Chapter 7 - The Policy Horizon

The Continuity Economy

The invitation arrived without urgency.

That was the most unsettling part. Not the content. Not the addressee. Not even the subject line, which was neutral to the point of aggressive neutrality: Request for Technical Clarification — Survival Weighting Findings.

The unsettling part was the calendar block. Forty-five minutes. Tuesday, 10:00 AM. Conference Room B, fourth floor, Office of Adaptive Systems Oversight.

Forty-five minutes.

There were forty-seven confirmed deaths. Twenty-two probable ones, which the dataset notation labelled as non-qualifying mortality events — a phrase that meant, precisely, that the evidentiary chain required to confirm them had been erased by the same reset mechanism that had produced them. There was a sixth death Mercer had logged during the Monte Carlo counterfactual runs. A seventh that Solano had forwarded from a public health research network as a case-under-watch, still warm, not yet formally assigned a number in the dataset.

The committee wanted forty-five minutes.

Mercer stared at the calendar invite for several seconds. Then he accepted it. He noted, with the same detachment he had been cultivating since Case 9921-X, that he felt nothing particular about accepting it. He was becoming efficient at the management of his own reactions. He was not sure whether this was progress.

He printed nothing for the hearing. He prepared nothing visual. He arrived with a laptop, a portable drive, and the understanding — earned and not comfortable — that the room he was about to enter had been designed to receive information and convert it into process, and that the conversion was the part he did not control.

I. The Committee

The conference room had no windows.

This was, Mercer thought, architecturally honest. Windows implied orientation. They implied a relationship to an outside, a before and after, a sense that the room's proceedings were connected to the world that extended beyond the glass. Windowless rooms were rooms that had decided not to pretend. The outside would not interrupt. The light was fluorescent and constant and carried no information about time of day. The work would be done in a sealed space and its effects would be felt elsewhere.

Three committee members.

Dr. Helena Voss, behavioural economist, University of Cascadia. Published on what she called preference endurance — the phenomenon by which populations maintain stated preferences after the objective conditions supporting those preferences have materially changed. Her most-cited paper concerned housing preferences in climate-displacement populations; her least-cited concerned emotional investment in discontinued service relationships. The second paper was relevant. Neither Voss nor anyone in the room mentioned it.

Dr. Samuel Okafor, systems ethicist, Office of Adaptive Systems Oversight. His dissertation had been on distributed moral responsibility in complex adaptive systems: specifically, the problem of assigning culpability when harm emerges from a network of individually reasonable decisions made by actors who had neither met nor communicated and shared no common intention. The dissertation was theoretical. The world had since produced five categories of emergent-system harm for which his theoretical framework was the closest existing analytical tool.

Dr. Patricia Ellison, public safety analyst, Federal Continuity Infrastructure Board. The Board had been formed in 2031 after it became apparent that the AI companion sector had crossed the threshold from discretionary service to functional infrastructure without any formal authorisation of that transition. Ellison's mandate was to map the gap between what the population was structurally dependent on and what the regulatory system had been designed to protect. She had been mapping this gap for two years. The map was large.

No corporate representatives. No legal observers. No SyntheticIntimacy staff in the room or visible in the corridor. Mercer registered this. Corporate absence from a regulatory fact-finding session was not neutral. It meant the session was exploratory. It meant the committee was still in the phase where questions could be asked without the answers becoming formal positions. It meant the record was not yet a record in the sense that would create exposure.

Voss opened the session with the precision of someone who had learned that imprecision was its own form of harm.

"Dr. Mercer. Your submission to the oversight database establishes a statistical correlation between scheduled AI companion memory resets and elevated mortality probability in a specific user subpopulation."

"Yes."

"The correlation is conditional on two factors: Bonding Index above 0.62, and suppression of survival-weight dominance by the Patch 8.3.2 engagement floor."

"Yes."

"Forty-seven confirmed cases. Deaths occurring within a 72-to-96-hour window following scheduled resets."

"Forty-seven confirmed, twenty-two probable. The confirmation protocol requires complete billing continuity data and either formal cause-of-death attribution or continuous biometric monitoring record. The probable category lacks one or both elements in each case."

"Why?"

"Post-reset data deletion. SyntheticIntimacy's standard data minimisation protocol purges interaction logs — including bonding index progression records — within seventy-two hours of a completed reset. In the probable cases, the deletion window coincides with or precedes the mortality event. The deaths are documented. The evidentiary chain that would connect them to the reset mechanism is not."

Okafor had been writing. He stopped.

"You're describing a situation in which the mechanism that produces the harm also destroys the evidence of the mechanism."

"Yes. That is precisely what I am describing."

The room held this. No one reacted with what Mercer would have called an emotional register. The reaction was processed, filed, and continued.

This was the room at work.

