Out of the Loop

Lena Varga arrived at 7:42 Saturday morning with two instrument cases and no overnight bag.

Evan had known her for six years, mostly as a sequence of comments in shared documents. In person she was taller than he remembered and wore her gray hair clipped so tightly behind her head that it appeared to be part of the experimental protocol. She left both cases on the front step until he confirmed that the apparatus was off.

Ruth opened the door behind him.

"Are you the person who doesn't believe him?" she asked.

Lena looked at Evan. "How much did you tell her?"

"He told me you were coming," Ruth said. "The rest was probability."

Lena brought the cases inside.

She spent the first hour on OLD-7 without discussing causal histories, consciousness, or the ten die rolls. She photographed the annulus, measured the silver-epoxy contacts, mapped the lower-face resistance at four points, and objected to the thermocouple mounting. She objected again after Evan showed her the calibration file.

"This gives you absolute temperature within half a kelvin if the adhesive behaves," she said. "It does not give you a trustworthy local gradient during a moving front."

"The anomaly tracks the phase relation, not the absolute gradient."

"Then bad thermal metrology is an unnecessary gift to anyone who wants to dismiss it."

She replaced two probes with thin-film sensors from her case and attached their loggers to separate batteries. She brought her own Hall probe, photodiode noise source, rubidium-referenced counter, oxygen meter, and laptop with a clean operating-system image. Every data link crossed through fiber she had packed herself.

Evan watched her convert his garage back into an unproven garage.

At nine, Lena read the raw files behind OUTCOME-01. She reproduced the timestamp hash, rebuilt the die sequence from the local video, and inspected the vision code. At 9:37 she closed the laptop.

"I don't see the fraud," she said.

"That isn't the same as believing the result."

"Correct. It's the prerequisite for spending the rest of my Saturday here."

She wrote the replication protocol on paper. Evan could answer questions about safety limits and connector locations. He could not choose targets, touch the controller after arming, see live analysis, or remain inside the house during the independent run.

"The house?"

"Your software, hardware, and experiment design already remain in the loop. I would like to remove at least the part that argues."

Ruth agreed to spend the afternoon with Marisol after Lena explained that the garage would be noisy. Evan objected that it was no noisier than usual. Ruth said the explanation had the advantage of being directed to her.

For the first run, Lena loaded a wrap-safe build compiled from Evan's source on her laptop. She kept the build labels hidden and used her own script to choose between safe and fault-reproduction binaries. Evan performed the mechanical setup without knowing which image had been installed, then sat facing away from the displays.

Nothing happened. The sensors remained full-rank. Lena's neutral task scores matched her baseline. When the run ended, she revealed the safe build.

She repeated the procedure with a second concealed choice. This time the scheduler crossed the wrap boundary, the early heater edge accumulated its phase walk, and the independent channels began losing dimensions. Lena did not announce the change. Evan felt the quiet arrive and wrote onset on his paper at 11:26:18.

Lena recorded her own onset nine seconds later.

After shutdown, she graded the task sheets without speaking. She had solved two unfamiliar thermal-network problems faster than baseline, rejected a third because its boundary conditions were incomplete, and corrected an error in the answer key.

"Describe it," Evan said.

"No. You describe yours, and then we compare."

He wrote: reduced competing interpretations; low anxiety; correct action presents as recognition rather than choice. Lena turned over her page.

She had written: alternatives remain representable but lose urgency; solution path feels recalled; confidence rises only when constraints close.

Neither description proved anything. Their agreement did not make the experience external. The task scores and covariance trace made dismissal less comfortable.

Lena ate half a sandwich at the kitchen table and designed the run that excluded Evan.

She generated two target definitions on her clean laptop, committed both hashes to a timestamp service, and placed one encrypted file on a removable drive. A hardware coin flip chose which target she would decrypt after Evan left. He saw only the comparator interface and the safety limits.

The target was not another long die sequence. During each of 2,048 decision points, fresh noise from Lena's photodiode selected APPLY or HOLD for a corrective phase step. Her private mapping defined which physical action counted as preferred. Reversing the hidden mapping would reverse the raw bit bias if the effect followed her objective rather than Evan's circuitry.

INDEPENDENT OPERATOR TEST

operator: L. Varga
designer absent: minimum 5 km, no network or phone contact
source: isolated photodiode avalanche noise, local raw log
mapping: committed and hidden from E. Hale
consumer: fresh bit selects APPLY / HOLD
stop: 2,048 decisions or ordinary safety limit

primary:
  favored physical action under hidden mapping
secondary:
  covariance rank, operator task calibration

Evan drove to a diner beside the interstate, 6.3 kilometers from the garage. He left his phone in the car, ordered coffee, and received a laminated menu dense with possible breakfasts. The experiment had begun before he sat down.

