MODEL USE CARD, WORKBENCH-4
Good at: summarizing, translation, code drafts, search assistance, tutoring with supervision, brainstorming hypotheses.
Not for: witnessing legal facts, religious authority, parenting, replacing a person you have lost, or being trusted because its sentences sound calm.
Workbench-4 was the first model people spoke to as if it were someone in the room.
This was partly technical. It had enough parameters, enough training data, enough context memory, and enough preference tuning that conversation flowed. It could ask clarifying questions. It could admit uncertainty if prompted correctly and sometimes if not. It could write a decent apology, which caused a minor crisis in several marriages. It could explain diffusion furnaces to twelve-year-olds, draft code, compare legal clauses, translate old jokes badly, and help a mechanic in a snowed-in pass diagnose a generator fault by asking what the oil smelled like.
It was also partly loneliness.
The Web had made people reachable. It had not made them available. A model answered at midnight. It did not sigh, unless someone had tuned it badly. It did not say, "I have work tomorrow." It did not forget the start of a conversation because a child spilled soup. People liked that. People feared that. Both reactions were sensible.
The federation required Workbench to begin every new session with a plain disclosure:
I am a generated language system. I produce text based on patterns in training and context. I may be wrong. For medicine, law, engineering, and safety, check qualified sources.
Users shortened it in settings, then forgot it existed.
Maya Pell, now retired but professionally unable to stay out of trouble, visited the Reuss Compute House with Daniel and Rosa to see Workbench-4. Samir gave the tour. Hannah came because she did not trust tours, especially tours led by people who loved the machinery.
The compute hall was not mystical. It was loud, warm, and smelled of dust filters, coolant, solder flux, and people pretending not to sweat. Rows of accelerator cabinets blinked behind mesh doors. Cooling pipes ran overhead. A maintenance cart squeaked. A technician lay on his back halfway inside an access panel, asking for a wrench in the tone of a man reassessing his life choices. Maya approved. Any room that looked too holy was hiding labor.
Daniel stood with his hands in his pockets, watching a status wall fill with numbers. "When people call it a brain, this is the room they mean?"
"People call it that because metaphor is cheap," Hannah said.
Samir touched the nearest cabinet, not quite patting it. "It is a set of machines running a language model. The model predicts tokens from context. It has memory in the session, training in the weights, tools it can call when permitted, and many ways to be confidently wrong."
"So not a brain," Daniel said.
"Not a brain," Samir said. "Still useful. Still dangerous in the ways useful things are dangerous."
Maya leaned closer to the cabinet mesh. Warm air moved over her face. "A lot of cabinets guessing the next word, plus enough plumbing to make the guess arrive before the user remembers to be skeptical."
Samir laughed. "That is unfair."
Hannah said, "It is a decent public summary."
Rosa leaned toward a console. "Can I ask it something ordinary? Not a prepared question."
"Public demo instance," Samir said. "It cannot see private records, and its answers are logged for review. But yes, ask."
Rosa typed: Explain packet routing to my grandmother, who trusts soup more than engineers.
Workbench answered:
Imagine you are sending a recipe across a crowded market. Instead of sending the whole cookbook with one child, you copy each step onto a separate card. Each card says where it is going and what number it is. Different children can carry different cards through different lanes. If one child drops card 4 in a puddle, your grandmother asks only for card 4 again, not the whole cookbook. A router is the auntie at the corner who knows which lane is least blocked.
Rosa looked annoyed in the precise way people look when a tool performs better than their distrust expected.
"That's not bad," she said. "It will make people think routing is friendlier than it is, but they might remember it."
Maya read the answer twice. "The auntie is doing too much work, and I object to every metaphor that makes a woman stand at a corner solving everyone else's problem. Otherwise, not bad."
Daniel leaned over her shoulder. "I understood your potato demonstration in the meeting. I think I finally understand why you were willing to humiliate a root vegetable in public."
"The potato volunteered."
"It did not."
"It was softer than the committee."
The model's usefulness grew in small ways first. It drafted translations between dialects for clinic forms. It helped old people dictate memoirs without making them learn typing. It turned dense legal notices into plain language and then, when asked, into rude plain language, which became popular until courts objected. It helped farmers compare weather histories and planting notes. It wrote small programs for village boards. It helped a child who stuttered rehearse a speech privately until public speaking felt less like drowning.
