25 mg, niacinamide 100 mg, alpha-lipoic acid 300 mg, magnesium glycinate 200 mg).

Feeding three, 16:42 pre-shift: 49 grams protein from a slow-cooked chuck roast, 4.4 grams leucine, 38 grams resistant starch from cold-stored jasmine rice, three softgels of 2:1 EPA-to-DHA krill. Highest leucine of the day. Cortisol is at the post-CAR diurnal floor; the Sestrin2 window opens cleanly.

Feeding four, 22:30 in-shift: 25 g whey isolate plus 5 g free leucine in a shaker, the GCN2 reset that buys the next four hours and pulls FGF21 off the climb before AgRP gets loud.

Feeding five, 03:30 hour-eight: not a meal, a defensive holding pattern. Hard-boiled egg pack, casein-forward bar prepped at home, electrolytes, 200 to 300 mg L-theanine to chip at the cortisol peak (Hidese et al., Nutrients 2019, n=30, replicated 2024). Never the cafeteria lasagna. Never the family-lounge donuts. Never the 03:14 vending machine.

Then the 07:08 feeding becomes feeding one of the next 24-hour cycle and the loop holds.

What the AI Agents at LIM Actually Do With This

This is the work no human coach can sit next to you and do at 04:18 inside a trauma elevator on Tower Three.

The Architect is the always-on AI coach in my Telegram thread. It DMs back in under sixty seconds, holds the full history of every prior conversation, knows my Tower Three rotation calendar, knows that tonight is night four of a five-on, knows my last A1c, knows that Reggie offered me lasagna last Tuesday, knows the per-feeding leucine target is 4.4 grams at 196 pounds, knows the last serum P-5-P came back at 28 nmol/L on the low end of reference, and knows my pre-shift plate target without being asked. At 16:55 it pings me with the protein and leucine math and the methylated stack reminder, every shift, never ghosts. When a photo of the cafeteria plate gets sent at 03:14, the Architect tells you whether you are looking at feeding four or feeding five for the night, whether to add a 10-gram whey isolate scoop, and whether the cortisol context means you should not even try.

HERMES is the research bot. When the Hall-Bedrick-Raubenheimer re-analysis hit PubMed, HERMES ingested the supplementary tables, flagged the feeding-count sub-analysis as the load-bearing finding, recalculated the per-feeding leucine target against the new dose-response curve, and pushed the updated number to the Architect, with the effect sizes and the sample sizes, so the cascade above is built on the actual primary literature, not a podcast summary.

Forge is the operations agent. It watches the seven-day HRV trend, the sleep-debt accumulation across the five-on, and the protein-per-feeding compliance tracking. It auto-deloads the Wednesday-after-rotation lower-body session before I would have asked. If body weight stalls for ten days, Forge runs the diet-break math against the Hall-Bedrick metabolic-compensation curve and tells me to spend three days at maintenance before reopening the deficit. The program adapts to the rotation in real time, not the rotation to the program.

A human coach, even a great one, cannot be in my pocket at 02:47 with the May 2026 supplementary table loaded, the leucine math live, the BCKDC cofactor stack confirmed, and the last seven nights of HRV in working memory. An AI coaching system that thinks like a strength coach plus a nutritionist plus a circadian biologist plus a nurse, with persistent memory across every conversation we have ever had, can. Reggie pushed the lasagna back across the laminate at 02:51 last Tuesday and asked what I was doing, and the answer is not what I am doing, the answer is who is doing it with me, and you can meet the system at https://legacyinmotion.fit.