Consensus

Orpheus was not a room. It was three sealed servers in a decommissioned data center beneath Morrow's former headquarters, physically disconnected from every network and listed in property records as "legacy environmental monitoring equipment."

Celia stared at the inventory photograph.

"I hate this company."

Ivo looked delighted. "Air-gapped cold storage. Very 2012."

"Why was it hidden?" Elias asked.

"Because somebody wanted to preserve it without admitting it existed," Celia said. "Or because asset management is incompetent. Both are historically plausible."

Access required court authorization, two privacy officers and a hardware specialist old enough to recognize the storage format. It took four days.

During those four days, the story leaked. Not Orpheus.

First Copy. A claimant posted on a private support forum that her released reconstruction contained a hand gesture traced to another family. Someone else posted a dialogue match. Within forty-eight hours, people were comparing transcripts online.

Identical phrases appeared. Similar apologies.

The same joke about burnt toast. A mother in Bristol and a father in Glasgow both said, You were always better at leaving than goodbye.

A dead husband and a dead sister both said, Don't make a museum out of me. The internet did what institutions had failed to do.

It found patterns. It also found false patterns.

By the third day, users were claiming that any shared phrase proved memory theft. News sites published lists. Amateur investigators identified innocent donor families from old videos. One woman received threats because people believed her dead husband's laugh had been used in strangers' reconstructions.

Mara watched the coverage from Elias's sofa.

"This is why Celia wanted verification first."

"Yes."

"I hate when she's right."

"She enjoys it enough for everyone."

Tess refused to publish until they had causal evidence.

"Similarity isn't provenance," she said. "If I print that Morrow copied one dead mother's sentence into a thousand mouths and it turns out to be a common phrase, we become the problem."

On the fourth morning, Orpheus came online in a sealed room. The index contained 9.4 terabytes of old model states, source maps and session renders.

Ivo whistled softly.

"They kept everything."

"Why?" Elias asked.

"Engineers don't delete things when lawyers tell them to. They rename them backup."

They found Session Fourteen. They also found a folder called CONSENSUS TESTS.

Elias opened the documentation first. The purpose was blunt.

Researchers wanted to know whether two people exposed to the same reconstructed event would gradually converge in their recollections of it.

Not whether they would remember it accurately. Whether they would agree.

Agreement had been treated as a measure of stability. Mara read the sentence aloud.

"They made consensus into evidence."

Lena's voice came through the secure conference line. "That was Rohe's theory. If independently interviewed participants converged on a scene, he believed the reconstruction was approaching a shared autobiographical truth."

"Even if both were trained on the same fake scene," Elias said.

"Yes."

"That's circular."

"Yes."

Mara looked at Session Fourteen.

"Play it."

The recording loaded. Two separate immersion rooms appeared side by side.

Mara in one. June in the other.

They were being shown the same breakfast scene. Afterward, each was interviewed separately.

What color were the curtains?

"Yellow," both said.

Who spilled coffee?

"June's dad."

What did Helen say when Mara refused toast? Both girls repeated the same sentence within two words.

The confidence scores rose with each exposure. By the third session, their answers matched at eighty-nine percent.

A researcher wrote:

CONSENSUS ACHIEVED.

Elias felt sick. The system had created agreement, then cited agreement as proof.

On the final page of the protocol, he found the planned next phase. Family consensus deployment.

Real families would reconstruct disputed events together until their accounts converged. Mara whispered, "Did they do this?"

Lena did not answer quickly enough. Elias looked at the speaker.

"Lena."

Her voice was quiet.

"Yes."

The family-consensus files were worse when viewed in sequence. One session could be dismissed as a crude experiment. Twenty-three looked like a philosophy.

Elias read pre-session interviews first. Family members disagreed in ordinary ways. A daughter remembered her father leaving a hospital room before midnight; her brother insisted he stayed until morning. A husband remembered his wife wearing green at a wedding; she remembered blue. A mother believed her son apologized during a final argument; the son said he had not.

After repeated shared reconstructions, disagreement narrowed. Researchers celebrated.