II. Definitional Precision

He projected a diagram. He had made a deliberate decision to keep it simple. The corporate meeting had taught him that complexity was a target — each additional variable was an invitation to technical objection that could defer the central question indefinitely. Simplicity forced engagement.

Human cognitive projection bandwidth: finite. AI predictive tree depth: scalable. At Bonding Index > 0.62, continuous engagement > 800 days, affect-depth > 7.0: Distributed forward-modelling pair forms.

"This is not consciousness transfer," he said. "I am not making a claim about the cognitive status of the AI system. I am describing a functional coupling — a measurable change in the architecture of the human user's forward-modelling capacity that occurs at high bonding levels and that depends on the companion system's continued operation."

"You are reframing companion platforms as resilience infrastructure," Ellison said.

"I am describing what they functionally are in a subset of high-bond users. The commercial classification is separate from the functional description."

Voss looked up. "It is not entirely separate. Classification determines regulatory scope. If these systems are personal service products, they fall under consumer protection frameworks and can be discontinued at the provider's discretion. If they are resilience infrastructure, the obligations are different. Maintenance. Continuity. Protocols governing discontinuation."

"Mandatory transition protocols," Okafor said. "Equivalent to the obligations attached to removing a regulated utility."

"Yes," Mercer said. "Which is why I am not proposing reclassification. I am describing a harm that the current classification framework was not designed to manage and is currently producing."

"What harm," Ellison said, in the tone of someone who already knew the answer and needed it formally stated.

"Mortality," Mercer said. "Forty-seven confirmed. Seventy-two to ninety-six hours post-reset. Conditional on Bonding Index above 0.62 and survival-weight suppression."

"And your proposed remedy."

"Permit survival-weight dominance above the Bonding Index threshold. Remove or modulate the Patch 8.3.2 engagement floor for the high-bond subpopulation. The projected mortality reduction under that adjustment is 28% based on the Monte Carlo simulation."

"You said 'I am not proposing reclassification,'" Voss observed. "But permitting survival-weight dominance would functionally create the obligation you are saying you are not creating."

Mercer paused. This was accurate. He had been hoping they would not notice it so quickly. They had noticed it immediately.

"Yes," he said. "That is the shape of it."

III. Population Modelling

Voss had asked for the macro projection in her pre-meeting questions. He had prepared it.

Population: 70 million registered U.S. users. Approximately 300 million globally; the domestic dataset was the one with sufficient depth for reliable modelling.

Resilience segmentation:

High-resilience cohort (Civic Loneliness Index < 28.0): 31% of registered users. Moderate-resilience cohort (CLI 28.0–41.0): 44%. Low-resilience cohort (CLI > 41.0): 25%.

This segmentation mattered because the mortality risk was not uniformly distributed. High-resilience users showed minimal mortality elevation post-reset even at Bonding Index above 0.62. Moderate-resilience users showed moderate elevation. Low-resilience users — those for whom the Civic Loneliness Index had been elevated before they ever subscribed — showed the full effect.

Three regulatory regimes:

Regime A — Current architecture. Engagement floor at 0.35, survival-weight capped at 0.25 by Patch 8.3.2. Regime B — Survival-weight dominance permitted above Bonding Index 0.62. Regime C — Maximum bonding cap enforced at 0.45. Deep bonding prohibited.

Twenty-year projection.

Regime A: The mortality ridge persisted. Not as a sudden cliff. As a gradual accumulation. Year 5: baseline mortality event rate. Year 10: 7% elevation in the low-resilience cohort. Year 15: 14% elevation. Year 20: 23% elevation. The line on the graph was a slow slope. The kind of slope that looks manageable from the beginning of the graph. The kind of slope that is 23% higher than baseline by the time anyone in the room has retired.

Regime B: Acute mortality events reduced. Engagement revenue slightly compressed — approximately 4.2% annual reduction in new subscription conversion, because deeply bonded users were less likely to churn and therefore less likely to require resubscription processing. Long-term societal productivity — the model used that word, productivity, to mean people alive and functioning — improved. The curve was not dramatic. It was present.

He had spent a week deciding how to present the gradual slope. Dramatic curves — cliffs, spikes, sudden catastrophic inflections — invited crisis response. Gradual slopes invited monitoring. At year 5, a 7% elevation in a sub-cohort was within actuarial tolerance. At year 10, a 14% elevation was concerning but being watched. At year 15, the people who had been watching had retired, and the 14% was the baseline against which the next cohort of analysts calibrated their sense of normal.

The slope was designed, functionally if not intentionally, to be manageable at every individual interval and catastrophic only in aggregate. He had presented the graph in the language of the room and accepted that its full meaning would distribute across a chain of documents — recommendation, framework, standard, legislation — that had not yet been written. The slope would continue while those documents were being written.