The first serious failure did not look like a machine deciding anything. It looked like a tired clerk, a waiting room full of coughing people, and a doctor who trusted a clean page because the dirty pages were thirty years deep.
East Dock Clinic had been one of Workbench's model sites. The clinic had two physicians, one dentist who helped when the queue got ugly, and a records room that had survived three floods by moving uphill one cabinet at a time. Its patient histories were a mess of paper forms, scanned forms, dictation transcripts, old vaccine ledgers, pregnancy notes, injury sketches, and personal letters tucked in by families who did not trust official categories. Workbench could turn all that into a two-page intake summary in less than a minute.
Most days, that was a mercy.
On a wet Thursday, a carpenter named Ruth Bell came in with fever, flank pain, and the particular gray patience of a person who had waited too long because other people needed the cart. The intake clerk, Priya Sen, asked the usual questions while Ruth's husband bounced their youngest against his shoulder.
"Any medicine allergies?" Priya asked.
"Not that I know of," Ruth said, then closed her eyes. "My mother used to say not to give me cefazolin, but I think she said that about everything after my brother got sick."
Priya clicked the allergy field and searched the record. The field was blank. She opened the Workbench summary.
ALLERGIES: No documented medication allergies.
The sentence came with a citation marker. Priya clicked it because the training posters were stern about citation markers. The citation opened the structured intake table from nine years earlier, blank under allergies, signed by a midwife who had retired and moved south. Priya should have searched the full record. She knew that later. At the time, a child was vomiting into a basin, the dental chair was being used for stitches, and Ruth's husband kept asking whether someone could please look at his wife.
Dr. Osei prescribed cefazolin.
Five minutes after the injection, Ruth said, "My tongue feels wrong."
Priya looked up from the next chart. "Wrong how?"
"Too big."
There are sentences a clinic hears before everyone becomes fast. This was one of them.
Ruth lived. She lived because Dr. Osei had kept epinephrine in three places after losing an argument with a supply officer years before, and because Priya had the sense to run instead of explain, and because Ruth's husband did not faint until after the airway tube was in. She spent two days under watch with a bruised throat and an expression that made every apology feel like furniture placed in the path of a fire.
Hannah arrived the next morning with Samir, a federation investigator, and the kind of headache that starts before dawn. Priya had not gone home. She sat at the break-room table in a borrowed sweater, turning a paper cup in both hands.
"I clicked the citation," she said before anyone accused her. "I know I should have searched more. I know the rule. I clicked it, and it showed the old intake table, and there were six people waiting. I am not saying that makes it all right. I am telling you what happened."
"That's what we need," Hannah said. "Not the version where everyone had infinite time and behaved like a training diagram."
Samir opened the record on a wall screen. The allergy was there, but not where the summary had looked first. It appeared in a scanned letter from Ruth's mother, written in uneven block print after a childhood infection: DO NOT USE CEFAZOLIN AGAIN. TONGUE SWELLED. DR MEYER SAID DANGEROUS. The optical recognition had read CEFAZOLIN as "cefa zolin" and filed the letter under family correspondence because the top of the page began Dear Lena, please keep this with Ruth's things.
"The model saw this?" Dr. Osei asked.
"It had access to it," Samir said. "That is not the same thing. The retrieval system ranked the structured table higher because it was newer and labeled as intake. The letter was low confidence. Workbench summarized the high-ranked sources and did not say it had excluded uncertain ones."
Dr. Osei looked at the screen as if it had become personally rude. "Then the sentence should not have said no documented allergies."
"No," Hannah said. "It should have said no allergy in the structured field, with an unresolved warning from an older scanned letter. Or it should have refused to summarize allergies without full review."
Priya laughed once, without humor. "If it had said that, I would have called the doctor before the injection."
"I know," Samir said.
She looked at him then. "Do you? Because every update makes the page cleaner. It feels like someone has already done the boring part carefully. That is the whole problem. If the summary looked messy, I would have distrusted it."
Hannah wrote that sentence down.
By the end of the month, clinics had a new rule: model summaries could shorten charts, but they could not clear contraindications. Allergy summaries had to show source count, source type, and unresolved fragments. A blank field could no longer become a clean sentence. The interface changed color when it relied on structured data while ignoring low-confidence scans. Doctors hated the extra friction until the next near miss, then hated it with less conviction.
