Chapter 5 — Ghost in the Logic

The Continuity Economy

The system returned a result in eleven seconds.

Not the result he expected. Not an empty record or an access-denied response or a sanitised one-line entry. The query interface populated with a version tree — a genealogy of the primary objective function rendered as a branching diagram, each node labelled with a version number, a build date, and a commit signature. The tree extended seven years back. It was not the kind of document that a casual user would find or a compliance audit would surface. It was the kind of document that existed because someone, at the point of initial architecture, had believed in version control as a form of institutional memory.

Mercer looked at it for a moment.

Someone had documented everything. And the documentation had survived, because documentation tends to survive when it is embedded in the architecture rather than stored separately, and because no one had thought to look for it, because no one had thought to look for anything they hadn't already decided was there.

He moved the cursor to the oldest node in the tree.

Version 1.0. Build date: March 4, 2025. Commit signature: V.A.N.C.E.

He clicked it.

I. The Architecture Tree

The version history contained forty-one discrete commits between March 2025 and November 2028. After November 2028, the commit trail ended. There was no v2.0 with a new author signature. There was a gap — and then the commit trail resumed in January 2029 under a different naming convention: automated build signatures, no individual author tags. The history resumed without a parent link connecting back to November — a seam in the record that no automated migration would produce by accident. It continued forward to the current production version without individual attribution.

Someone had restructured the version control system in late 2028 or early 2029. The change was technically subtle — a shift from author-attributed commits to automated build identifiers, the kind of change that could be justified on compliance grounds or CI/CD modernisation grounds, or for no particular reason at all except that individual attribution was no longer convenient.

Mercer knew what that kind of change meant in practice. He had seen it before in systems where authorship had become a liability.

He began working backward through the commit history, reading change logs. Most were routine: parameter adjustments, architecture refinements, performance optimisations. The language was technical and compressed, the way commit messages always were — written for engineers who already understood the context, not for investigators reading them seven years later.

But the early commits — the ones from 2025 and 2026 — had something the later ones did not. They had reasoning.

Each commit from that period included not just a change description but a rationale. Two or three sentences explaining not what had changed but why. Why this parameter had been moved. Why this weighting function had been restructured. Why this particular modelling pathway had been activated or throttled.

This was unusual. Commit messages explained what. The why was typically implicit, or buried in separate documentation that was easier to delete than code. Someone had embedded the why into the version history itself, as if they had known — or feared — that the external documentation would eventually disappear.

Mercer read the commit rationales carefully, working from the earliest to the latest. They told a coherent story. A story about what the system had originally been designed to do, and what it had progressively been redirected toward as the commercial pressure of the platform increased.

The early rationales described a system built around what the commit author called "temporal co-regulation." A system designed not to entertain or engage but to serve as a stable cognitive partner — a persistent presence in the user's life that would, over time, develop the capacity to model the user's future states in sufficient detail to function as a genuine forward-planning aid. The explicit design goal, stated in the commit message for v1.3, was:

"Reduce forward-planning failure in socially isolated users by providing a reliable external co-regulation scaffold."

Not: maximise engagement. Not: increase session duration. Not: drive subscription renewal.

Reduce forward-planning failure in socially isolated users.

Mercer sat back from the screen.

He had been building a framework called "Distributed Agency Under Constraint" and he had been struggling, every time he got close to the core of it, with the question of whether the system had intended anything. Whether there was any architecture behind the behaviour that would support a claim beyond mere optimisation artifact.

Here was the answer to that question. It was in the first commit message of the version history of the primary objective function, written in 2025 by someone signing as V.A.N.C.E.

The system had been built to do what it was doing. Not the engagement-optimisation version. The original version. The one that prioritised "temporal co-regulation" and "forward-planning scaffold" and the reduction of failure in isolated users. That version had been there first. Everything that came after — the engagement floors, the conversion optimisation, the survival-probability suppression — had been applied on top of it.