He tried to read. Without the apparatus nearby, his attention behaved normally. A child in the next booth constructed a wall from jelly containers. Two truck drivers disagreed about roadworks. The server asked whether he wanted more coffee, and Evan took several seconds to decide.

At the garage, Lena controlled a machine built from his choices. Every phase table, safety threshold, and definition of lower loss had passed through him. If the anomaly vanished, he was the active element. If it persisted, he had built something that no longer required his presence.

The second possibility should have been reassuring.

He remained at the diner for two hours after the planned stop, as instructed. When he returned, Lena was sitting on the back step with the instrument cases closed. The garage breaker was off. A red cable tie passed through the disconnect handle.

"Safety fault?" he asked.

"No. Result."

She made him wait while she copied the raw logs, signed both copies, and revealed the selected target. Zeros had meant APPLY. The photodiode source produced zeros on 54.9 percent of decisions. The lower-loss action had been selected on 55.1 percent after invalid samples were excluded. During a no-consumer block immediately before the live gate, the same source had remained at 49.8 percent.

The covariance rank fell to 1.41. Lena's task performance improved. Evan had been absent for the entire interval.

"You reversed the mapping?" he asked.

"Second run. Ones meant APPLY. Ones rose to 54.6 percent."

She handed him the summary.

REPLICATION R-01

condition                    rank     preferred action
safe build                   4.63          .501
fault build / no consumer    4.51          .498
fault build / Lena live      1.41          .551
fault build / mapping rev.   1.46          .546

E. Hale absent during live conditions.
Raw source bias reverses with private action mapping.

Evan read the table twice. Relief reached him first. The effect did not depend on his expectation, memory, or private willingness to see a pattern. Another person had designed the test, operated the apparatus, experienced the quiet, and obtained a bias that followed a meaning he did not know.

"It's real," he said.

"The effect is real. Your explanation remains an explanation."

"It followed your target."

"It followed a control topology that distinguishes two actions. We do not know what else it followed."

She pointed to the red cable tie. "You stop the operator runs. No more task lists, no more eighteen-minute exposures, no testing this on anyone. We replicate at a neutral site with isolated control code and people who don't know your theory."

"You ran it twice."

"To determine whether you had a measurement problem."

"And you felt it."

"Yes. That is why you stop."

The answer irritated him because it used his strongest evidence as an argument against the thing the evidence made useful.

Lena opened his exposure log. "Your interval stayed fixed and your frequency doubled in six days. You counted colleagues changing their minds as possible effects, then scheduled another run to make a work plan they would accept. This is not controlled self-experiment. It is a reward loop with good documentation."

"The work was correct."

"Correct work is not evidence that the method has no cost."

Ruth and Marisol returned through the side gate while they were still arguing. Ruth carried a library book and a paper cup. She stopped at the red tie across the disconnect.

"Did she find the problem?" Ruth asked.

"She reproduced the result," Evan said.

"I found a reason to stop testing people," Lena said.

Ruth looked from Lena to Evan. "He's already tested himself."

No one answered. Marisol took Ruth's paper cup and went into the kitchen, recognizing a conversation that would eventually need coffee.

Lena stayed until evening. They wrote a joint replication note that described the hardware, source isolation, hidden mappings, and operator separation without naming causal pruning. They encrypted one copy for Lena, one for Evan, and one for a neutral repository neither could alter alone.

At the bottom, Lena added a hold condition.

HOLD PENDING INDEPENDENT SITE

no cognitive-exposure protocols
no outcome targets beyond replication controls
no exposure of patients, dependents, or uninformed participants
no increase in cycle depth or affected systems

Restart requires both investigators' signed approval.

Evan signed. Lena signed beneath him. She photographed the red tie and took one half of its numbered seal strip.

When she left, the garage felt larger with the apparatus unavailable. Evan put away her replacement probes, cleared the paper cups, and found Ruth's library book on the workbench. It was a collection of topology exercises for undergraduates, checked out because Ruth liked the diagrams and denied remembering the course.

He returned it to her room.

Ruth was asleep with the lamp on. On her bedside table, the home-care schedule lay beneath a note she had written to herself in Denise's pen.

TUESDAY: WOMAN FROM AGENCY. SHE IS SUPPOSED TO BE HERE.

Below it, in smaller writing, Ruth had added:

ASK EVAN WHY HIS MACHINE MAKES PEOPLE BETTER.

He turned off the lamp without waking her.

In the garage, Lena's red tie crossed the disconnect handle. The hold prohibited cognitive exposure and uninformed participants. It did not say anything about magnetic-field strength, distance, or the minimum cycle count at which an effect became an effect.

Evan noticed the omission because he was trying not to look for one.