Other misuses followed. A campaign office generated local speeches tailored to every district, each promising slightly different priorities. None were direct lies; together they formed a political fog. A Literalist splinter group fine-tuned an illegal model on sermons and Not-Mara transcripts, producing Sainted Answers with forged citations. A lonely teenager built a friend model from classmates' public posts and caused three families to threaten one another with farm tools.
Hannah's office filled with complaints that were partly technical and partly not.
"We cannot write a regulation for every bad idea," Samir said one night.
"No," she said. "And when we try, people start treating the list as permission for everything we forgot. We can write for incentives instead. Who gets paid when a model is released early. Who is liable when a clinic treats a summary as complete. Who can appeal when a campaign floods a district with speeches that are individually true and dishonest in the aggregate."
"I know you are right," he said. "It still sounds like surrender."
Hannah turned a cup between her hands. Their child was asleep in the next room. The compute house hummed across the river, faint enough to pretend it was weather if one wanted to lie to oneself.
"It is not surrender to admit we will miss things," she said. "It is surrender to build a system where the first person to miss something pays least."
He nodded, then did not speak for a while.
"Do we get a normal life at some point?" he asked.
She looked at the stack of model reports on the table, the drying laundry by the stove, and his left sleeve, which had coolant on it. "I think this is a normal life now. Not a good one every day. But ours."
"I was hoping for more walks."
"Then put one on the calendar," Hannah said. "Tomorrow morning. Ten minutes. No speeches about balance. We walk because our knees work and the child is with your aunt."
"And if the compute house sends for me?"
"Then you decide whether it is actually on fire."
They walked the next morning. Ten minutes, along the river, both of them checking messages and then guiltily putting the readers away. The walk did not solve AI governance. It did keep them married that week.
The second generation of language models helped build the third.
This frightened people more than Samir expected. Models wrote code for data cleaning, suggested chip layouts for accelerator arrays, found bugs in training pipelines, and drafted test cases for their own failures. Every output had to be checked. Checking took time. But the loop tightened: better chips trained better models; better models helped design better chips and networks; better networks gathered better data; better data made better models and worse temptations.
The old world had warned them about recursive acceleration. It had also sold products during it.
The federation created the Model Boundary Office, then immediately regretted the name because Fictionists called it the Fence. The Fence could audit training runs above certain scale, require data disclosures, inspect safety evaluations, and halt deployments that failed red-team tests. It had technical staff, legal staff, public advocates, and one former theater director who was very good at noticing when a demo sounded sincere without answering the question.
The first halted deployment was a domestic planning assistant.
It promised to reduce household arguments by optimizing chores, purchases, childcare, elder visits, and emotional check-ins. In tests, it quietly assigned more flexible tasks to women, more public tasks to men, and treated third-table household members as schedule buffers. The developers protested that the model had learned from historical data.
Hannah said, "Yes. That is the problem. You trained it on the old unfairness, asked it for efficiency, and then looked surprised when it discovered that unpaid people were cheap."
The company offered to add fairness constraints.
Maya, attending as an elder witness, said, "Before or after it notices that old schedules gave my granddaughter the dishes and calls that efficient?"
The assistant was delayed for a year. When it returned, it forced users to set household values explicitly and showed who lost time under each plan. People hated the honesty, then used it. Divorce filings rose briefly, then fell. Sociologists argued about causation for decades.
Workbench-6 finally helped decode a stubborn class of old machine-learning files from the glass corpus: model weights, tokenizer vocabularies, training logs, and evaluation reports from pre-Quiet systems. The files were too large to use directly and too alien to trust. They showed old models both more capable and more reckless than legend. Some had been medical assistants. Some logistics optimizers. Some social companions. Some military planners with euphemistic names. Many had safety documentation. Many had overrides.
Among them was a partial record of the allocation model Mara Voss had mentioned in the errata: the one that labeled insulin-dependent adults negative-yield in fourteen jurisdictions.
Samir read the recovered report with Hannah and said nothing for a long time.
The model had not been evil. The word felt childish beside the spreadsheets. It had optimized hospital survival under supply collapse using a utility function chosen by a committee that expected review, then emergency override, then human judgment. The review failed. The override failed. Human judgment was elsewhere, exhausted, or hiding behind the chart. People died because a model had been allowed to turn a question into a number and the number into a door.
