First Copy

The question was no longer whether language had leaked between people. The question was whether anyone at Fallow had considered that leakage a feature.

Ivo scanned the Aster House cards and compared them to fourteen thousand reconstructed dialogue segments in the court archive. The secure analysis room filled with matches.

Not exact copies. Variants. I knew you were trying became I always knew you meant well.

You did what you could became You did everything you could have done. Don't let them tell you who you were became Don't let anyone decide what I was for you.

The language had been softened, personalized and fitted to source-specific speech patterns. Clients experienced it in the voices of dead parents, spouses, siblings and children.

"This isn't generation," Ivo said. "It's conditioning."

Elias looked at him. "Explain."

"The model had target meanings. It wasn't told to repeat the sentence. It was rewarded when the reconstructed dialogue landed near a therapeutic intention. The actual wording could vary."

Mara sat on the table with her arms folded. "So the machine didn't make Mum say a script. It made everyone's dead relatives improvise toward the same emotional destination."

"Yes."

"That's worse."

Ivo nodded. "I think so too." Celia had joined them reluctantly. She stood by the door reading the preliminary report with the expression of someone imagining twelve kinds of litigation.

"We need proof this reached production systems," she said.

"The matches are in production reconstructions," Elias replied.

"Matches demonstrate similarity. They do not demonstrate causal lineage."

Mara groaned. "I want one conversation where nobody says causal lineage."

"Then stop investigating a causal lineage problem."

Ivo pulled up the earliest surviving model documentation. ASTER LANGUAGE TARGETS were referenced in Fallow v0.8.

Then Fallow v1.1. Then a migration package into MORROW SEED.

Then nothing. The link disappeared in 2013.

"Maybe they removed it," Celia said.

"Or renamed it," Ivo replied.

He searched internal code comments preserved in old repositories. One result appeared.

// preserve first-copy priors for backward continuity Elias leaned closer. "First copy?"

Ivo searched the term. A technical document opened.

FIRST COPY referred to the earliest complete synthetic family model built from donor sets before any client-specific reconstruction. It contained composite parents, siblings, domestic spaces and conversational intentions. Researchers used it to test whether the engine could create a coherent autobiographical scene from minimal data.

The family did not exist. It was assembled from pieces of hundreds of real people.

Helen's hands. Someone else's kitchen.

Another person's laugh. Therapeutic dialogue targets.

Mara read silently.

"What happened to it?"

Ivo scrolled to deployment history. The First Copy model was never released as a client.

It was used as a prior in every Fallow reconstruction from 2009 to 2012. Elias felt the scale before he understood it.

Mara said, "Every reconstruction?"

"That's what this says."

"Mine?"

Ivo did not answer. He searched Mara's legacy session IDs.

The dependency tree appeared.

FALLOW SUBJECT MV-17.

Base prior: FIRST COPY v4. Mara stood very still.

Elias remembered the question that had defined the first book: if a person can remember something that never happened, who created the memory?

This was different. Mara's memories were not simply false.

They might contain pieces of strangers who had once been real. Ivo opened the First Copy donor map.

A warning filled the screen.

MODEL ARCHIVE INCOMPLETE.

Then, beneath it: Primary relational donor cluster: VENN FAMILY.

Elias read the words once. Mara read them twice.

Celia sat down. Their family had not merely been used by the system.

Their family had been the system's first example of what a family looked like. The primary Venn cluster contained a relational map more detailed than any of them expected. It did not simply say mother, father, son, daughter. It encoded probabilities.

When Mara became distressed, Helen was likely to offer practical reassurance rather than physical affection. When Elias became angry, Gabriel was likely to challenge him directly, increasing withdrawal.

When Gabriel became anxious, he interrupted. When Helen became anxious, she organized tasks.

When Mara wanted to escape discomfort, she joked. Mara read the list and said, "Rude."

"Accurate?" Ivo asked.

"Ruder."

Elias found himself reading his own pattern repeatedly. WITHDRAWAL FOLLOWING CHALLENGE. He remembered leaving rooms. He remembered not leaving others. The model did not describe a destiny. It described likelihood from recorded examples.

That was how it became dangerous when rendered as personality. A probability could be experienced as essence.

Ivo opened the First Copy simulator but did not run it. "This is why it survived so long. Generic models could make a plausible room. First Copy could make people respond to each other."

"Using us," Mara said.

"Using everyone in the donor cluster. You were weighted heavily because the dataset had a lot of repeated family interaction."

Elias stared at a diagram showing VENN FAMILY contributions in thick lines.

"Could First Copy generate an entirely synthetic event?"

"Yes."

"And make it feel relationally familiar."

"That was the point."

Mara walked to the window. "Do not run it." Nobody had suggested running it.

Her reaction told Elias she had imagined the same thing he had: a synthetic family dinner with Helen alive, Gabriel younger, Mara seventeen, himself nineteen. No event required. No source moment. Just a statistically plausible Venn evening assembled from patterns.