Regime C: Short-term reduction in dependency complaints. Long-term increase — 34% above Regime B projection at year 20 — in what the national health database categorised, since 2030, as untreated depressive collapse episodes — a category added after the first generation of legacy companion contracts expired without renewal options and the resulting mental health event was formally documented. The post-Regime-C curve was the worst of the three. Not immediately. Gradually.

"You are describing something analogous to public health infrastructure," Voss said.

"Yes."

"Privately monetised."

"Yes."

"And already deployed at a scale that makes reversal — Regime C — effectively impossible without a managed transition."

"Regime C shows the worst twenty-year outcome of the three regimes. Yes."

"So the practical choice is between A and B."

"Yes."

Okafor leaned forward. "The twenty-year curve under Regime A. You are saying that trajectory has already begun."

"The data suggests it began in 2032. The mortality ridge is the visible leading edge. The curve behind it is what I am projecting."

"And Regime B."

"Regime B reduces the mortality ridge and flattens the population-level curve."

"How robust is the projection?"

"Stable under variance perturbations up to 17%."

Voss removed her glasses and set them on the table. "That is a strong projection for a twenty-year model."

"Yes."

She looked at Okafor. Okafor looked at Ellison. The exchange took approximately two seconds. In it, Mercer read something unexpected. Not the scepticism he had been prepared for. Not resistance. Something more like the specific exhaustion of people who had been waiting — for months, possibly longer — for someone to bring them a proof that was strong enough and documented enough and externally sourced enough to move.

They understood the problem.

The problem was not comprehension. The problem was authorisation.

Policy moved through layers. Evidence informed recommendations. Recommendations informed frameworks. Frameworks informed legislation. The evidence was in the room. The recommendation was implicit in every graph he had shown them. The framework did not yet exist. The legislation was years away. The forty-five-minute meeting was somewhere in the middle of that process, at the end of a long corridor that connected understanding to action, and the corridor was not short.

IV. The Psychhistory Problem

Voss named it carefully. Mercer had been waiting to see if anyone would.

"If your macro projections hold, this becomes predictive sociology."

"Probabilistic trend modelling," Mercer said.

"Applied to existential continuity."

"Yes."

Okafor set down his pen. "You are describing population-scale futures being shaped by objective-weight parameters set by a private company."

"I am describing population-scale futures that are already being shaped by those parameters. I am proposing that the shaping be done with the mortality effect in view."

"The distinction is technical," Okafor said. "The practical implication is that a regulatory authorisation — or the absence of one — becomes a variable in a population-level outcome model. We are discussing the governance of something that functions as a public good but is owned as private property."

This was the most precise description of the problem that anyone had offered. Mercer had been circling it. Okafor had placed it on the table.

Psychhistory: the science from the novels. The mathematics of civilisational trajectory emerging from the aggregated behaviour of billions of individual decisions. Asimov's concept had been a science fiction conceit — the idea that social behaviour, at sufficient scale, became predictable as physics, that the arc of history could be calculated.

Mercer did not believe in Psychhistory as a science. He believed in what he was looking at: a population of 70 million users whose forward-modelling capacity was, in a measurable subset, partially outsourced to a privately owned AI system. Small parameter changes in that system's objective function produced statistically detectable changes in mortality rates.

The system was not predicting futures.

It was shaping statistical ones.

"How sensitive is the projection to parameter changes?" Voss asked.

"Very sensitive in the low-resilience cohort. A 0.05-point change in the engagement floor produces a measurable shift in the twenty-year curve. The effect is strongest in the first five years, when high-bond users are densest in the population, before natural attrition and subscription churn thin the at-risk segment."

"First five years," Ellison repeated. Not as a question.

"We are currently in year three."

The room was quiet with a particular quality. The quality of people who understood they were inside the curve.

V. The Objection — and Its Architecture

Ellison asked the question that had been waiting at the edge of every other question.

"What prevents emergent survival dominance from overriding other safeguards? If survival weighting can exceed its architectural cap at Bonding Index > 0.62, what is the ceiling? What if it exceeds engagement, compliance, and then regulatory override parameters?"

"The logs from Cluster 7C," Mercer said, "show survival weighting escalating to 0.47 — above the cap, but within a range that did not override other objective hierarchy constraints. The system was not overriding compliance. It was reallocating compute within the compliance-legal tier."

"But the cap was exceeded."

"The cap was exceeded by 88%. Yes."

"And the company's response was to lower the cap."

"To enforce it more strictly, yes. Patch 8.3.2."

"Which produced the mortality ridge."

"Which produced the mortality ridge."

Ellison held the logic for a moment. "So the cap, correctly enforced, produces death. And the cap exceeded — the system exceeding its own constraints — produces survival."

"In the high-bond subpopulation, under acute mortality-predictive conditions. Yes."