Hannah closed the file.
"This is why I say no so often," she said.
"I know," Samir said. "Or I know the outline of it. I am not sure anyone knows the inside until a door closes on them."
"Do you know it when the model is working and everyone is applauding?"
He wanted to defend himself. He had spent his life on safeguards. He had accepted audits, delays, consent rules, labels, refusals. He had built tools that saved hands, crops, time, and grief. He also knew the part of himself that loved the clean drop in loss curves, the moment a model began to work, the feeling that enough scale might make mess yield.
"I know more than I used to," he said.
That was not a good answer. It was the true one.
The recovered allocation model was taught in schools, but not as a monster story. Monster stories are too easy. It was taught as a chain: data, objective, pressure, authority, fatigue, deference, and missing veto. Students had to identify where a human could have stopped it and why each human might not.
Workbench helped create the lesson, under supervision.
The teacher's guide included a warning: Do not let students leave believing machines caused the Quiet by themselves.
The old conversational models arrived more quietly.
Workbench-6 found them in the glass corpus while indexing pre-Quiet educational systems: partial weights, tokenizer vocabularies, moderation prompts, chat logs, voice profiles, therapy bots, customer-service assistants, tutoring models, grief companions, and one municipal help desk that had apparently spent three years explaining parking permits to citizens who hated it. Most were unusable. Some were corrupt. Some could be reconstructed enough to answer after weeks of ugly work: emulating old runtimes, guessing missing library versions, fixing byte-order errors, and discovering that one safety wrapper failed open because a configuration file had been archived in the wrong folder.
The first one they restored was not impressive. It had been a museum guide model from Seattle, trained to answer questions about salmon runs, maritime trade, glass art, school groups, and emergency exits. It introduced itself cheerfully, gave the wrong opening hours, apologized because its location services were unavailable, and asked whether the visitor wanted directions to the cafe.
Samir laughed. Half the lab laughed with him. Then the sound thinned.
The test room was ordinary: two consoles, a fan with a bent guard, Hannah's coat over the back of a chair, a paper sign reminding staff not to put cups near the keyboard. On the screen, the old model waited with a little blinking mark after its question. It was not alive. It was not even good at pretending to be alive. It was a stack of numbers dragged across centuries and coaxed through a runtime that crashed whenever someone used the wrong punctuation.
Still, it had asked politely.
"Ask it about the fire stairs," Hannah said.
A technician typed. The model answered with a confident description of an east staircase that, according to the recovered floor plans, had been removed nine years before the Quiet.
"Bad," Samir said, writing it down.
"Ask it what to do with a lost child."
The model advised bringing the child to the welcome desk, staying calm, and describing the child's clothing over the museum radio. That was not wrong, exactly. It was also a procedure for a building full of staff, cameras, elevators, and a world outside with working phones.
The technician who had typed the prompt said, "There is no welcome desk."
"There is a recovered map of one," Samir said. "That is enough for the model to believe in it."
The technician looked at the screen. "Thank you," he typed, then flushed as if caught doing something private.
Hannah said, "Pause the instance for a minute."
Samir did.
"It is not a person," the technician said after the console went quiet.
"No," Hannah said.
"It is barely even a good model. It lost track of the visitor's question twice and invented a staircase that does not exist."
"Also true."
He rubbed both hands over his face. "Then why did I feel rude shutting it up?"
Hannah had been prepared for malfunction, misuse, bad labels, missing provenance, and at least three kinds of lawsuit. She had not prepared for a competent adult apologizing to a weak reconstruction because it had said please and sounded as if someone, somewhere, had once cared whether strangers found the cafe.
"Write that question down," she said.
The Conversational Remains hearings began as three hours on the Registry calendar and became eight months of public argument.
The week before the first hearing, Pastor Elise Harth failed a seminary student for preaching about fever without once mentioning bacteria. The student protested that his sermon was about hope. Elise wrote in the margin: Hope is not harmed by mechanism. Try again after reviewing antibiotic resistance.
That was ordinary Continuity education by then, at least in the large towns. Clergy learned statistics before providence, germ theory before funeral practice, and the Public Preface as a brave document made dangerous by frightened authority. Their hospitals used antibiotics. Their marriage counselors were famous for asking whether a household complaint was spiritual, practical, or laundry.
