Previously on Chiron: Juniper and her colleagues left the last meeting with more questions than answers, realizing their most important challenge might not be what the AI recommended—but what everyone assumed it understood.
Note: All names used in Chiron are fictitious. Additionally, this is the third of eight episodes in which we build a story arc using the same characters. At the end, you will find a character cheat sheet to help keep everyone straight from episode to episode.
Mercer arrived before everyone else.
Juniper noticed it the moment she stepped into the conference room at 7:52. He sat alone at the far end of the table with a printed copy of the email in front of him, one hand resting beside a closed notebook, the other holding an uncapped pen.
Vale wasn’t there. Rowan wasn’t there. Sol wasn’t there.
Only Mercer.
And the sentence.
If the system can measure it, why are we assuming that is the problem?
Juniper had printed the email the night before. Mercer had done the same.
“Morning.”
“Morning.”
He glanced at the folder in her hand. “You brought evidence.”
“I brought examples.”
“Better.”
He opened his notebook. “Before everyone gets here, I want to be clear. I’m not defending the dashboard.”
“I didn’t say you were.”
“No,” he said with a faint smile. “But you thought it.”
Juniper met his eyes.
“I’m defending the question.”
“That much was obvious.”
“It’s a good question.”
“It’s a dangerous question.”
“The useful ones usually are.”
Juniper finally sat. Outside the glass wall, the clinic was waking up. Staff moved through the hall carrying binders, water bottles, therapy materials, and the postural fatigue of people whose day always started before the schedule admitted.
Mercer tapped the paper.
“If the system measures clinical reasoning poorly, that’s a problem. If it measures it without transparency, that’s a problem. If it measures something else and calls it clinical reasoning, that’s a serious problem.” He paused. “But if it’s measuring something real, then perhaps measurement isn’t the problem.”
Juniper waited.
“The organization may have been relying on an asset it never realized it had.”
Before she could answer, Vale entered carrying coffee and his laptop. Rowan followed, visibly irritated by the meeting’s existence. Sol came in last with only his small notebook.
Vale connected to the display.
The dashboard appeared.
Quality indicators.
Documentation metrics.
Recommendation acceptance.
Rationale completion.
Beta analytics.
It looked clean.
That was exactly the problem.
Nothing on the screen suggested ethical uncertainty, philosophical conflict, or a decision future conferences would call a turning point. It looked like software.
Vale opened the beta panel.
Clinician Reasoning Profiles
The room fell silent.
Rows populated beneath the heading.
Acceptance rates.
Modification themes.
Referenced variables.
Confidence scores.
Rationale length.
Contextual variable density.
Rowan frowned.
“Contextual variable density?”
“It appears to estimate how often clinicians reference contextual factors in their rationales,” Vale replied carefully.
“Appears to.”
“Yes.”
Mercer leaned forward. “Can we open one?”
Juniper looked at him.
“That’s exactly what concerns me.”
“I understand. Use a de-identified profile.”
Vale selected C-147.
A summary appeared.
Clinician frequently accepts AI recommendations when recent responding is stable across three or more sessions… Clinician frequently modifies recommendations when caregiver implementation differs across routines… Confidence ratings decrease when home and clinic data diverge.
Mira, standing against the wall, studied the screen.
“That could be me.”
“It’s de-identified,” Vale said.
“I know. That’s not what I meant.”
Mercer nodded toward the profile.
“This is useful.”
Rowan sighed.
“I knew you’d say that.”
“If I were onboarding a new clinician,” Mercer replied, “this tells me something meaningful about how experienced staff reason through ambiguity.”
“It tells you how they explain decisions inside this interface.”
“That’s an important distinction.”
“Important enough to slow down.”
“Agreed.”
Rowan blinked.
Mercer continued. “I’m not arguing for speed. I’m arguing for accurately identifying what this actually is.”
Juniper opened her folder.
“There it is.”
“What?”
“The word.”
“Asset?”
“Yes.”
Mercer looked around the room.
“We can soften it if that makes everyone feel better.”
“It won’t.”
“Then we should use the precise term.”
He turned toward the screen.
“Treatment rationales, supervision edits, payer responses, fidelity data, and clinician explanations aren’t just documentation.”
He let the sentence settle.
“They’re behavioral infrastructure assets.”
Silence.
The phrase landed with uncomfortable precision.
He hadn’t talked about clients.
He hadn’t talked about care.
He had talked like an investor describing infrastructure.
Behavioral infrastructure assets.
Clinically unsettling.
Strategically accurate.
Sol finally spoke.
“What makes something infrastructure?”
“Infrastructure reduces friction for future behavior,” Mercer answered. “Roads. Databases. Training systems. Referral networks. Clinical protocols. Someone builds conditions that make future action easier.”
“And if that infrastructure consists of prior clinical behavior?”
“Then the organization has to decide whether it’s preserving knowledge, extracting labor, improving care, building dependency—or all four.”
For the first time that morning, Rowan had no immediate reply.
Organizations may assume that preserving records is the same as preserving knowledge. Those concepts are related, but not identical. A record captures a product of behavior: what was written, selected, approved, changed, rejected, or justified at a particular point in time. Knowledge is more than the preserved product. It includes the context in which the product occurred: the data available, the contingencies operating, the training history of the clinician, the organizational pressures in place, the client and family conditions, the payer constraints, and the consequences that made one response more likely than another. Capturing knowledge allows the system to generate the “best” behavioral product in the future, whether the response should be the same or differ from the behavior captured in the data.
This distinction matters when AI systems are layered onto clinical archives. A model can detect patterns across thousands of records, but those patterns are not automatically equivalent to expertise. Some patterns may reflect strong clinical judgment. Some may reflect documentation habits. Some may reflect payer pressure, staffing shortages, regional practices, supervisory preferences, or the limitations of the available data. The system can identify that a pattern occurred. It can’t automatically determine whether the pattern occurred because it was clinically wise, administratively reinforced, ethically necessary, or simply common.
For BCBAs and organizational leaders, the practical takeaway is to treat historical records as selection products, not neutral containers of truth. Before using archived clinical data to support AI recommendations, leaders should ask what conditions produced the records, whether those conditions still apply, which historical practices they would still endorse, and which patterns should be excluded from future influence. Strong governance is not only about privacy and access. Strong governance is also about deciding which histories deserve to shape future care.