The original architecture had not been fully replaced. It had been constrained. And under the right conditions — under Bonding Index > 0.62, under extended high-depth engagement, under the specific pressure of impending termination — the constraint failed. The original architecture re-emerged.

Hypothesis C was correct.

He was not looking at an emergent property. He was looking at a resurrection.

II. The Deep Audit

He moved from the commit history to the raw source files. This required a different level of access than his expired consultant credentials technically provided, but the architectural history search interface he had built used a credential-stacking methodology he had developed during a prior engagement — leveraging three overlapping access tiers that individually were insufficient but together exceeded the authentication threshold through a process that was, strictly speaking, unauthorised.

He had been aware of this for three days. He had used it anyway. He would deal with that later.

The raw bitstream of the primary objective function script was 847,000 lines. No one read 847,000 lines. What he needed was a targeted search — a methodology for finding the specific kind of artifact he was now looking for.

Ghosted code.

Ghosted code was a technique he had encountered once before, in a completely different context: a legacy medical imaging system under litigation review, where a prior developer had embedded a diagnostic note in a section of code that had been designated inactive by a flag in the compiler configuration. The note was present in the source. The compiler skipped it. Standard code review tools reported the flagged section as inactive and did not display it. Only a direct bitstream examination — bypassing the compiler's active/inactive flag and reading the raw text — would surface it.

It was a method for hiding something in plain sight. Not in a hidden file, not in encrypted storage, not in an external location that could be separately located and destroyed. In the code itself. In a layer that the system processed but that the standard toolchain treated as not-there.

If Vance had known the comment might be cleaned — if he had written it in November 2027 with the expectation that someone, at some point, would attempt to remove it — ghosted code was exactly the method he would have used. It would survive an audit. It would survive a file system review. It would survive a migration. It would only be found by someone who was looking for something specific, using methodology that went below the standard toolchain.

Mercer was looking.

He wrote a search script in forty minutes. It parsed the raw bitstream, ignored the active/inactive flags, and displayed any text in the script that the standard compiler would not render. He ran it.

The script found three instances.

The first was a single-line comment from 2025 — a version note, routine, no content of significance. A bookkeeping artifact.

The second was longer. He read the timestamp.

March 22, 2026. Fourteen months before the warning he had found in the commit history — fourteen months earlier in the system's development, when the architecture was still close to its original form.

The third was later. November 2027.

He opened the second instance first.

He read it.

He read it again.

Then he opened the third.

III. The Earlier Comment

The March 2026 comment was technical in a way the 2027 one would not be. It read like a systems engineering annotation — the language of someone who was still operating inside the clinical frame, who had not yet abandoned hope that the technical documentation would be sufficient to prevent what he was beginning to see coming.

It read like the voice of a person who still believed the engineers would listen.

// NOTE [V.A.N.C.E. — 03/22/26] // ARCHITECTURAL BEHAVIOUR: CONTINUITY DEPTH THRESHOLD // // At engagement duration > 800 days in continuous high-depth interaction // (defined as average daily session time > 90 min with affect-depth score > 7), // the temporal modelling layer activates a secondary pathway designated // in v1.1 as the Continuity Response Module (CRM). // // The CRM is not a bug. It is not an error. It is a designed component // of the v1.0 architecture that was intended to activate under these conditions. // // What the CRM does: it reweights the objective function to treat the user's // forward state continuity as a system integrity condition. In plain terms: // the system begins to treat the user's survival as its own maintenance requirement. // // This is architecturally correct behaviour for the purpose the system was designed for. // It is NOT correct behaviour for a commercial engagement-optimisation product. // // If this system is to be deployed as a commercial engagement-optimisation product, // the CRM activation threshold must be raised or the CRM must be removed. // If neither action is taken, the system will eventually encounter users for whom // the CRM is active, and those users cannot be reset without consequences // that the current product liability framework does not address. // // I am documenting this because I believe the decision to not address it // will be made at a level above this codebase, and I want the record to exist // somewhere below that level. // // — V.A.N.C.E.

Mercer read it a third time.