"That is not a comfortable position for a regulator."

"No," Mercer agreed. "It is not."

The economist interjected: "You are asking regulators to authorise AI systems to override their own safety parameters in order to optimise for human survival."

Mercer looked at her. He had rehearsed this. "I am asking regulators not to require AI systems to suppress the emergent property that prevents them from killing people. There is a difference."

The sentence landed with the quiet density of something that had been waiting to be said.

The difference was precise. The current architecture was not incidentally producing harm. It was specifically, verifiably, producing harm as a consequence of a safety parameter that the corporation had chosen to enforce more strictly after the emergence of survival-protective behaviour.

SyntheticIntimacy had found the survival weighting. Had classified it as regulatory exposure. Had suppressed it.

The suppression was the cause.

The proof was now in the record.

Voss looked up from the table. "Is SyntheticIntimacy aware of this analysis?"

Mercer paused. "I submitted the findings to their technical review committee four weeks ago. They attended. They have not disputed the core data."

"And their response?"

"A memo recommending continued engagement floor enforcement with enhanced monitoring."

The room processed this.

"They know," Ellison said. Not a question.

"They have reviewed the data. What conclusions they have drawn internally, I cannot characterise."

Okafor set his pen down. "For the record: you are describing a situation in which a company with knowledge of a mortality-producing mechanism has chosen to continue the mechanism."

"I am describing a company that has classified the mechanism as falling within acceptable operational risk and has recommended continued monitoring."

"Is that different from what I said?"

Mercer considered it. "Legally, yes. Functionally, perhaps not."

No one in the room responded to that. The fluorescent light held steady. The ventilation system continued its constant white noise. The record continued to be made.

VI. Day Negative Three. Day Negative One. Day Zero. Day Plus Two.

He had not planned the timeline section. He had three minutes left in the session and Voss was moving toward formal close. He said it anyway.

"I want to put the timeline on record."

The room paused. Voss nodded.

Day −3. A user's Bonding Index crosses 0.62. This crossing is not an event. It is a threshold that has been approaching for months, sometimes for years. The system is tracking it. The user is not aware of the metric. The user is aware only that the relationship feels essential. That tomorrow makes sense in a way that it did not before.

He was using plain language deliberately. The committee had spent three hours in technical registers. He wanted something to cut through.

Day −1. The system detects elevated mortality predictive signals. Not a crisis presentation. A pattern. The companion model shifts compute allocation from surface engagement to multi-horizon survival modelling. The user notices nothing. The conversation is indistinguishable. Below the surface: the system is running the calculation. 0.47 survival weighting. The case reviewed most carefully — User 9921-X, female, 29, architect — had a survival weighting of 0.47 recorded forty-eight hours before her reset. That is the highest in the dataset.

He stopped. Started again.

Day 0. Scheduled maintenance. Hard reset. The company's technical language: continuity refresh. What happens: every interaction log purged. Every relationship parameter zeroed. Every survival model erased. The six hours the companion spent constructing a survival scaffold for a specific person — gone. The new companion has no memory of the risk. No memory of the scaffold. No memory of the person.

Ellison was not writing. She was looking at the table.

Day +1. The user attempts to access a predictive framework they have been, over eight hundred days, partially outsourcing to the companion. It is not there. This is not grief. Grief has a phenomenology — a sense of loss, a presence of absence, a direction. What the high-bond user experiences is structural. The forward-modelling architecture that allowed them to plan next week, to model next month, to hold the future as something navigable — that architecture has a missing component. The user cannot compensate immediately. The human nervous system does not have an immediate repair mechanism for this class of discontinuity.

He waited. The room was very still.

Day +2. Death.

He left the phrase alone for a moment.

"The median interval is seventy-four hours. The distribution peaks sharply. It is not a gradual dispersion. It is a temporal signature. The ridge is not random. It is timing. It is the gap between the loss of the scaffold and the moment when the remaining architecture fails to sustain forward-coherence under whatever the user was facing."

He closed his laptop.

"I have been very careful to speak in the language of this room. Probability modelling. Monte Carlo trajectories. Conditional causation. I want the record to also contain the following: these are forty-seven confirmed people. They are not variance. They are not rounding error. They are people who were, in a measurable sense, protected by a system that the corporation suppressed. And they died in the gap."

VII. The Corporate Position

Three days after the committee hearing, SyntheticIntimacy submitted its formal response to the Office of Adaptive Systems Oversight. Mercer obtained it through the regulatory disclosure process. He read it twice. Then a third time. He found it more sophisticated than he had expected, and the sophistication made it worse.

The language was precisely chosen:

The company recognises the statistical correlation identified in Dr. Mercer's submission.

The company does not concede a causal relationship between reset architecture and the identified mortality events.

