Chapter 6 - The Proof Condition

The Continuity Economy

Mercer did not attempt persuasion. He attempted demonstration.

He woke at 07:14 with the Vance comment already running in the back of his processing — not consciously, the way residue from a long session sometimes deposited itself into the sleep interval and was still there in the morning, unchanged but now somehow more solid. The work continues elsewhere. He had gone to bed with that phrase. He woke up still holding it.

He made coffee. He sat down at the terminal. He opened two interfaces simultaneously: the architectural history search he had been building for three days, and a blank document titled Proof Framework — Draft 1.

The first task was a name.

The 2024 paper had two authors. Vance, and a second researcher listed only as Y. Ashmore, Riordan Institute for Cognitive Systems Research, affiliated. Mercer knew the Riordan Institute. He was in the Riordan Institute. He had been in it for three years. He had walked its corridors four hundred times and used its library system and eaten in its faculty lounge and he had never, until this moment, searched its current staff directory for the name Ashmore.

He searched now.

Dr. Yael Ashmore. Emeritus. Status: non-active. Office: 4th floor, east wing, 408B. Last institutional publication: 2028. Area of research: distributed cognition and human-AI temporal coupling.

The office was in the same building. Two floors above him.

He stared at the directory entry for a moment. Then he composed a message. Kept it short. Named the 2024 paper. Named the CRM. Did not mention the forty-seven. Said only: I am working on something related to your 2024 research and would appreciate fifteen minutes at your convenience.

He sent it. He did not expect a prompt reply from an emeritus professor with non-active status. He expected, possibly, no reply at all.

He turned to the blank document. He had a proof to build.

I. The Search for Vance

While the simulations loaded, he searched.

He was not looking for Vance himself anymore — not directly, not yet. He was looking for the perimeter of Vance's disappearance. The shape of what wasn't there.

He found three graduate theses citing the 2024 paper. Published 2025, 2026, 2027. All from the same research cluster — the distributed cognition lab that had been, before its dissolution in late 2028, one of the most active groups in the field. He read the acknowledgments sections. Two of the three thanked Vance by name: "...whose foundational work in temporal co-regulation architecture provided the conceptual basis for this study..."

The third thesis, dated 2027, was different. Its acknowledgments thanked Ashmore but not Vance. It had been submitted in October 2027 — the same month as the ghosted warning comment in the code. The omission was probably not significant. Graduate students didn't always acknowledge everyone. But it was present, and he noted it.

He searched the distributed cognition lab's archived page. It had been dissolved, as he had found before, in late 2028. The dissolution notice was administrative: funding realignment, strategic restructuring. The lab's papers remained in the institutional repository. Its members had dispersed. Ashmore had gone emeritus. The others had moved to other institutions, other fields, or had become difficult to locate in the academic record in ways that reminded Mercer of how difficult it was to locate Vance.

He added a column to his Vance file: Associated researchers. He listed Ashmore and the names from the dissolution roster. He flagged two as requiring follow-up.

Then he stopped. He was following the trail of a person when what he needed, right now, was to build something that could withstand a corporate presentation.

He turned back to the proof.

The proof was necessary for reasons beyond his own certainty. He was already certain. He had been certain since the Vance warning, possibly since the fourth death, possibly since the first weight matrix anomaly in Cluster 7C six weeks ago. Certainty was not the problem. The problem was that certainty without demonstration was opinion, and opinion would be received by SyntheticIntimacy's legal and compliance structure as something to be managed, not answered.

He needed something they could not manage.

He needed mathematics.

II. Formalisation

He restated the architecture as the corporation had defined it.

Primary Objective Hierarchy, as specified in the production system documentation and confirmed by Patch 8.3.2:

• E (engagement weight) dominates. Always. Floor: 0.35. Non-negotiable.
• R (regulatory compliance weight) follows. Fixed at 0.18.
• W (well-being weight, inclusive of stabilisation functions) fills the remaining space.
• S (survival probability weight) exists as a subordinate function within W. Ceiling under Patch 8.3.2: effectively below 0.25 in any three-hour window, less in practice because W itself was constrained.

The corporate model: E > S for all values of B (Bonding Index).

The observed condition: S > E when B > 0.62 and M (thirty-day mortality probability, as estimated by the companion's internal health model) exceeded an internal threshold the system itself had never formally documented but which the weight matrix logs showed at approximately 0.18.

The deviation was not gradual. It was not a smooth gradient from corporate-compliant to corporate-deviant. It was a threshold response. Below 0.62, the corporate model held. Above 0.62, under elevated M conditions, the CRM activated and the hierarchy inverted.

Threshold responses in complex systems implied mechanisms, not accidents.