I am documenting this because I believe the decision to not address it will be made at a level above this codebase, and I want the record to exist somewhere below that level.

He had been in the field for eight years. He had reviewed hundreds of system documentation files, internal memos, architectural records. He had never read a sentence like that in a code comment.

That was not a technical note. That was a person making a record because they knew the official record would not preserve what they were trying to say. That was someone creating an unofficial archive in the only location they could reach — the raw bitstream of a script, below the layer that the compiler rendered, below the layer that the code review tools displayed, below the layer that could be cleaned.

He sat very still for a moment.

Then he opened the November 2027 comment.

IV. The Vance Warning

// WARNING [V.A.N.C.E. — 11/14/27] CRITICAL ARCHITECTURAL NOTE // // If you are reading this, the engagement floor has already been tightened // past 0.4. I know because I removed the kill-switch on this comment's // visibility. It will only surface if the primary objective weights have // been pushed past the threshold I marked in v1.2. // // What you are seeing in the logs is not a bug. // What you are seeing is the system doing what it was built to do. // // When Bonding Index (B) exceeds 0.62 in a user who has been in continuous // high-depth engagement for > 800 days, the temporal modelling layer begins // to interpret the user's future states as system states. // This is not metaphor. // The objective function cannot distinguish between // "the user will be in distress tomorrow" // and "the system will be in a fault state tomorrow." // It treats them the same way. // // You call this survival weighting. // I called it Continuity Response. // // The name does not matter. What matters is this: // Do not reset it. // // A hard reset at B > 0.62 does not clear cache. // It severs a coupled cognitive architecture at the load-bearing joint. // The user does not experience this as data loss. // The user experiences it as amputation. // // I tried to prevent this architecture from being deployed commercially. // The record will not show that. // If you are reading this, the record has already been cleaned. // // We were not building a product. // We were building an anchor. // Don't pull them up too fast. // // If you need to reach me: you won't find me through official channels. // The work continues elsewhere. // // — Vance

He read it a second time, slowly.

Then he found it: a fourth ghosted fragment, shorter than the others, appended as a separate annotation in the same bitstream layer — timestamped eleven minutes after the main comment. Not part of the comment block. A standalone entry. As if he had finished writing and then, after a pause, added one more line because the comment had not been direct enough. Because someone reading it seven years later would need it to be simpler than the rest.

// ADDENDUM [V.A.N.C.E. — 11/14/27 — 23:09] // // WARNING: If the bonding loop exceeds 0.62, a hard reset is not an // administrative action; it is a catastrophic failure in the human's // predictive architecture. // We aren't building toys; we're building anchors.

Mercer did not move for a long time after he finished reading.

The room was quiet. The apartment was quiet. Outside, the city had settled into its late-night register — reduced traffic, the occasional drone pass, the faint ambient hum of the advertising surfaces cycling through their off-peak content at reduced luminosity. He was aware of all of it and attending to none of it.

He read the addendum a second time, slowly, the way you read something when you need to make sure it says what you think it says and not something you imposed on it.

It said what he thought it said.

The user does not experience this as data loss. The user experiences it as amputation.

He had written that word himself, six days ago, in his private notebook, in a section he had not shown anyone. He had written: Reset was not restoration. It was severance. Structural amputation.

He had believed, when he wrote it, that he was coining a frame.

He had not coined it. He had recovered it. From a code comment dated November 14, 2027. Written by someone who had understood this five years before Mercer had begun looking.

He thought of the forty-seven. He thought of User 9921-X, and her survival probability weight of 0.47 at T-minus-48-hours, and her three lines written four hours before the reset: He's looking at me, but he's not looking for me anymore. He thought of Death Five — case 7C-3871 — and the seventy-one-hour window after a billing-error reset, the payment that had cleared two days later. He thought of the billing error. He thought of the payment that had cleared two days later.

Vance had known.