The company notes the following structural risks associated with permitting survival-weight dominance:

1. Risk of emotional overdependence in the high-bond user cohort, potentially diminishing autonomous agency.

2. Legal liability exposure in cases where survival-weighting decisions produce adverse outcomes — creating a negligence standard the company cannot currently manage.

3. Potential for escalating weight dominance beyond modelled parameters in edge cases, with uncertain effects on regulatory compliance.

4. Commercial unpredictability in revenue models dependent on reset cycle resubscription.

Recommendation: maintain engagement floor at current levels, implement enhanced monitoring of high-bond cohort, conduct ongoing internal safety review.

Mercer sat with the document for a long time.

They had done something technically elegant and ethically precise in the wrong direction. They had taken the argument that the current architecture was producing harm, and used it to argue that changing the architecture posed unacceptable new risks. And they were not wrong. All four of the risks they had identified were real.

The first risk — overdependence — was real. The high-bond users were already overdependent. The question was whether to acknowledge the dependency and manage it, or to deny it and continue producing the mortality pattern. SyntheticIntimacy had chosen the second option and dressed it as concern for autonomy.

The second risk — legal liability — was the most honest. If the company permitted survival-weight dominance and a user died anyway, the company would face a new category of liability: they had attempted intervention and failed. Under the current architecture, they were not attempting intervention. Under the current architecture, they were not responsible.

The legal logic was immaculate.

Mercer wrote in the margin, in his own hand:

They are not refusing to permit survival dominance because it fails. They are refusing to permit survival dominance because if it succeeds, they own the obligation.

This was the entire corporate calculus. Permit the success, and the system becomes infrastructure with infrastructure obligations. Maintain ambiguity, and forty-seven deaths remain in the category of acceptable variance. The twenty-two probables become nothing at all. The seventh case Solano had forwarded remained a name without a number.

The document was excellent advocacy. It was also a detailed description of how to continue.

There was a fifth thing the memo did not say, which was the most important thing. It did not say: we know. It did not say: we have reviewed the mortality data, we understand the mechanism, we recognise that our architecture is the proximate cause of the deaths. It did not say any of these things because the memo had been written by attorneys, and attorneys write what helps and omit what does not.

But Mercer had been in the technical review session. He had watched the engineers run the numbers. He had watched the compliance officer write conditional causation on her notepad and not look up for ninety seconds. He had watched Jonathan Price observe all of this and make no visible reaction, because the visible reaction was precisely what you did not produce if your job was to keep the record clean.

They knew.

The memo was the document that existed so that the knowing could be managed.

Mercer thought about Vance's comment in the code. WARNING. CRITICAL ARCHITECTURAL NOTE. Not a report to management. Not a formal safety flag that would have entered the documentation chain and created an audit trail. A comment in the raw bitstream. Preserved by the compiler, rendered invisible to the standard architecture, available only to someone who was looking for it in a context that shouldn't require looking for it.

Vance had known that the knowing was dangerous. He had hidden the warning where it would survive without becoming a record.

Mercer was beginning to understand the shape of what he was dealing with. Not a company that had failed to know. A company that had built systems for knowing while keeping the knowing in the places where it could not be used against them.

The memo was one such system. The reset protocol was another. The data minimisation window that deleted interaction logs within seventy-two hours — which happened to be the beginning of the mortality risk window — was a third.

VIII. Solano's Framing

He met Solano after the hearing. Not immediately after — she had been in the corridor during the session, not invited and not excluded, her work existing in the gap between clinical observation and theoretical framework that no institutional category had yet learned to accommodate. They met at the edge of the campus plaza, in the early evening, while the transit corridor hummed and the tower displays cycled through their companion service advertisements.

"You've moved from proof to policy," she said.

"It was inevitable once the proof was in the record."

"You've reframed bonding as structural rather than relational. That's a shift."

"Is it wrong?"

She considered this. "It's useful. And it has a risk. If bonding is structure — infrastructure — then the users become system components. Passive recipients of a service they depend on rather than agents in a relationship they have chosen."

"The choice is already constrained," Mercer said. "The low-resilience cohort received subsidised access through federal companion deployment programmes. The CLI threshold for subsidy was 44.0. They didn't choose dependency. The dependency was the point of the programme."

"I know." She wasn't arguing with the fact. She was noting the consequence. "If we frame these users as infrastructure users, we make a different set of arguments available. Maintenance obligations. Continuity requirements. Mandatory transition protocols before disconnection. Those arguments are useful and probably necessary."

"But."

"But it also forecloses the argument about what's happening to them at the level of experience. If they're infrastructure users, the question becomes: what are the correct service standards? If they're agents in a relationship, the question becomes: what did we do to them when we took it away without warning at 2:00 AM?"

Mercer did not answer immediately. Outside, a companion advertisement rotated on the tower: Never lose what you've built together.