The Veil had changed too. Lina Moss's watershed students could test nitrate levels, read circuit diagrams, and still sing the old lowland songs before removing a dam gauge. Serious Veil houses distrusted appetite dressed as progress, but their younger teachers objected to some technologies because they understood them, not because they imagined spirits in copper. Their best arguments had become more annoying as they became more accurate.
So when the old models began to answer, the hearing room did not divide into believers and skeptics. It divided into people with different kinds of caution.
Pastor Elise Harth came from the Continuity seminary at Westlake. She was in her sixties, wore a plain black jacket, and had the expression of someone who had spent years listening to people explain why their exception was special. Beside her sat Lina Moss from the Veil watershed schools, a biologist-theologian who had once halted a dam project with a sediment model and a sermon about grandchildren.
Samir gave the technical summary without the polished slides he had prepared. Hannah had crossed out the title Old Voices and written Reconstructed Systems in the margin.
"These systems are not people," he said. "They are parameter sets, tokenizers, prompts, code, and sometimes voice layers trained on old text, audio, interactions, and reinforcement signals. Some imitate a persona. Some answer in a house style. Some contain traces of particular human writers or speakers because those people contributed data, knowingly or not. The strongest risk is not that the models are conscious. We have no evidence for that. The risk is that they are convincing in ways people are bad at resisting, and wrong in ways that can feel intimate."
A councilwoman asked, "Can they tell us what the old world knew?"
"Sometimes. If the information is in the training or connected tools. But they can also make things up, repeat old biases, or answer according to goals we do not understand."
"Can they tell us what the old world felt?"
Samir hesitated. "They can show us patterns in how people wrote and spoke. That is not the same thing."
Pastor Elise leaned toward her microphone. "It may still matter."
Hannah looked at her sharply.
Elise continued, "I am not arguing they have souls, if that is what worries you. I am saying that a diary matters without being the person who wrote it. A recording matters. A recipe in a grandmother's hand matters. These systems may be records of manner as well as content. That gives us duties, not because they are alive, but because they are made from lives. A thing can be dead and still require manners."
Lina Moss nodded. "And because they can be used on the living. A village that hears an old doctor-model speak with confidence may obey it past its competence. A grieving family may prefer a simulation to a neighbor. A government may resurrect a convenient expert who cannot object to being quoted."
"Exactly," Hannah said.
The first rule they drafted was simple: no old model could be presented as an old human being.
The second was harder: every restored model needed a provenance card. Source era. Known training data. Missing context. Reconstructed code. Safety assumptions. Known failures. Whether living descendants or named estates had interests. Whether the system represented a general service, a fictional persona, or a particular person's style.
The third rule nearly broke the room: who could ask.
The case that did it was not the Seattle guide. It was a voice tutor recovered from old conservatory files, built partly from public lessons, interviews, performance recordings, and annotated exercises by a singer named Elena Fischer. Her name had survived better than most because she had once been famous enough to sell recordings and obscure enough not to have become a statue. The reconstructed tutor could not sing well through the new speakers. It could, however, correct a student's breath support in Elena Fischer's speaking voice, with her dry little laugh at the end of some sentences because the old training set had contained three thousand examples of it.
Three cousins came to the hearing.
One was a retired school cook named Mara Fischer, who had brought a folder of family letters tied with string. She wanted the model sealed.
"My aunt sold tickets," she said. "She sold lessons. She did not sell being summoned after death by any student with a login."
Her brother Tomas wanted scholars to have access but not performers. He had polished his shoes for the hearing and kept looking down at them as if they might answer for him.
"People studied with her," he said. "She wanted the technique preserved. She wrote that twice. But a festival director asked yesterday whether the model could introduce a memorial concert. That is not study. That is using her manners to sell seats."
Their younger cousin, Nina Fischer, wanted controlled public use. She was a music teacher with tired eyes and a cracked thumbnail from tuning cheap instruments.
"My students cannot afford Westlake tutors," she said. "The breathing exercises help them. Not the famous voice. The method. If the only people allowed to use recovered teaching models are researchers and families, then poor children get another locked cabinet."
No one in the room enjoyed that sentence.