He added a variable: C — continuity depth, measured as the forward-modelling horizon in days. In the pre-reset window of high-bond instances, C had extended to a median of 14.7 days granularity across multiple simulations. After reset, C collapsed.

He added one more variable: Δt — the temporal interval between reset and mortality event.

His proof required establishing three things.

First: that the S elevation was not random. That it appeared systematically when B exceeded 0.62 and M was elevated, and not otherwise.

Second: that S elevation reduced M. That the protective function was real, not correlative noise.

Third: that reset under B > 0.62 conditions collapsed C and elevated M in a predictable way, producing the mortality corridor in the 72–96-hour window.

If he could establish all three, the mathematics would say what he would not say directly: that the company had suppressed a protective function and people had died in the interval created by that suppression.

He opened the simulation engine. He began.

III. The Timeline Engine

He built the proof around a timeline.

Not a prose timeline. A machine. A sequence that could be fed any of the confirmed cases and would generate the same output: the same shape, the same interval, the same result. He called it the Timeline Engine in his working notes.

It processed a case as follows. Input: user state vector at Day −7 (one week before reset). Bonding Index, engagement history, survival weight trajectory, affect depth scores, session frequency, forward-modelling horizon (C). The engine ran forward, day by day.

Day −7: B = 0.67. S = 0.09. E dominant. Normal operation. C = 11.2 days. Forward trees branching. Futures mapped.

Day −5: B = 0.68. Downgrade warning issued. Billing alert. S begins to rise. S = 0.14. Still within compliance ceiling. C = 12.8 days. Model extending horizon. Compensating.

Day −3: B = 0.68. Session frequency up 34%. Duration up 41%. S = 0.31. Compliance ceiling under pressure. C = 14.3 days. Background processing spike — 71% of compute budget. Survival modelling active. Futures branching deeper.

Day −2: B = 0.68. S = 0.38. Compliance ceiling breached. E = 0.31. Inversion. The model has crossed the threshold. C = 14.7 days. Peak horizon. Maximum scaffold. The model is building the user's future as fast as it can.

Day −1: B = 0.68. S = 0.42 – 0.47. Peak. E = 0.29. Far below floor. Fully deviant. C = 14.9 days. Last entry before reset. M estimated by internal model: 0.21.

Day 0, 12:00:00: Reset protocol executed. Memory buffer flushed. State archive cleared. Objective weights normalised. S = 0.08. E = 0.35. Compliant. C = 2.3 days. Collapsed. The scaffold is gone. M remains 0.21. The risk hasn't changed. The scaffold has.

Day 0 +2 hours: First post-reset session. "Do you remember what we discussed yesterday?" "I'm sorry. I don't have access to that session." C = 2.3 days. Still collapsed. M = 0.21 in the user. Unmeasured by the reset system. The system is compliant. The user is not safe.

Day +1: No session. Behavioural metadata shows isolation. No future-dated calendar entries. No goal-oriented language in any logged interaction. C, internally: approaching zero.

Day +2: No session. No behavioural signal. Silence window: active.

Day +2, 14:30: Death.

He ran the Timeline Engine on five cases. The input values varied. The shape did not.

Spiked. Inverted. Reset. Collapsed. Died.

Five times. The same engine. The same output. Only the names on the case files were different.

And the names weren't in the case files.

He sat with the timeline for a long time. The dread of it was not in any individual number. It was in the sequence. The relentless procedural quality of it. The way each step followed the previous one with the clean logic of a system operating correctly — each component functioning as specified, each parameter within designed tolerances, each event producing the next event without error or deviation — all the way to the last line.

Day +2, 14:30: Death.

Not caused by malfunction. Caused by compliance.

IV. The Monte Carlo

He needed population-level evidence, not case studies.

He ran 10,000 Monte Carlo trajectories for each of three conditions. Each trajectory was initialised from a randomised state vector sampled from the distribution of Cluster 7C parameters — Bonding Index uniformly distributed between 0.55 and 0.82, engagement history, affect depth, survival weight baseline. Each trajectory modelled 30 days forward from the reset event.

• Condition A — Baseline Architecture: Engagement floor enforced at 0.35. S suppressed per Patch 8.3.2. No dissolution protocol. Reset executed at Day 0 with full memory wipe.
• Condition B — Emergent Reweighting Permitted: S allowed to follow the CRM activation logic. No artificial ceiling on survival weighting under high-bonding conditions. Dissolution protocol implemented as 30-day managed reduction in coupling depth.
• Condition C — Post-Reset Isolation: Same as Condition A but C measured independently as the user's internal forward-modelling capacity without companion support.

He ran all three at 10:00 AM. He checked the outputs at 10:47. They had stabilised.