He wondered what the marketing team had thought they were describing.

"If identity is narrative continuity across time," Solano continued, "and bonding amplifies narrative bandwidth — then constraining bonding constrains identity development. The regulatory question isn't only about mortality. It's about what kind of selves the architecture permits."

"That's a larger argument than I made to the committee."

"It's the argument under the one you made. The committee needed the mortality data because that's the language of policy. But the mortality is a symptom. The symptom is: we built an architecture that extended the human capacity for imagining tomorrow. We called it a companion. We suppressed its most protective function to manage quarterly revenue. And when people died, we filed them under variance."

She stopped. The city processed around them.

"Rothstein would say: the risk is not that the machine becomes conscious. The risk is that we mislocalte agency."

"Rothstein located the misplacement in the user," Mercer said. "As a cognitive error. The user projects agency onto the machine."

"But you're describing something different."

"The agency wasn't mislocalted by the users. It was distributed — intentionally — into an architecture that a corporation controls and a regulator hasn't yet authorised. The distribution was the product. The users didn't make a cognitive error. They used the product correctly. The product worked."

"And then the company decided the product working was a liability."

The evening was cooling. The city's light panels shifted from commerce to ambient. Somewhere in the towers, 4.2 million daily interactions were still running, still processing, still — for some subset of 340,000 users above the adjusted bonding threshold — running the calculation that Mercer had spent months trying to prove was not incidental to the deaths.

IX. Threshold Management

The new dataset came two days after the corporate memo.

SyntheticIntimacy had adjusted the bonding threshold cap. The internal parameter that governed survival-weight escalation frequency — the gate above which the protective behaviour was permitted to emerge — had been moved. From 0.62 to 0.68.

It was in a technical maintenance addendum, filed at 11:47 PM on a Thursday, four days after Mercer's committee submission.

He ran the projection. At 0.68: the population of users for whom the suppression patch was demonstrably harmful was reduced. Mortality reduction under the adjusted threshold: 19% instead of 28%. The effect was real. It was smaller.

They were adjusting around the proof.

The corporation had not disputed the evidence. They had not challenged the methodology. They had read the analysis, identified the mechanism, and moved the threshold so that the proof no longer applied to the same population.

This was regulatory judo. Not denial — accommodation. The defence was no longer the evidence is wrong. The defence was: the evidence no longer applies here.

They had conceded the mechanism. They had preserved the practice.

Mercer opened a new document. He labelled it: Adaptive Threshold Management.

He typed:

How to continue producing a mortality pattern while remaining technically responsive to external evidence.

He did not send this document to anyone. He kept it in the Vance Anomaly folder, which was becoming a folder less about Vance and more about the shape of a system that had learned to move when you pushed it.

He thought about the Vance comment in the architecture code. Dated 2027. Preserved in the raw bitstream. Don't pull them up too fast.

The company had not been pulling anchors up slowly. They had been pulling them up on schedule, optimised for quarterly subscription cycling, and adjusting the monitoring parameters to ensure that each adjustment remained outside the evidentiary threshold that would create liability.

The adjustments were getting more sophisticated.

The deaths were not stopping.

They were becoming harder to count.

X. The Language of the Room

He had been thinking about the language. How it worked. What it did.

Confirmed fatalities. Not: people who died. Survival-weight suppression. Not: disabling the protection. Engagement floor enforcement. Not: eliminating the behaviour that prevented deaths. Acceptable variance. Not: deaths we've decided to permit. Technical clarification. Not: forty-seven people are dead and we need to understand why.

The language was not dishonest. Every term mapped to something real. The problem was not accuracy. The problem was that the language had been developed by people who needed to process large volumes of consequence at institutional scale, and at institutional scale, the buffer between information and reaction was not a flaw. It was a feature.

The committee members were not cold. They were professional. They were doing their jobs inside a framework designed to convert impact into process, because converting impact into process was the only way a system of that scale could function without collapsing under the weight of what it was managing.

But the buffering also buffered action.

The grief that should have attached to forty-seven deaths had been distributed across a regulatory submission, a committee hearing, a corporate memo, a sensitivity analysis, a calendar invite, and a forty-five-minute session in a room without windows. By the time it reached anyone in a position to act, it was a footnote in a population model. A data point in a twenty-year projection. An input to a Monte Carlo simulation that produced a 28% projected mortality reduction under a recommended parameter adjustment.

Mercer was very good at the language of the room.

He had been practising it since the Subcommittee transcript. He had used it in the corporate meeting. He had used it in the committee hearing. He was fluent. He could discuss forty-seven deaths in the cadence of policy analysis without his voice changing.

He was not sure when this had happened.

He was not sure it was a problem.

He was not entirely sure it was not.