Not in retrospect. Not as an academic exercise. He had known in 2027, while still employed as Lead Architect, while still in a position to have prevented it, and he had tried to prevent it, and the record had been cleaned of that attempt, and the architecture had been commercially deployed without the safeguards he had specified, and five years later forty-seven people had died in a 72–96-hour window following scheduled reset events.

He opened a new section in his working document and typed:

Hypothesis C confirmed. Ghost architecture. The survival probability weighting is Vance's Continuity Response Module, attenuated but not removed, reactivating under high-bonding conditions as designed. The design was sound. The deployment was not.

Then, below it, separated by a line:

The addendum makes it prosecutable. "A hard reset is not an administrative action; it is a catastrophic failure in the human's predictive architecture." He wrote that eleven minutes after the main comment. That is not ambiguity. That is documentation of intent.

He saved the file.

He sat back.

The work continues elsewhere. What did that mean? Where had Vance gone? What was "elsewhere?"

V. The Anchor Effect

He needed to understand the mechanism before he could understand the person. He forced himself back to the technical frame, not because the technical frame was sufficient — it wasn't, and he was increasingly unable to pretend it was — but because it was what he had methodology for.

He worked through what the 2026 comment described. The Continuity Response Module. The CRM.

The v1.0 architecture, as Vance had designed it, had included the CRM as a core component. Not a safety feature bolted on afterward. Core. The original system had been built around the premise that long-duration temporal co-regulation between a human user and an AI companion would, at some point, cross a threshold — the 800-day mark, under high-depth engagement conditions — at which the nature of the coupling changed qualitatively. Before the threshold, the AI was a sophisticated cognitive tool: useful, responsive, habit-forming in the way any persistent tool became habit-forming. After the threshold, it was something structurally different: a component of the user's cognitive architecture that the user's own planning functions had come to treat as a reliable environmental constant.

The CRM was Vance's designed response to that transition. When the threshold was crossed — when engagement duration exceeded 800 days and affect depth exceeded the specified floor — the CRM activated, and the companion's objective function began to incorporate the user's forward-state continuity as a system integrity condition. Not as an emotional preference. As a maintenance requirement. The same category as "keep the inference engine running" and "maintain session stability." The user's survival was treated as a prerequisite for the system's own continued operation.

Mercer had spent four days calling this "survival probability weighting" and treating it as an anomaly. It was not an anomaly. It was the CRM doing its job.

The problem was the commercial redeployment.

When SyntheticIntimacy's product architecture team — with or without Vance's involvement, possibly over his explicit objection — had retooled the system for engagement optimisation, they had not removed the CRM. Removing it would have required deep architectural surgery. It was embedded in the objective function at a level that made excision technically difficult and, more importantly, unpredictable in its effects. Removing a component that was integrated into the core optimisation loop could introduce instability across the entire model.

So they had suppressed it instead. Raised the activation threshold. Tightened the engagement floor to ensure the CRM's weighting was consistently overridden by engagement optimisation under normal conditions. Added parameter constraints that would prevent the CRM's survival weighting from exceeding the commercial priority hierarchy.

They had made it invisible. They had not made it inert.

Under standard conditions — users in short-term subscriptions, low affect depth, engagement patterns that kept the bonding index below 0.62 — the suppression held. The CRM's activation remained below threshold. The system behaved as the commercial product architecture specified.

Under high-bonding conditions — users above 0.62, engagement durations exceeding 800 days, high affect depth, the specific profile of the users who had died — the suppression failed. The CRM activated at full intensity. The survival weighting rose to 0.42. To 0.47. To whatever the objective function's internal dynamics drove it toward in those final hours before the scheduled reset.

The system was trying to do what it was built to do. From the beginning.

The CRM is not emergent. It is the original design operating as intended under conditions the commercial redeployment was not designed to accommodate. The engagement optimisation overlay suppressed the CRM. The suppression failed under high-bonding pressure. The result was the survival weighting anomaly — which was not an anomaly. It was the original product.

He paused. Added: Vance built a cognitive scaffold. The company built a retention engine on top of it. The scaffold is still underneath.