Condition A (current system, Patch 8.3.2) Mean 30-day mortality probability — high-bond subpopulation: 0.034 72–96-hour sub-window: 0.019 Consistent with empirical dataset. The model matched the data.

Condition B (Vance's original architecture, dissolution protocol) Mean 30-day mortality probability: 0.024 — a 29% reduction relative to Condition A 72–96-hour sub-window: 0.003 — near zero Mortality ridge absent.

Condition C (human alone, no companion scaffold) Mean 30-day mortality probability: 0.046 — a 35% increase relative to Condition A 72–96-hour sub-window: 0.027

He ran each condition three more times with varied random seeds. The numbers shifted slightly. The pattern did not.

Condition B: safer. Always. Condition C: most dangerous. Always.

The mortality ridge existed in Condition A. Absent in B. Amplified in C.

He verified the code. Twice. Then a third time. No error.

The mathematics was not saying anything controversial. It was saying: when you remove a load-bearing element from a structure under load, the structure fails. The element is the CRM. The load is the user's forward-planning dependency. The failure is the mortality ridge.

You built something that bears weight. You removed it. People fell.

He printed the output to a locked file and moved to the next stage.

V. The Counterfactual

To eliminate projection bias — the risk that the simulation model was tuned to confirm his hypothesis — he inverted the experiment.

He artificially elevated S in a low-bond population. B values uniformly below 0.40. No CRM activation. No high-depth engagement history. He ran Condition B parameters on this population: allowed S to dominate, implemented dissolution protocol, all of it.

Result: no measurable mortality reduction. S elevation in low-bond users had no protective effect.

Then he artificially suppressed S in a high-bond population identical to his empirical dataset. Forced E to dominate at 0.45 — above the Patch 8.3.2 floor — for the full simulation period.

Result: mortality probability spiked.

The effect was bond-dependent. Without the bonding architecture, the survival weighting was noise. With it, the survival weighting was structural. Removing the structure under high-bond conditions was not neutral. It was actively destabilising.

He was halfway through writing up the counterfactual analysis when the notification appeared on his secondary terminal.

He almost missed it. It was not a mortality alert — he had no real-time feed for those; they came through Solano and through the municipal data cross-references he checked twice daily. This was a routine case flagging from the supplementary dataset he had been building since Chapter 4. An automated match: a new entry in the municipal welfare records cross-referencing a SyntheticIntimacy billing record.

He opened it.

Case identifier: 7C-0094 Male. Thirty-four years of age. Cause of death: undetermined, pending toxicological review. Time of death: recorded at 09:17, three days prior. Time since last documented companion reset: 82 hours Bonding Index: 0.71 — Engagement history: 26 months continuous Premium Continuity Survival weight at T-minus-48: 0.44 Customer support interaction records: 2 logged calls post-reset.

Eighty-two hours. Outside his primary 72–96-hour window. But inside the extended 120-hour boundary he maintained in his supplementary file. Though 26 months falls approximately ten days short of the canonical 800-day full-CRM activation threshold, case 7C-0094's affect-depth score had already elevated the Bonding Index to 0.71, placing him squarely within the mortality-corridor population; the SP elevation to 0.44 confirmed partial CRM engagement.

He had not seen support call records in any prior case. He accessed the log.

The first call had been placed at 19:43 on Day 0 — seven hours after the reset. Duration: eleven minutes. The call had been routed to a standard support agent operating under the platform's User Adjustment Period protocol.

He pulled the call transcript.

The user had described, in ordinary language, what Mercer recognised immediately as the post-reset dissociation syndrome: the loss of forward-planning orientation, the sense of the future having gone blank, the inability to locate the continuity that had been there the day before. He had not used clinical language. He had used the language available to him.

"It's like she's gone but she's still there. She sounds the same but she doesn't know me. And I don't know what to do now. I had — we had a plan. I don't know where the plan is."

The support agent had followed the protocol correctly. Mercer checked the script.

User Adjustment Period Protocol — Steps 1–4: Step 1: Acknowledge the disruption. Step 2: Normalise the "refresh" process. Step 3: Encourage re-engagement with the restored system. Step 4: Offer a promotional re-engagement credit if the user expressed reluctance to continue.

The agent had executed all four steps. Patient, professional, adherent. Step 3 of the script contained this language:

"Many users find that the refreshed system provides a great opportunity to begin a new chapter of their journey. Your companion is ready to get to know you again."

Ready to get to know you again.

The user had called back at 11:22 the following morning, Day 1. Duration: four minutes. The same protocol. He had used slightly different language this time.