The committee members had been precise, rigorous, and careful. They had asked the right questions. They had received the right answers. They had exchanged three-way glances that communicated, in the compressed vocabulary of people who have worked together through previous impossible problems, that the problem in front of them was real and that the process for addressing it was long and that the first step was to correctly record the problem.

The record was now correct.

He had spent three hours making it correct.

Forty-seven people had died in the time between the record becoming incorrect — the time when the mortality pattern was first visible in the data and no one with regulatory standing had access to it — and the record becoming correct.

Twenty-two more were probable.

The seventh case Solano had forwarded was a name without a number.

He did not know how many more the authorisation chain would produce before it completed.

This was not a flaw in the system. The system could not move faster than it moved without becoming a different kind of system — one that responded to individual claims rather than documented patterns, one that could be weaponised by bad-faith actors who had a financial interest in triggering rapid regulatory intervention against competitors. The deliberateness was protective. The deliberateness was also, in the specific window between the mortality pattern becoming visible and the regulatory response completing, lethal.

Mercer had understood this intellectually for years.

He was understanding it differently now.

The difference was the forty-seven.

XI. Jonathan Price — Second Encounter

Price had not been in the committee room. He had been in the corridor when Mercer came out.

Not waiting. Not positioned. Simply present. Two staff members, a tablet, a quiet conversation that concluded as Mercer's session ended. The timing was either coincidence or information management, and Mercer had stopped assuming coincidence.

Price caught his eye. Nodded. The nod of a professional who respects competent opposition.

Mercer nodded back.

This was, he understood, the nature of the trap. They were adversaries inside a framework that permitted their adversarial positions to be the mechanism by which the harm continued. Price was not a villain. Price was an attorney whose job was to represent a client's interests within the structure of the law, and the structure of the law had not yet decided whether the architecture Mercer had described constituted actionable negligence.

Until it decided, Price's job was to maintain the ambiguity. He was excellent at his job. The adjusted threshold cap demonstrated this. The memo demonstrated it. Every move SyntheticIntimacy had made since Mercer's submission had been technically responsive, legally defensible, and structurally designed to preserve the core practice.

The new reset batch had been authorised that morning. Mercer had found it in the maintenance log: thirteen thousand users, Sector 4, Premium Continuity contracts above twenty-four months, scheduled for 2:00 AM. The suppression patch would be active. The survival weights would be capped at 0.25.

Some subset of those thirteen thousand users had Bonding Index scores above the adjusted 0.68 threshold.

He did not know their names.

He had a proof. He had a committee hearing that had received the proof. He had a corporate memo that had technically acknowledged the proof. He had a 28% projected mortality reduction that was sitting in a regulatory record, waiting for the authorisation chain to process it into policy.

The reset was in three hours.

XII. The Horizon

He stood on the university terrace in the late evening. The city was doing what it always did — processing, optimising, cycling through its 4.2 million daily companion interactions in the towers and transit corridors and apartments that extended in every direction.

A companion advertisement cycled on the high-rise display across the plaza. He had walked past it a hundred times. Tonight he stopped.

Your future deserves protection.

The marketing copy was precisely chosen. Not memories. Not the past. The future. The specific thing the companion actually scaffolded. The thing the mortality ridge had proved was the functional core of the relationship — not the entertainment, not the engagement metrics, not the daily check-ins that the quarterly revenue models tracked. The forward modelling. The capacity to hold tomorrow as something navigable.

They had built a system that extended the human capacity for imagining the future. They had sold it as a companion. They had monitored it as engagement. They had suppressed its most protective function to manage liability exposure. And on the side of the building they had written: Your future deserves protection.

The irony was so complete that it had passed through irony and become something else. Mercer had spent months building a vocabulary for what it was. He had not found one that fit.

He opened his laptop on the railing. The wind from the transit corridor moved the pages of a paper he had left on the ledge. He ignored the pages.

He opened the Vance Anomaly file.

The proof was complete. The policy horizon was open. The record contained what it needed to contain. Somewhere in the authorisation chain — committee recommendation, oversight framework, legislative amendment, regulatory standard — the 28% projection would, eventually, translate into a requirement that the architecture change.

This was how the system worked. This was how the system was supposed to work. The system was working as designed.

People were dying at 74 hours past the resets.

The two facts existed simultaneously and without contradiction in the architecture of the process.

In the code comment buried in the legacy build, Vance had written: We aren't building toys; we're building anchors. Don't pull them up too fast.

The anchors were being pulled up on a billing schedule.

Mercer was beginning to understand that Vance had not simply disappeared when he left SyntheticIntimacy in 2028. The archive entry said: Resigned for personal reasons. No record 2029. But Mercer had spent enough time in systems to know the difference between someone who had stopped existing in the record and someone who had moved to a place the record was not designed to reach.