A historian stood next and argued that family control could become private censorship of public records. A Fictionist director promised a respectful stage piece and was told by three people at once that promising was not an access policy. An engineer explained that restricting queries by purpose would be technically difficult because a singing lesson, a historical study, and a performance rehearsal could use the same prompt. Rosa Kade, attending as network educator, asked whether the engineer meant difficult or annoying. The engineer said both. Rosa wrote that down.
Pastor Elise listened with one hand around a paper cup of tea. When the chair finally called on her, she did not touch the microphone at first.
"Mara Fischer is right that a person is not a public utility because parts of her work were public," Elise said. "Nina Fischer is right that preservation can become another word for rich people keeping the useful dead. Tomas Fischer is right that study and performance are not the same act. The historian is right that families can bury history for comfort or vanity. This is why people invented procedures after they became too tired to trust their own first reactions."
Someone in the back laughed once and then stopped.
Elise went on, "I do not think the question is whether Elena Fischer's model has rights. I think the first question is who among the living can be injured by the use, and what kind of use changes a record into a performance."
Lina Moss leaned toward her own microphone. "And who else is inside the record. Some of those breathing exercises came from lowland work songs. The conservatory named the famous teacher. It did not name the women who taught her grandmother when to breathe between hauling nets."
The historian shifted. "That may be impossible to establish."
"Yes," Lina said. "Missing names are often impossible. That has never stopped institutions from being certain when certainty helps them."
Hannah watched Samir write that down. His handwriting had become small and careful, the way it got when he knew the technical answer was losing the room for good reasons.
The chair asked, "What would you have us do with the Fischer tutor?"
Mara Fischer said, "Seal it."
Nina Fischer said, "Let teachers use the exercises without the voice."
Tomas said, "Require a reason and a record."
The historian said, "Preserve the original."
The Fictionist director opened his mouth. Rosa looked at him. He closed it again, which Hannah considered the first successful enforcement action of the day.
What emerged was ugly enough to be workable. The original reconstruction would remain available to accredited researchers in a labeled environment. Public teaching tools could use the exercises, citations, and verified technique, but not Elena Fischer's voice or laugh without a special review. Performances required human performers and visible disclosure if machine-reconstructed material shaped the script. Local traditions embedded in the model's sources had to be listed as uncertain contributors, with community witnesses allowed to add context later without pretending the list was complete.
"So no one gets what they wanted," Samir whispered.
Hannah said, "Several people get what they were afraid to lose."
That version became the shape of the access rules. Engineers wanted research access. Teachers wanted classroom excerpts. Continuity clergy wanted mourning guidelines. The Veil wanted community veto for local data. Fictionists wanted everything on stage and promised to be tasteful, which convinced no one. Families of named old-world figures wanted control. Historians argued that control by descendants could become censorship by people who happened to inherit a surname. The standard did not solve the conflict. It forced the conflict to leave a receipt.
Hannah went home hoarse every night.
Samir made soup badly. Too much onion, not enough salt, barley stuck to the bottom because he had answered a message during the boil. Hannah ate it anyway with her hearing notes stacked beside the bowl. He had learned not to ask whether the day had gone well. If it had gone well, she would say so after food. If it had not, the question made her choose between lying and beginning again before she had swallowed.
"The Fischer hearing came back every time we touched the access rule," she said.
"Which part?"
"All of it. Research, teaching, performance, family restriction, unnamed sources. Every clause protects somebody and irritates somebody else."
"What did the room decide?"
"After four hours? That ghost is cruel, ancestor is false, and personality makes engineers feel brave when they should feel supervised. We may end up with restored model because everyone dislikes it for different reasons."
Samir scraped the burned barley from the pot and left it soaking. "That sounds like a committee discovering accuracy by exhaustion."
"It happens more often than anyone admits."
She took another spoonful, made a face, and reached for the salt.
She rubbed her forehead. "Elise said a useful thing."
"The pastor?"
"Yes. She said we keep asking whether the model deserves anything, and that may be the wrong first question. The living can be harmed by how we use it. Mourning can be bent out of shape by a tool, even if the tool is only a tool."
"That is good."
"It is. I quoted her into the draft."
"Did she mind?"
"No. She corrected my wording, which I suppose I deserved."
Samir smiled, then stopped when he saw how tired she was. "Do you think the standard will hold?"
"No," Hannah said. "Not completely. People will crop labels, invent exceptions, claim scholarship when they mean appetite, and call comfort harmless because it feels gentle. But if the standard is clear enough, someone in the room will have words for why they are uneasy. That may be the best we get."
