The Witness Pilot

The legal pilot began with a robbery that never mattered to anyone outside the research team. A convenience-store clerk witnessed a masked man enter, threaten him, take cash and leave. The crime had been recorded by two cameras, allowing researchers to compare what the clerk remembered with what actually happened. No suspect was prosecuted through the study. The case was chosen precisely because ground truth existed independently.

Jonah showed Elias the redacted report in Saira Malik's office.

"This was 2010," Jonah said. "Fallow had started getting inquiries from legal consultants about whether guided reconstruction could help witnesses recover details after traumatic events."

The clerk initially remembered the robber wearing a black jacket, standing near the drinks refrigerator and saying, Open the drawer. The video showed a dark navy jacket. He stood two feet farther left. Audio was too poor to determine exact words.

Researchers built several probable reconstructions using the video, witness report and store layout. In one version the robber said, Open the till. After three guided sessions, the clerk's confidence that the phrase had been Open the till rose from uncertain to near-certain.

The video still could not verify it.

"Maybe that was the actual phrase," Elias said.

"Exactly," Jonah replied. "Could be. That's why the study is frightening. If the generated sentence happened to be right, the process still made the witness more certain for the wrong reason."

Another participant initially could not remember whether a suspect held an object in his right or left hand. Reconstruction used the right hand because that was statistically more likely from the available camera angle. Weeks later the witness confidently identified right.

Ground truth from another camera showed left.

"How many?" Elias asked.

"Twenty-three participants. Small sample. Enough to kill the legal program."

"But not the clinical one."

"Because clinical outcomes were different. A grieving person doesn't need to testify about which hand a dead parent used to hold a cup."

"Until the memory enters a family dispute. Or an abuse allegation. Or police ask them later."

"Exactly. Memory doesn't stay inside the use case you designed it for."

Saira opened a later document. In 2012, Fallow researchers proposed a permanent visual marker for generated details inside every reconstruction. Clinical staff objected that the marker reduced immersion and therapeutic benefit. A compromise placed the marker at session start and end.

Users later remembered seeing the scene. They did not always remember seeing the warning.

"Why didn't this end everything?" Elias asked.

Jonah looked at him as if the question was too familiar. "Because the therapy helped people. I keep telling you. People stopped having nightmares. Trauma patients could enter rooms they'd avoided for years. A woman who had lost her child used reconstruction to recover ordinary mornings that had become inaccessible behind the final hospital week. She said it gave her daughter back a life instead of only a death."

"And were the mornings accurate?"

"Some details. Some inferred. The mother knew that at the beginning. Five years later? We don't know."

Elias thought of his own mother. Red dressing gown, blue dressing gown, perhaps green according to Mara.

"Did Mara know about the witness pilot?"

Jonah nodded. "That's what changed her from suspicious patient to whistleblower. She found a summary in a shared research folder. She realized the team already knew confidence could be engineered independently of accuracy."

"You gave it to her."

Jonah looked down. "I left a directory accessible. Whether that was intentional depends on which lawyer is asking."

"You were twenty-seven."

"Twenty-six. Not an excuse."

"Why help her?"

Jonah took time.

"Because she asked a question none of the senior people liked. If a patient remembers an inferred detail later, do we mark the therapy successful or the record corrupted? Nobody could answer without admitting the metric was doing moral work we had hidden inside statistics."

Elias looked through the report again. At the bottom was a handwritten note from Adrian Rohe.

Confidence gain is not fidelity gain. Do not pursue evidentiary deployment.

"He knew," Elias said.

"Yes."

"And stopped it."

"Yes."

"Then years later approved consumer design that reduced source reminders."

"Also yes."

Elias leaned back.

Rohe was easier to understand when imagined as a hypocrite. Harder when the same person had correctly identified the risk, stopped one application, then gradually accepted the mechanism elsewhere because the context felt different.

Institutional failure did not require ignorance. Sometimes it required remembering a warning only inside the category where it had first been written.

Saira closed the file. "This is the strongest evidence we have because it doesn't depend on anyone's autobiographical account."

"Will it be public?"

"Redacted."

"Rohe will argue it proves the company acted responsibly by ending the pilot."

"He should. That's part of the record."

Elias looked at her.

"You don't want a simple case either."

"Simple cases make bad safeguards. If we pretend every decision was malicious, future institutions learn only to avoid looking malicious. I want rules that survive good intentions."

That sentence stayed with Elias. Rules that survived good intentions.

Fallow had been built from people trying to reduce suffering. Morrow had been built from people trying to make reconstruction useful.

Elias had built Continuity trying to protect Mara. Mara had disappeared trying to protect herself.

Good intention was not innocence. It was simply one more fact.