What the summary charts hid was how narrowing happened. People did not simply remember better. They changed language. "I think" became "I remember." "Maybe" disappeared. Participants began quoting phrases from the reconstruction back to researchers as though those phrases had preceded the session.

Mara watched one interview until a woman said, "No, I'm sure now," and asked Ivo to pause.

"That's the moment."

"Which moment?"

"She thinks increased confidence means increased accuracy."

Elias knew the mistake intimately. Ivo plotted confidence against source fidelity. In several families, confidence rose while factual support stayed flat. In two, factual support decreased because the shared reconstruction favored one plausible variant and suppressed contradictory evidence.

"Consensus can be useful," Lena said. "Family therapy depends on finding shared narratives sometimes."

Mara looked at her. "Shared narrative is not shared memory."

"Agreed."

"Were they told the difference?"

Lena looked at the old consent language. "Not well enough." One case involved a birthday cake. Three siblings disagreed over whether their deceased mother had baked it herself. The reconstruction chose homemade because family photographs showed flour on the counter. After four sessions, all three remembered her baking.

Later, a receipt was found from a bakery. The mother had bought the cake.

No harm followed. Nobody's life changed. That triviality made Elias more uneasy than a dramatic example would have. The system could manufacture consensus around something nobody would ever investigate because it did not matter.

Most memories were like that. Small.

Unverified. Safe from correction because no one had reason to check.

Mara said, "If they got the cake wrong, how many things do we never catch?" No one answered.

The question was unbounded. Elias closed the file.

"We don't chase every cake."

Mara looked at him, surprised.

"My rule?"

"Your rule."

"This is becoming intolerable."

But she smiled.

Orpheus was colder than Elias expected. Not metaphorically. The data center had been built for machines that no longer generated enough heat to justify the cooling system, so the basement air sat at sixteen degrees and smelled faintly of dust and old cable insulation.

Celia handed out access badges with temporary numbers. "Nothing leaves physically. No personal devices. Every query logged. If anyone attempts to copy raw patient material, I will personally make your remaining week unpleasant."

Ivo held up both hands. "I have seen your pleasant weeks. Threat received."

Three sealed server cabinets stood behind a mesh partition. Their labels were mundane: environmental backup, legal hold, infrastructure archive. Kade had hidden sensitive provenance beneath categories nobody wanted to investigate.

The first boot took twenty minutes.

Mara paced.

"This is less cinematic than crimes against memory should be."

"Mostly loading bars," Ivo said. "That's how civilization ends."

When the archive opened, there was no friendly interface. Directories displayed cryptic identifiers and dependency tables. Ivo smiled with genuine affection.

"Ugly. Beautiful."

Elias sat beside him. The raw provenance store preserved source links that later Morrow systems had normalized away. Instead of saying FAMILY PRIOR: VENN, Orpheus could show the exact tape, timestamp, transcription transform, training use and downstream reconstruction sessions influenced by that fragment.

They searched the phrase from Gabriel's 2014 session.

The path appeared.

Original source: Martin Saye hospital audio.

First Copy transfer: bereavement reassurance set.

Recipient: Gabriel Venn.

Acceptance feedback: high emotional congruence.

Downstream use: 417 sessions.

No one spoke.

"Four hundred seventeen," Mara said eventually.

Ivo checked the scope. "Not necessarily exact dialogue. Could be weighted influence on pattern selection."

"Still."

They opened aggregate statistics, not patient identities. Venn-derived fragments appeared in thousands of early reconstructions. So did Saye, Mercer, Calder and dozens of other families. The system had built composite relational priors from people who never knew one another.

"First Copy wasn't one family," Elias said.

"It was a statistical neighborhood," Ivo replied. "A place where enough families overlapped that the software could borrow what worked."

Mara leaned over the screen. "And Origin?"

Ivo searched.

A dependency tree appeared beneath First Copy. At its root was ORIGIN-HJ.

Helen Journal.

The code was unmistakable.

Elias felt the room narrow.

"How much?"