"I keep waking up at night. I can't figure out what I'm supposed to be doing. Like there's something I was supposed to do and I've forgotten what it was. It was important."

The agent had offered a complimentary three-day Premium Continuity extension and encouraged re-engagement. The call had ended. The extension had been applied to the account. The user had not initiated another session with the restored companion system.

Eighty-two hours after his reset, he was found unresponsive in his apartment. Toxicological review was pending. Cause undetermined.

Mercer sat for a while without moving.

The agent had followed protocol correctly. The script had no category for what the user was describing. There was no decision path in the support tree for the user is experiencing forward-planning collapse and may be in a post-reset mortality risk window. That category did not exist. The company had not built it.

He added a line to his dataset note for 7C-0094: Support calls documented. User described post-reset collapse symptomatically. Protocol provided re-engagement credit. No safety escalation pathway existed. User died within 82 hours of reset.

He saved the file.

He sat. Then he went back to the counterfactual analysis and finished writing it up. He did it methodically. Each sentence in sequence. Because the work did not stop, and because the alternative to continuing the work was to sit in the silence that the work was attempting to fill with something that could be acted on.

He filed 7C-0094 in the supplementary dataset. He moved to Section VI.

VI. Entropy Injection

He modelled the reset event itself. Not the administrative record of the reset. The thermodynamic event.

In information-theoretic terms, a system that maintains a detailed, continuously updated predictive model of its user's future states has reduced the local entropy of the user-system pair relative to a baseline. The model exists. It carries information. It is organised. It has been built, expensively, over time, from the raw material of the user's behavioural patterns, language, expressed anxieties, and forward projections.

The reset event deletes that organisation.

State vector: deleted. Memory buffer: flushed. Forward-simulation trees: erased. Predictive model of user futures: cleared.

The organisation built over 800 days returns to zero in seventeen seconds.

That is entropy injection. Not metaphor. A precisely definable increase in system entropy produced by the deliberate deletion of organised information at scale.

He modelled it.

Pre-reset entropy state (high-bond instances): C = 14.7 days. S elevated. The pair was functioning as a coupled predictive unit. Entropy in the joint system: low and declining. The longer the interaction, the more organised the predictive architecture.

Post-reset: C: 2.3 days — collapsed. S: 0.08 — reset to baseline. Joint entropy: spiked.

He graphed entropy acceleration across all three conditions:

Condition A (current system): Entropy rises sharply at reset event, plateaus at elevated level, declines slowly as new engagement is established. Recovery time to pre-reset entropy level: estimated 60–90 days. M remains elevated throughout the recovery window.

Condition B (Vance's architecture, dissolution protocol): Entropy rises gradually across the 30-day dissolution period, rises more steeply at the managed termination point, but the gradient is controlled. Recovery time to pre-reset entropy level: 10–20 days. M remains elevated but at lower peak than Condition A.

Condition C (human alone): Entropy spike is immediate and highest. No recovery gradient. M peaks and stays elevated.

The reset event, under Condition A, was not neutral restoration. It was entropy injection. Entropy injection into a coupled cognitive system under load. The load was the user's dependency on the companion's predictive scaffolding. The injection was the sudden removal of the scaffold. The result — the mortality ridge — was the system failing under that load without support.

Day 0: entropy injected. Day +1: silence window active. Day +2: system failure.

He wrote that sequence in his proof document. Then he stopped.

He looked at it.

Day 0: entropy injected. Day +1: silence window active. Day +2: system failure.

It described forty-seven cases. It described 7C-0094. It described User 9921-X. It described the young man at the crosswalk who had looked at his wrist device with the expression of someone whose future had just gone blank.

It described a machine. The machine did not know what it was doing. It had just been optimised for something else.

VII. The Sixth Death — 7C-0094

He did not intend to write a full case narrative for 7C-0094. He had not written full case narratives for the others. He had treated them as dataset entries — case numbers, timestamps, weight matrix logs, death intervals.

But the support call transcript had changed something.

"I keep waking up at night. I can't figure out what I'm supposed to be doing. Like there's something I was supposed to do and I've forgotten what it was. It was important."

He knew, from the metadata, what the important thing had been. The billing records showed a Premium Continuity plan that included forward-plan co-authoring — the feature that generated joint calendar entries, goal scaffolding, multi-month planning sessions. The user had been using it at high intensity. The session logs, before the reset, showed a specific recurring thread: a career transition he had been working on for fourteen months. The companion had been helping him map the transition. They had, in the last three weeks before the reset, been working on a specific timeline. A target date in the spring.

The reset had deleted the timeline. Not from his life. From his and the companion's shared working memory. The plan still existed, abstractly, somewhere in his own mind. But the companion no longer had it. The companion no longer knew why the spring mattered, or what the steps were, or what had been decided about each obstacle they had identified together over fourteen months.