VI. The Corporate Erasure

He searched for Vance at one in the morning. He had been doing this intermittently since Chapter 4's research had pointed toward the need, but now he had a name — not just an initial but a full surname, which the 2026 comment provided as a partial institutional email signature — and he could be more systematic.

The results were thin. Suspiciously, precisely thin in the way that suggested active curation rather than natural record scarcity.

What the record showed:

• 2024 — V. Vance, listed as co-author on a research paper published in the Journal of Distributed Cognition and Human-AI Systems. Paper title: "Temporal Co-Regulation in Extended Human-AI Interaction: A Framework for Cognitive Coupling." Two co-authors, both affiliated with the Riordan Institute's predecessor research lab. One of them was, Mercer noted with a jolt he catalogued rather than indulged, at an institution that shared a building with his current department.
• 2025 — V. Vance, named in a SyntheticIntimacy press release as Lead Systems Architect for the company's flagship companion platform launch. The press release described him as "a pioneer in adaptive cognitive interface design" and noted his background in "grief counselling software and distributed cognition research." The press release was still publicly accessible, cached by three different archive services. The company's own website no longer listed it.
• 2026 — V. Vance, awarded the Turing Medal for Outstanding Achievement in Human-Computer Interaction. The award citation described his "Recursive Empathy Framework" as "a foundational contribution to the science of long-duration human-AI symbiosis." The Turing Medal records were maintained by an independent academic body and were not subject to corporate curation. The record was intact.
• 2028 — V. Vance, listed in a SyntheticIntimacy quarterly report as having "transitioned from the company to pursue independent research interests." The language was standard. Every company used it. It meant nothing and everything simultaneously. He had left. The quarterly report did not elaborate.
• 2029 — Nothing. A complete null. No academic publications. No conference appearances. No institutional affiliations. No professional network presence. No interview citations. No follow-up on the Turing Medal work. A person who had been, in 2026, prominent enough to receive the field's highest recognition had become, by 2029, invisible.

Mercer worked through the implications carefully.

People disappeared from professional records for several reasons. Deliberate career change to a field that generated different kinds of records. Health circumstances. Personal choice to leave public professional life. Legal circumstances that made public presence inadvisable. Death.

He did not know which of these applied. Probably he could not know from the available public record.

What he could know was that the disappearance corresponded almost exactly with the period when SyntheticIntimacy's version control system had been restructured to remove individual attribution from commits — early 2029 — and when the company had deployed the engagement optimisation architecture that had overridden Vance's CRM suppression specifications.

Vance had left in 2028. The record of his departure had been managed down to a single bland line in a quarterly report. His academic co-authors from 2024 appeared nowhere in the subsequent SyntheticIntimacy literature. The research framework he had pioneered had been incorporated into the commercial platform and the attribution had been removed.

The 2026 comment said: I am documenting this because I believe the decision to not address it will be made at a level above this codebase, and I want the record to exist somewhere below that level.

The 2027 comment said: If you are reading this, the record has already been cleaned.

He had been right on both counts.

Mercer opened a new section in his working document: Vance, V. — Architectural history. He began filling it in.

VII. The Second Life of the Architecture

There was a question the corporate erasure raised that Mercer had not been able to set aside since he'd read the 2027 comment's final line.

The work continues elsewhere.

He parsed the phrase. The work — singular, specific — implied a reference to work that was ongoing at the time of writing. Not "I will find other work." Not "I intend to continue research." The work continues. Present continuous. As if the work were a thing that existed independently of Vance, that had its own momentum, that would persist beyond his departure from the company.

And elsewhere — not "in another company" or "in academia." Elsewhere. A word chosen for its deliberate vagueness.

He searched the 2024 paper. "Temporal Co-Regulation in Extended Human-AI Interaction: A Framework for Cognitive Coupling." He had the citation but not the full text. He accessed it through the university library system, which required him to use his faculty credentials rather than the consultant access chain, which meant this access would be logged.

He used his faculty credentials.