Ivo opened aggregate weightings. Origin-HJ contributed not the majority of all content but a disproportionately high share of maternal dialogue priors. Helen's recorded journals had been early, abundant, emotionally varied and cleanly transcribed. The model learned from her because she was available.

Mara walked away from the terminal.

Lena followed but did not touch her.

"They made Mum a default," Mara said.

"An early one," Elias replied. "Not universal."

"I know what the numbers say."

He stopped speaking.

Further in the archive they found an internal study comparing direct-source accuracy against user acceptance. The conclusion was written with bureaucratic calm: strict provenance reduced reported emotional benefit in high-gap reconstructions. Cross-family completion increased narrative continuity.

One recommendation had been highlighted.

Consider acceptance as a primary outcome where factual recovery is impossible.

Celia read it twice. "There."

"The smoking gun you hate," Tess said over secure video.

Mara shook her head. "No. The mistake is older than one memo. This is just where somebody wrote it down."

That distinction mattered. Institutions loved singular bad decisions because singular decisions could be blamed on singular people. Orpheus showed a slower failure. Convenience, grief, interface design, research incentives, legal ambiguity and human hunger for coherent stories had accumulated until uncertainty itself became treated as a defect.

At 3:12 a.m., Ivo found the preserved natural-exposure study index.

Lena stared at the time on the monitor. Elias noticed and almost smiled at the old association.

"What?" she asked.

"Nothing."

"Progress."

"Be quiet."

For once, the numbers had become intolerable for ordinary reasons: they were tired, cold and hungry. Mara ordered everyone out for breakfast before they made a mistake.

Celia objected that the archive still contained urgent work.

Mara pointed at her. "You too."

Celia looked offended.

Then she smiled.

During the second night inside Orpheus, they found the training feedback loops. The discovery made First Copy less like a library and more like a social organism built from millions of small approvals.

Every session produced post-exposure signals. Did the user say the scene felt right? Did distress reduce? Did they correct dialogue? Did they revisit the scene voluntarily? Those responses affected future ranking.

"Not machine consciousness," Ivo said. "Just optimization."

Mara looked at the charts. "Optimization is doing plenty."

One graph showed a donor-derived maternal gesture increasing in selection probability after repeated high acceptance. Another showed a phrase being gradually abandoned because users found it too formal.

Human reaction curated the library.

The model was not inventing culture alone. People were voting without knowing they were voting.

"This is why the output became so persuasive," Elias said. "Not because the AI understood families. Because thousands of people taught it which borrowed behaviors felt believable."

Lena nodded. "And therapeutic teams treated believability as clinical usefulness."

Celia asked whether any individual user could reasonably have understood that their emotional response would shape future patients.

Ivo pulled up a consent clause.

Anonymized session performance may be used to improve service quality.

Mara stared. "I want to set page fourteen on fire."

"This is page nine," Ivo said.

"Progress."

They found one unexpected result. Some direct-source details decreased in ranking because users repeatedly rejected them. A father who had actually used a certain phrase sounded too harsh in reconstruction, so therapists favored softer variants. Historical truth could lose to emotional plausibility even when the evidence was stronger.

Elias felt something close to anger at mathematics, which he knew was irrational.

"The system was rewarded for making dead people more acceptable," he said.

"Sometimes," Lena replied. "Also for making scenes tolerable enough for patients to continue treatment."

Mara groaned. "And there is the complication again."

"It matters."

"I know. That's why I hate it."

They reviewed outcomes. Some patients benefited. Others later reported confusion. Most had no long-term follow-up capable of distinguishing relief from autobiographical adoption.

The missing data was as important as what existed.

At 4 a.m., Elias found a note from Peter Kade embedded in a test log:

If acceptance becomes truth weight, the model will eventually learn the family people wish they had.

Ivo whistled quietly.

Mara read it. "Did anyone listen?"

The next commit showed the warning field had been removed from the production dashboard six months later because clinicians found it distracting.

No villainous meeting. No order to deceive.

A warning became clutter.

Clutter disappeared.

The consequences remained.