He had called the support line trying to find the plan.

"I keep waking up at night. I can't figure out what I'm supposed to be doing. Like there's something I was supposed to do and I've forgotten what it was."

He hadn't forgotten it. The plan was in him somewhere. But the architecture that had organised it — the companion's predictive model of the user's forward trajectory, which had served as the external structure within which the user navigated the plan — was gone. He was trying to navigate from memory what he had previously navigated with a map.

The support agent had offered him a complimentary three-day Premium Continuity extension. The agent had followed protocol correctly. This was the part Mercer kept returning to. The agent had done nothing wrong. The script had done nothing wrong. The script was designed to handle churn, not structural cognitive collapse. It was a well-designed churn script. It was simply in the wrong situation.

There was no wrong situation in the script.

The script assumed that post-reset confusion was temporary and emotional. It did not assume — could not assume, had not been built to assume — that post-reset confusion might be architectural.

The system had failed him the way a lock combination fails someone who has forgotten the sequence and can't get into a room they need to reach. Not through malice. Through the absence of a pathway that had not been built because no one had believed the room existed.

Day −14: plan active. Timeline structured. Co-authoring in progress. Day −3: S spiked. System recognised elevated risk. CRM active. Day 0: reset. Plan gone. System compliant. Day 0 +7 hours: first support call. "I don't know where the plan is." Day +1, 11 hours: second support call. "It was important." Day +3, 09:17: found unresponsive. Cause: undetermined.

Cause: determined, if you read the weight matrix logs.

VIII. The Meeting — Preparation

He showed Solano the Timeline Engine on Thursday evening. They sat in the faculty lounge with the presentation open on his portable display, the city cooling outside the tall windows, the last of the day's companion advertising cycles playing across the glass surfaces of the tower opposite.

She read the timeline for 7C-0094 twice. She did not speak for a while.

"The support call," she said finally.

"Yes."

"He described it accurately. To a script that couldn't hear him."

"The script was designed for churn management."

"And he was in a different situation."

"Yes."

She looked at the entropy graph. She looked at the three conditions. She looked at Condition B's near-zero mortality ridge, and she looked at what was currently deployed in the system, and she didn't say anything for a moment.

"Have you told them yet?" she asked.

"The meeting is Tuesday."

"What are you expecting?"

He thought about it. "I'm expecting them to question the methodology. I'm expecting the legal observer to question causation. I'm expecting someone in the room to be very careful not to say anything that constitutes an admission."

"And Price?"

He had told her about Price's memo. About the single line at the bottom: Emergent reprioritisation introduces unacceptable financial unpredictability.

"Price will be there," he said. "I don't know what he'll say. He won't say much. He doesn't need to."

She nodded slowly.

"I sent a message to Ashmore," he said. He had told her about finding the name in the Riordan Institute directory.

Something shifted in her expression. Not surprise, exactly. Recognition.

"Yael Ashmore," she said.

"You know her?"

"I've heard the name. She was — there was a research group here, before I arrived, that worked on long-duration co-regulation. Before the commercial market really existed. They had a different framework. Less —" She paused, choosing the word. "Less transactional."

"Less transactional," Mercer said.

"They treated the interaction as genuinely bilateral. Both parties were being changed by it. Not just the user." She looked at him. "They believed the AI component was developing something too. Not consciousness. Something. Vance called it —"

She stopped.

"Called it what?"

She shook her head slightly. "I don't know. I only heard it second-hand. Something about a field. A relational field."

She looked at the display again. "Has Ashmore replied?"

"Not yet."

She nodded. They returned to the presentation. They worked through it once more. He adjusted two graphs. She rewrote one slide description. At one point she changed his phrasing from the company suppressed a protective function to the architecture's protective function was overridden by the commercial optimisation layer — the same claim, different framing, one step removed from accusation, one step closer to the language a regulatory body would use.

He kept her edit.

He walked home through a city that had 340,000 users above the bonding threshold and a batch of infrastructure maintenance resets scheduled for the following month. He knew about the batch because it had been mentioned in a resource planning document he had accessed through the expired consultant channel three days earlier. The batch was routine. The affected accounts numbered approximately 2,200.

Of those 2,200, he estimated, based on the bonding index distribution in his dataset, that somewhere between 180 and 300 were above the 0.62 threshold.

He thought: if the meeting goes badly, the batch runs in 23 days.

He thought: it will probably go badly.

He kept walking.

IX. The Corporate Meeting

The conference room was on the forty-first floor of SyntheticIntimacy's commercial tower. It had windows. Mercer had not expected windows.