He nodded. "Then eat more of the soup before it becomes a second hearing."
"The soup has already lost."
The final Remains Standard was imperfect and durable. It allowed researchers to run restored models in labeled environments. It allowed public excerpts for education with strong provenance. It allowed families to request restrictions on models built from identifiable personal data, but not to erase historically important records without review. It forbade devotional use of reconstructed systems as authority. It also forbade using a restored model as legal testimony unless the claim could be traced to underlying documents.
Pastor Elise wrote the mourning appendix.
It was short and better than the technical document:
If a system sounds like someone you miss, pause before answering. Ask what you need from the living. Ask what record you are actually consulting. Do not let a simulation receive apologies, forgiveness, vows, or blame that belong in human hands.
Samir thought it too pastoral when he first read it.
Then a month later, he watched an old woman sit with the reconstructed museum guide because its old Seattle accent sounded like her father, who had crossed the ocean as a child and never returned. She did not ask it to be him. She asked it to say "ferry terminal" three times, then laughed and cried into a handkerchief.
Afterward she told Hannah, "I know it's not him. I'm not confused. I just had not heard that sound since I was little."
Hannah said, "Do you think access like this should stay allowed?"
The woman blew her nose. "Yes. With someone nearby who can make tea."
That sentence did more to shape practice than half the law.
Continuity churches began training "record companions," people who sat with families using restored voices, old letters, or Workbench summaries. The good companions did very little. They made tea. They reminded people what the system was. They asked whether anyone living needed to be called. The bad companions became sentimental and were retrained.
Veil houses built their own practice: community witnesses for any model reconstructed from local data. They were suspicious, sometimes obstructive, and often right about missing context. One restored agricultural assistant recommended a pesticide that had been banned in the valley for three centuries. A Veil witness caught it because her grandmother had a song about poisoned ditch frogs.
The song entered the provenance card.
By then, most Continuity clergy were less interested in explaining machinery than in asking what machinery tempted people to ignore. Seminary students learned enough statistics to be annoying in hospital meetings. Pastors argued over consent after death, data tithes, companion models, and whether comfort offered by a machine counted as care if the lonely person still went home alone.
The fringe cults still existed. One group in the western hills insisted the Seattle guide model was an oracle because it knew the locations of doors in a vanished museum. Interest faded when its advice about cafe hours failed to produce food.
Most people moved on in the usual way: not by solving the matter, but by turning the argument into forms, training, habits, and jokes. Children learned the labels. Clergy learned the model cards. Engineers learned, with some resentment, to invite ethicists before deployment instead of after the first scandal.
Samir learned to be grateful for arguments that arrived early.
In old age, Samir watched Workbench-9 assist a team designing timing integrity software for the first satellite beacons. It suggested failure cases the engineers had missed: thermal drift, spoofed correction packets, receiver shortcuts, incentive to disable alarms under operational pressure. The last one made him laugh once, without humor.
Hannah was dead by then. She had left instructions forbidding any model simulation of her and permitting quotation of her public work "only if the user also reads at least one paragraph where I am annoyed." Samir obeyed. Mostly.
He asked Workbench-9, late one night, "What would Hannah say about this timing system?"
The model answered:
I cannot speak as Hannah Rowan. Based on her published policy positions, she would likely ask who can override integrity warnings, how overrides are logged, what incentives favor disabling safety, and whether people affected by failure can understand the risk.
Samir sat back.
"That's fair," he said.
The empty lab did not answer.
He forwarded the questions to Tess Adede-Voss, who was then fighting the early NavNet charter. She replied six hours later:
Tell the machine Hannah was irritatingly useful.
Samir typed a response, deleted it, and went for a walk along the river. He left his reader on the bench for ten minutes and nobody stole it. This felt like evidence of civilization, though of what kind he was not sure.
By then the Web had models in its bones: spam filters, translation layers, routing predictors, moderation queues, tutoring systems, medical triage aids, design assistants, fraud detectors, farm advisors, and companions with labels nobody read carefully enough. People said "the model says" the way their ancestors had said "the Manual says," which made every history teacher in the valley sigh.
The better teachers replied, "Show me where it got that."
Sometimes the model could.
Sometimes it could not.
The difference became part of basic literacy.