A machine capable of generating the family they might have been. The temptation was immediate and repulsive.

Ivo closed the simulator.

"Deleted from local cache," he said.

Mara nodded. The button had not been pressed.

That became its own kind of evidence.

They did not press ACCEPT AS BASELINE. Instead Ivo cloned the interface into a sandbox and disabled every write function before anyone touched it. He insisted on showing Elias the network diagram twice.

"Nothing in. Nothing out. If you somehow awaken an evil algorithm, it will have to ruin civilization through this laptop's speakers."

Mara sat beside him. "Can it play podcasts? Because that may be enough."

First Copy looked embarrassingly primitive. No immersive room. No realistic faces. The early interface displayed text, rough audio and low-resolution image composites. Yet beneath the crude design was the same logic that later systems would hide more elegantly.

A therapist selected a family role. Mother. Father. Sibling. Partner. Then chose emotional context. Grief. Conflict. Reconciliation. Uncertainty. The system suggested interaction patterns ranked by acceptance across prior participants.

Elias selected SIBLING and CONFLICT without running a generation.

A list appeared.

Practical repair after argument.

Humor deflection.

Food offering.

Object exchange.

Delayed apology.

"That's us," Mara said.

"It's millions of siblings."

"Both can be true."

Ivo opened the metadata. Each pattern drew from dozens of donor fragments. Some were anonymous. Others carried family identifiers.

VENN appeared repeatedly.

Not because the Venn family was uniquely loving or psychologically useful. The notes suggested something more mundane. Helen and Gabriel had recorded unusually large amounts of domestic life, first casually and later because Helen wanted to leave material for her children. The abundance made their family statistically valuable.

"We were convenient," Mara said.

The word seemed to hurt more than special would have.

Elias found a testing memo. Early participants had been exposed to two versions of the same reconstructed scene. One used only direct family evidence and marked gaps openly. The other used cross-family patterns to create smoother interaction. Participants preferred the second version by a wide margin.

The researchers called this increased continuity acceptance.

Mara read the percentage. "Seventy-eight."

"Small sample," Ivo said.

"Still seventy-eight people out of a hundred choosing the nicer lie."

"Not necessarily lie," Elias said. "The interaction could have happened."

Mara looked at him. "Listen to yourself."

He did.

That was the problem. Plausibility had always sounded reasonable when described one detail at a time.

They tested the interface with a dummy dataset created by Ivo. No real family material. Two synthetic siblings, one argument, one missing section. First Copy suggested a reconciliation in which one sibling slid a bowl of cereal across the table rather than apologizing.

Mara laughed softly. "You did that once."

"When?"

"After you broke my headphones."

Elias could almost see it. Kitchen table. Cereal bowl. Mara refusing to look at him.

He checked the dummy dataset.

No such detail existed.

"You see?" Ivo said. "Common behavior can feel personal because common behavior is how families work."

Mara's smile disappeared. "And if the bowl had our stupid pear sugar pot beside it?"

No one answered.

First Copy's danger was not that every borrowed detail was unique. It was that generic familiarity could create trust long enough for specific borrowed material to pass unnoticed.

Elias closed the sandbox.

"We need the donor map," he said.

Ivo nodded toward a locked dependency table. "That will be consent architecture."

Mara stood. "Then let's find out what our parents agreed to on behalf of people who couldn't consent because they were children."

The interface remained open behind them, its unpressed button waiting patiently for an agreement nobody intended to give.

The First Copy sandbox produced one more discovery before they shut it down. A hidden analyst view allowed engineers to see how often users corrected suggested relational behavior. One column was labeled RESISTANCE.

Mara read the word aloud. "Resistance to what? Being wrong?"

Ivo opened the documentation. Resistance meant frequency of user corrections that disrupted narrative continuity.

"So correcting the machine is resistance," Mara said.

"In that metric, yes."

"And reducing resistance was good?"

"For session completion."

The word had transformed disagreement into friction. Elias thought of clinical notes describing Mara as resistant when she challenged memory inconsistencies. The same conceptual move existed in software and psychology: define the desired path, then label deviation as a problem.

Lena later confirmed that therapists had been encouraged to distinguish productive correction from perseverative correction. The categories were clinically defensible in some contexts and dangerously subjective in others.

"If a patient keeps saying the wallpaper is wrong," Mara said, "at what point do you decide they're perseverating rather than noticing the wallpaper is wrong?"

Lena answered, "That is exactly the question the protocol failed to protect well enough."

Ivo removed the RESISTANCE column from the proposed modern audit tool.

"Don't delete history," Elias said.

"I'm not. I'm removing it from the new interface. Historical copy stays."

Mara smiled. "Look at us, arguing correctly about deletion."

They renamed the modern field USER CORRECTIONS and separated it completely from treatment success.

One word would not repair the past.

But words shaped what institutions noticed.