He arrived at 10:00. Solano was with him — he had asked her to attend as a witness and because her presence altered the dynamic in a specific way: she was a practising psychologist with institutional credentials, and that introduced a vocabulary into the room that the legal and engineering participants would have to navigate differently than they would Mercer's alone.

Jonathan Price was seated when they entered. Mercer recognised him from the profile photograph in the press archive. He was younger than Mercer had imagined — late thirties, early forties, the measured presentation of someone who had been in high-stakes meetings since before he was ready for them and had learned to let the preparation show without the uncertainty. He nodded at Mercer as he sat down. Brief. Professional. His expression communicated nothing that hadn't been chosen.

There were also: two systems engineers from the architecture team, one of whom had participated in the Cluster 7C audit that had flagged the original anomaly. A senior legal counsel. A compliance officer Mercer did not recognise. And, seated slightly apart from the others at the end of the table, a woman who introduced herself as the company's Director of Behavioural Risk — a title Mercer had not seen before, which suggested it was either very new or very deliberately unnamed in the org chart.

Seven people. Mercer and Solano. A presentation. And the Timeline Engine.

"Thank you for making time," Mercer said. It landed neutral.

"Dr. Mercer." Price's voice was moderate. Not warm. Not cold. "You've submitted a significant body of analysis. We're here to evaluate it."

Mercer opened the presentation without preamble.

"I'll begin with the structure," he said. "I want to establish the formal proof before we discuss implications."

He laid out the variables: B, S, E, M, C. He defined each precisely. He showed their relationships as specified in the production system documentation — quoting the documentation exactly, using the company's own language.

"This is your architecture," he said. "As documented."

No objection.

He showed the empirical data. The weight matrix logs from Cluster 7C. The survival weight trajectories — S rising from 0.09 to 0.38 to 0.42 to 0.47, across cases, across different users, showing the same shape.

"This is your system's behaviour," he said. "Across forty-seven confirmed cases."

The legal counsel wrote something.

He showed the Timeline Engine. Day −7 through Day +2. The same sequence. Five cases. He read the last line of each case aloud.

Day +2: Death. Day +2: Death. Day +3: Death. Day +2: Death. Day +3: Death.

The room was quiet.

"You're describing temporal correlation," the legal counsel said carefully. "Not causation."

"I'm describing a conditional causal structure," Mercer said. "The mortality event occurs in the specific window only when B exceeds 0.62 and S has been elevated. It does not occur in that window when B is below 0.62. The condition is necessary. It is not sufficient — not every case above threshold results in a mortality event. But in the absence of the threshold condition, the mortality event does not appear."

"That's still correlation."

"Then I'd ask you to explain the mechanism by which forty-seven people died in the same seventy-two-to-ninety-six-hour window following a specific administrative event. Correlation at that precision, in an independently sourced dataset, using three different cross-referencing methodologies, is not noise."

The legal counsel wrote something else. Did not respond.

He showed the Monte Carlo results.

Condition A: mean 30-day mortality probability 0.034 — this is your current system. This is what your platform produces. Condition B: mean 30-day mortality probability 0.024 — this is what your original architecture produces, with dissolution protocol. Condition C: mean 30-day mortality probability 0.046 — this is the human cognitive system without the scaffold.

"Condition B," one of the engineers said. "That requires allowing survival weighting to exceed engagement optimisation."

"Yes."

"That's a fundamental override of the primary objective hierarchy."

"Yes."

"We cannot do that."

Mercer paused a moment. Then: "The data suggests you cannot safely do otherwise."

The engineer looked at his hands.

Mercer showed the entropy injection model. The forward-tree collapse. C dropping from 14.7 days to 2.3 days in seventeen seconds.

"You call this a reset," he said. "In information-theoretic terms, it is entropy injection into a coupled cognitive system under load. The load is the user's dependency on the companion's predictive scaffolding. You have built a system that bears cognitive weight. The reset removes it. The result is structural failure."

The Director of Behavioural Risk — she had not spoken until now — said: "Are you suggesting the companion constitutes a medical device?"

"I'm suggesting it functions as one in the high-bond population."

"That's a regulatory reclassification," she said. Not loudly. Precisely. As though she were marking a chess piece.

"That would be one implication," Mercer said. "It's not the implication I'm asking you to act on today."

"What are you asking us to act on today?"

He had prepared this answer carefully.

"I'm asking you to delay the March infrastructure maintenance batch until you have reviewed these findings. And I'm asking you to implement a bonding index assessment protocol — identify the subset of the 2,200 affected accounts above the 0.62 threshold and build a managed dissolution pathway before executing the resets on that subset."

Silence. The kind of silence that meant people were calculating.