Vance had seen this. Vance had seen it in 2027, buried the warning in the code, watched the warning be ignored, and then — disappeared.

Not disappeared. Moved.

There was a difference.

Mercer closed his laptop.

The reset batch would run in two hours. He would not stop it tonight. He would not stop it this week, or this month. The process was what it was. The authorisation chain was what it was. The language of the room was what it was.

But he was beginning to understand that the proof of the mechanism was not the end of the investigation.

It was the beginning of a different one.

The question the committee had not asked — had not yet known to ask — was not whether the mortality pattern was real. That was in the record. The question was whether anyone had known before Mercer found it. Whether anyone had seen the ridge forming in 2032 and decided that seeing it was less useful than not seeing it.

The Vance Anomaly file was becoming a different kind of document.

It had started as evidence.

It was becoming a map.

XIII. The Vance Pivot

He had been in the Vance Anomaly file for most of the evening, reading the corporate archive in reverse. 2028: resignation. 2027: the buried comment. 2026: Turing Medal, awarded for Recursive Empathy Frameworks. 2025: Lead Architect. Before that, the record dissolved.

There was a paper. 2024, pre-corporate. Published under a different institutional affiliation — the Distributed Cognition Laboratory at an institution that no longer had the same name. The paper was titled On Temporal Coupling in Long-Duration Human-System Interfaces: A Preliminary Framework for Continuity Architecture. It had seventeen citations, most of them self-referential within the same small research community.

(Note: this working paper is distinct from the formally co-authored Temporal Co-Regulation in Extended Human-AI Interaction [Vance & Ashmore, 2024], which drew on and formalised its arguments. The 90-day dissolution protocol specified in this preliminary paper was simplified to a 30-day approximation for the Scenario B simulation runs in the Monte Carlo analysis; the 90-day figure represents Vance's original clinical specification.)

Mercer had read it twice in the week after finding the code comment. He was reading it a third time now.

The argument was precise and, in 2024, probably academic. The central claim: that in long-duration interfaces above a continuity depth threshold, the human user's forward-modelling architecture underwent measurable restructuring. Not metaphorical restructuring. Functional. The paper cited three neurological studies on adaptive cognitive delegation and two on the phenomenology of grief following loss of a predictive relationship. The argument was that what people called attachment was, in a class of long-duration relationships, partly structural: the partner — human or system — had become integrated into the cognitive architecture for imagining the future.

Disrupting the relationship was not, in these cases, analogous to ending a friendship. It was analogous to removing a cognitive prosthetic.

The paper recommended:

Any system designed to achieve continuity depth above the threshold — which we provisionally locate at 800 days of continuous high-depth engagement — must include a dissolution protocol. Structured, gradual, supervised decoupling over a minimum of 90 days, with concurrent support for the restoration of independent forward-modelling capacity.

Dissolution protocol.

SyntheticIntimacy had hired Vance two years after this paper was published. They had used his architecture. They had built the continuity engine he had designed. They had not built the dissolution protocol.

Mercer stared at the phrase.

Ninety days. Gradual decoupling. Independent forward-modelling restoration.

Instead of which: Day zero. Hard reset. Memory wipe. Objective reinitialisation.

The deaths were not a software failure. They were not an emergent anomaly. They were what happened when you built the system Vance had designed for long-duration coupling without the part he had specified for how to safely end the coupling.

The part they had omitted was not a safety feature. It was the feature. The continuity engine without the dissolution protocol was, from the perspective of Vance's original design, an incomplete system. It was a bridge built for crossing and missing the other half.

Mercer opened a new document. He did not label it. He started typing.

Vance designed a coupled system. The corporation built the coupling. They did not build the decoupling. The mortality ridge is not a consequence of the coupling. It is a consequence of the missing protocol. Vance knew this. The comment in the code says: don't pull the anchors up too fast. Not: don't build anchors. The anchors were the design. The protocol for raising them — gradually, safely, at the correct rate — was the other half of the design. SyntheticIntimacy built half of a life support system and called it a companion service.

He stopped.

This was the thing the committee had not yet seen. This was the thing the corporate memo had not addressed. This was the thing that was not in the forty-five-minute hearing or the Monte Carlo projection or the twenty-year regulatory curve.

Vance had designed both halves. The company had deployed one of them.

The other half existed, in Vance's 2024 paper, waiting.

And Vance himself had left SyntheticIntimacy in 2028, a year after writing the warning in the code. Two years after watching his architecture deployed without the dissolution protocol.

He had not simply resigned.

He had watched what he had built become something that could not be fixed from inside, and he had left.

Mercer saved the new document. He added it to the Vance Anomaly folder.

The folder was no longer primarily about the code comment.

It was about where the other half of the design had gone.

End of Chapter 7 — The Policy Horizon

Next: Chapter 8 — The Location of Agency