His fingers found the protein bar without his eyes leaving the middle monitor. The wrapper tore the rest of the way open along its seam, and he bit off a corner and chewed without tasting it. Peanut butter, probably. He'd bought a box of twelve from the CVS on Broadway three weeks ago and was on number nine or ten, which meant he'd been averaging one every two to three days, which meant he'd been in this office past dinner roughly forty percent of his evenings this month, which was actually lower than he'd estimated. Small mercies. He set the bar on a stack of printed consent forms and forgot about it.

The dose-response curves on screen two were wrong.

Not wrong in a dramatic way. Not the kind of wrong that announces itself. Wrong in the way a drawer sticks when the humidity changes — you push and it resists and you push harder and then you stop and think about the drawer itself, the wood, whether something has shifted in the frame. His 50mg cohort and his 200mg cohort tracked the way pharmacology said they should: more drug, more effect. Linear, clean, boring in the best possible sense. But the 10mg cohort — his lowest dose, his control-adjacent group, the patients getting barely enough caladrite to register on a mass spec — were outperforming the 200mg group on two of six cognitive measures. Verbal fluency and working memory. Not by a little. By a margin that made his regression line look like it had been drawn by someone who didn't understand axes.

James pressed his chin to his chest and rotated his neck slowly, one vertebra at a time, feeling each pop or refusal to pop. The courtyard outside his window was dark. The broken bench was still there, one slat hanging at an angle that suggested it had been broken by someone sitting on the arm rather than the seat. Facilities had put a cone next to it in January. The cone had blown over in March and no one had stood it back up.

He ran the analysis again, excluding the three patients in the 10mg group who'd reported concurrent caldrex use for sports injuries. The curve didn't change. He ran it again, controlling for age, sex, baseline cognitive score, and time since enrollment. The curve got worse — meaning the effect got stronger, meaning whatever the low-dose patients were experiencing, it wasn't explained by the obvious confounders.

He opened his notebook. The page was dated today, Thursday, and his notes from this morning's lab meeting were at the top in his handwriting — compact, fast, the letters leaning forward like they were late for something. He wrote: 10mg > 200mg on VF + WM. Robust to exclusions. Not artifact.

Then he stared at the sentence.

If the inverse dose-response was real, the drug wasn't doing what his model said it was doing. His entire Phase II was built on a monotonic assumption: more caladrite, more cognitive effect, up to some ceiling. That was how drugs worked. That was how nearly everything worked. You didn't get stronger coffee by using fewer beans. But calcium signaling — and caladrite's mechanism was almost certainly calcium-mediated, the nanoparticles slotting into ion channels like keys into locks — calcium signaling was nonlinear at low concentrations. He knew this. He'd cited it in his own grant application. At low doses, you might get a qualitatively different kind of activation. Not more of the same. Something else entirely.

His pulse was up. He could feel it in his wrists.

The thing that happened next was the thing he would have trouble explaining later, if he ever tried to explain it, which he wouldn't: before the dread arrived, before the obvious implication landed, he spent approximately four seconds composing a discussion section in his head. We report a paradoxical inverse dose-response relationship in neurograde caladrite's cognitive effects, suggesting a biphasic mechanism of action at the calcium channel level. It would be the most interesting finding in neuroimmunology this year. It would make Elena Mwangi's linear dose-escalation study look like it had missed the entire point. Four seconds of that. Maybe five.

Then the rest of it caught up.

He'd taken caladrite six weeks ago. A single intranasal dose from his own lab supply, lot SFZ-2026-0117, processed from Sofala extraction batch — he pulled up the provenance file, a habit, checking the chain like a man checking his watch when he already knew the time — batch 7.2.09, ENC Zambezia subfield, routed through the ChemChina processing node in Beira. He'd chosen 200mg because his own data, at the time, suggested 200mg was optimal. The razor-sharp twelve days that followed had confirmed his choice, or seemed to, or felt like confirmation in a way that was difficult to distinguish from actual confirmation.

But if 10mg was the sweet spot. If 200mg was doing something else to the brain, something that felt like enhancement but operated through a different mechanism.

He looked at his hands on the keyboard. Steady. His three o'clock tremor had been gone for weeks, actually, and he'd attributed that to cutting back on coffee, which he had done, which was a real thing with a real mechanism, which was probably the explanation.

Probably.

He picked up the notebook again and flipped backward. Three months ago, February, his handwriting was rounder, the letters more separated, a mild rightward slant he'd had since grade school. He flipped back further. September, six months ago, before the dose. Looser. More uncertain on the page, the pen pressing harder on downstrokes the way it did when he was unsure of what he was writing while he wrote it. The handwriting of a man who revised his thoughts mid-sentence.

His handwriting now was faster and more certain and, he noticed with a feeling he chose to classify as clinical interest, objectively better. More legible. More consistent. As though his hand had stopped negotiating with his thoughts and started taking dictation.

Something beeped in the hallway — the autoclave in Lena's section finishing its cycle, running unattended, which meant Lena had loaded it before leaving and trusted it to do its work without her, which was not how Lena usually operated, which probably meant nothing.

He should email Elena Mwangi. She'd requested his raw data two weeks ago and he hadn't responded. The request had come through the standard collaboration portal, formal, routine, the kind of thing you responded to within 48 hours if you were being collegial. He'd read it, flagged it, moved on. Fourteen days. That was long enough to be a message whether he meant it as one or not. He'd do it tomorrow.

James looked at the curve one more time. The 10mg patients, getting barely a whisper of the drug, their brains lighting up like they'd found a frequency the higher doses blew right past.

He closed the notebook. Opened it again. His pen was in his hand. He didn't remember picking it up.

He wrote: check confounders.

The handwriting was beautiful.