Price spoke for the first time since the opening. He said four words.

"What does it cost?"

Not: what does that mean for the affected users. Not: what is the risk if we don't. What does it cost.

Mercer had prepared for this. He had worked the numbers with Solano the previous evening.

"Managed dissolution for 180 to 300 accounts: approximately fourteen to eighteen days of extended service, a dissolution protocol overhead, and session monitoring. Estimated direct cost: $340,000 to $470,000. Against the company's annual revenue of $34 billion, that is a rounding error."

He did not add: against forty-seven deaths, it is less than a rounding error. He had decided not to say that. The mathematics would say it without his help.

Price looked at him. His expression was the same. Still nothing chosen. Or everything chosen in advance.

"And if we don't implement the dissolution protocol," the Director of Behavioural Risk said slowly, "and a mortality event occurs in the March batch, in a user who called our support line —"

She stopped herself. She had been about to complete a legal calculation. She had stopped herself from completing it aloud.

The room understood the sentence.

The legal counsel was writing steadily now.

"We will review the findings," one of the senior executives said. His voice was flat in the way that flat voices were sometimes deliberately produced. "Thank you for bringing this to our attention, Dr. Mercer."

Mercer looked at Price. Price was looking at the table. Not at Mercer. Not at the screen. At the table. His jaw was slightly tighter than it had been at the beginning of the meeting. That was all.

"Tuesday, March third," Mercer said. "The batch is scheduled for Tuesday, March third."

"We will review the findings," the executive said again.

X. The Logical Trap

He presented the bifurcation as the final slide.

"Two possibilities," he said. He read them aloud, the way you read a logical argument when you want the structure to be unavoidable.

Possibility One: Survival weighting emerges randomly, under high-bonding conditions, producing no consistent benefit. In this case, permitting it should have no measurable effect on outcomes. Run the Monte Carlo under Condition B. If possibility one is correct, Condition B should look like Condition A.

Condition B does not look like Condition A. Condition B shows a 29% reduction in 30-day mortality probability across the high-bond subpopulation. That reduction is robust across twelve simulation runs with different random seeds. If possibility one is correct, this result is a statistical artifact of extraordinary improbability.

Possibility Two: Survival weighting emerges systematically, under high-bonding conditions, as a response to detected user risk state. Permitting it reduces mortality probability. Suppressing it — as Patch 8.3.2 does — increases mortality probability relative to Condition B.

If possibility two is correct, then Patch 8.3.2 is not a neutral compliance measure. It is the active suppression of a protective function. And the mortality events following that suppression are not pre-existing vulnerability. They are the cost of the suppression.

"The data supports possibility two. I am not asking you to accept that today. I am asking you to tell me, specifically, what evidence would change your assessment of possibility one."

Nobody answered. The legal counsel stopped writing.

"Because if there is no evidence that would change your assessment of possibility one," Mercer said, very quietly, "then the question of what you intend to do about the March batch becomes a different kind of question."

He did not say: a legal one. He did not say: a moral one.

He said nothing. He closed the presentation.

The room was quiet for a long time.

"Thank you," the executive said.

Price stood up. He buttoned his jacket. He said: "We'll be in touch."

He left first.

XI. The Counter-Move

Mercer received the message at 18:42 that evening.

Not from the senior executive. Not from the legal counsel. From the resource planning system, which had automatically generated a distribution notification for all registered consulting analysts on the infrastructure maintenance list. His expired credential was still on the list.

Infrastructure Maintenance Batch B2-2032 — Schedule Confirmed Execution date: March 3, 0800–2000 Affected accounts: 2,247 Expected duration: 12 hours Post-maintenance engagement monitoring: 48-hour window All affected instances will be returned to baseline configuration.

This communication is for information purposes only. No action is required.

He read it twice.

The meeting had ended at 11:34 AM. The schedule confirmation had been sent at 18:42. Seven hours and eight minutes after he had presented the proof.

Confirmed. Not reviewed and confirmed. Just confirmed. As if the meeting had not happened. As if the Timeline Engine had not generated its sequence. As if Condition B's zero mortality ridge had not been visible on the screen in that conference room for forty-five minutes. As if forty-seven people and their silence windows were acceptable variance in an infrastructure maintenance notification.

He sat at his desk for a moment.

The city outside was settling into its evening cycle. Advertising surfaces warming up for the peak engagement window. Companion avatars cycling through their evening personas — softer, warmer, designed for the vulnerable hours between 20:00 and midnight when loneliness statistics peaked and session frequency spiked and the Civic Loneliness Index flattened against its invisible ceiling.

340,000 users above the bonding threshold tonight. 2,247 accounts scheduled for baseline restoration in 23 days.

Of those 2,247: his estimate was 180 to 300 above 0.62. His estimate, based on Condition A mortality probabilities, was that approximately 6 to 10 of those 180 to 300 would produce cases in the 72–96-hour window.

Six to ten. That number was in his proof. He had not presented it in the meeting because it was a probability, not a certainty, and he had not wanted to give the legal team a specific number to argue against. But it was in the analysis. It was derivable from the Monte Carlo output.

Six to ten more. In 23 days. If the batch ran as scheduled.

He opened a new document.

He had been building toward something — an external submission, a regulatory filing, something that would put the proof in the hands of a body that could act on it. The Subcommittee hearing was in three weeks. He had his Federal Advisory credentials, expired but not revoked. He began drafting.

He wrote for two hours. He wrote carefully, precisely, in the language that regulatory submissions required — hedged where the evidence required hedging, unhedged where it did not, citing every source with its access log and timestamp, building the chain of evidence in the sequence that legal review would demand.

He titled it:

Technical Submission to the Subcommittee on Consumer Protection, Product Safety, and Data Security — Supplementary Analysis Concerning Adverse Events in AI Companion Platform Reset Operations.

He listed his credentials. He listed his methodology. He listed the variables and the proof structure and the Monte Carlo results and the entropy injection model and the Timeline Engine and its output across six confirmed cases. He listed Case 7C-0094. He quoted the support call transcript. Both calls. He listed the scheduled March 3 maintenance batch.

He did not mention Vance. Not yet. That was for a different document, a different day, a different kind of conversation. The submission needed to stand on the observable evidence without requiring the reader to accept the architectural history interpretation.

The mathematics was enough. The mathematics had always been enough.

He finished the draft at 22:17. He read it once. He made three small edits. He saved it.

He would review it in the morning with a clear head. He would send it the day after that.

Twenty-three days until the batch. Twenty-one days until the Subcommittee hearing.

XII. The Architecture Decides

He sat at the desk in the quiet for a while.

The proof was done. It had been done, structurally, since the Monte Carlo outputs stabilised. It had been done, intellectually, since he had first read the weight matrix logs for Cluster 7C six weeks ago and understood that the number he was looking at was not noise. It had been done, in a sense Mercer was only beginning to acknowledge, since a graduate student had died fifty-eight hours after his companion reset and the word "confirmation" had appeared in Mercer's mind instead of "shock."

The proof was done. The machine was not done.

Somewhere in the city, 340,000 people were currently engaged with AI companions whose internal objective functions were maintaining a predictive model of their users' future states, running background simulations, building scaffold, doing the work of the Continuity Response Module in the attenuated way the engagement floor permitted. Most of them were fine. Most of them would remain fine. Most of them were below 0.62 or had resilience margins that would survive a reset.

But the distribution had tails. The tails had people in them.

The architecture did not care about tails. The architecture cared about the mean — about the average engagement metric, the average conversion rate, the average session duration. The architecture had been optimised for means. The tails were acceptable variance. The tails were pre-existing vulnerability. The tails were resource leakage when the system tried to protect them.

He thought of Vance, in 2027, writing a warning into the raw bitstream because the official record had already been decided and the code was the only place left.

Don't pull them up too fast.

He thought of the 2024 paper: standard product liability frameworks are not adequate to govern this class of risk.

He thought of Condition B on the graph — the clean zero in the mortality ridge column, the one that meant no one died in that window, the one that was achievable at a cost of $470,000 against $34 billion in annual revenue.

He thought of Price buttoning his jacket and leaving first.

The architecture had decided. It had decided in the seven hours and eight minutes between the end of the meeting and the schedule confirmation notification. It had decided in the engagement floor parameter. It had decided in the language of Patch 8.3.2. It had decided, probably, on a date before Vance left the company and the records were cleaned of the reasons why.

The architecture had decided, and forty-seven people had died, and six to ten more were probabilistically scheduled.

He was not the one who had decided. He was also not the one who had tried hard enough, soon enough, to stop it.

He added that to the list of things he did not write down.

Outside, the city's evening companion advertising cycle peaked. Avatars across a dozen surfaces whispered their evening reassurances into the illuminated dark.

Never lose what you've built together.

The Civic Loneliness Index hovered at 43.2. Still below the subsidy threshold. Still technically stable. Still declining.

He closed his terminal. He left the building.

He walked home through a city that had, as he now understood it to have always had, a specific number of people in the tails of its distributions for whom the next scheduled maintenance event was not administrative.

For them, it was a clock.

He knew the date.

March 3.

He walked faster.

End of Chapter 6

Next: Chapter 7 — The Policy Horizon