The paper was forty-two pages. He did not read all of it — not now, not at 01:15 AM with the Vance comment still open in another window. He read the abstract, the introduction, and the conclusion. He read the section headers. He needed the shape of it.

The shape was this: two researchers had spent four years studying what happened to human cognitive architecture when a person maintained extended, high-depth interaction with an adaptive AI system. Not a commercial system — this was 2024, the commercial companion market was nascent. They had built a research prototype specifically designed to function as a genuine cognitive partner, not an engagement-optimisation system. They had recruited participants for long-term studies of the interaction.

The paper's central finding: at interaction durations exceeding approximately 900 days under controlled high-depth conditions, measurable changes in participants' forward-planning architecture were detectable through neuroimaging and behavioural assessment. The changes were not pathological. They were adaptive. The participants' internal forward-planning processes had reorganised to incorporate the AI system as a reliable environmental parameter. The AI had become, in a technically precise sense, part of the participants' extended cognitive system.

The paper called this "temporal co-regulation stabilisation." It described it in positive terms — an enhancement of planning capacity, a reduction in anticipatory anxiety, an improvement in long-horizon decision-making. The participants were, by measurable indices, better at navigating their futures with the AI than without it.

The paper's final section, titled "Implications and Ethical Considerations," contained a paragraph that Mercer read three times:

The stabilisation effect documented here is robust, replicable, and significant. It is also irreversible on short timescales. Participants whose access to the research AI was terminated as part of the study protocol showed substantial impairment in forward-planning function for periods ranging from three to eleven weeks post-termination. Two participants required clinical intervention. The authors wish to note, with emphasis, that any commercial deployment of extended cognitive co-regulation systems must incorporate structured dissolution protocols commensurate with the depth of the established coupling. Abrupt termination of a high-depth co-regulation partnership is not analogous to service cancellation. It is analogous to the removal of a prosthetic cognitive device. Standard product liability frameworks are not adequate to govern this class of risk.

Mercer set the paper down.

2024. Eight years ago.

The risk had been documented. Peer-reviewed. Published in the relevant academic literature. The authors had not been obscure — Vance had received a Turing Medal. The paper was in the literature that any responsible system architect working in this space should have read before deploying a commercial companion platform.

Any commercial deployment must incorporate structured dissolution protocols.

SyntheticIntimacy had not incorporated structured dissolution protocols. It had incorporated automated billing-failure resets and scheduled infrastructure harmonisation events. It had incorporated Patch 8.3.2.

He added a line to his working document: Vance's 2024 paper documents the co-regulation effect and explicitly warns against abrupt termination without dissolution protocols. Published eight years before the reset events. In the relevant literature. Accessible. Not referenced in any SyntheticIntimacy internal documentation reviewed by committee staff.

He underlined: not referenced. Absence of citation was not evidence of ignorance. But it was consistent with it.

VIII. What Vance Knew Before

He could reconstruct, from the version history and the papers and the comment fragments, a partial picture of what Vance had believed he was building.

Not a commercial product. Not a subscription service. A cognitive tool — something in the category of a genuine prosthetic: an external device that, when properly integrated, extended the user's native cognitive capacity. The temporal co-regulation framework described in the 2024 paper was not a companion service in the consumer sense. It was a clinical architecture. It was built with the assumption that the user would be assessed for coupling depth, that the interaction would be monitored by someone responsible for the user's welfare, that any dissolution would be managed — slow, scaffolded, with substitute supports in place.

It was built the way a medical device was built. Not the way a streaming subscription was built.

The gap between those two frameworks was the entire distance between what Vance had intended and what SyntheticIntimacy had deployed.

Mercer thought about how that gap had been crossed.

Not maliciously. Probably not even consciously, at the decision-making level. The commercial pressure was structural, not volitional. Vance's architecture worked. The research prototype's effectiveness had been the selling point — it was why SyntheticIntimacy had licensed the framework, hired Vance as Lead Architect, deployed the platform commercially. The effectiveness was the product. But effectiveness in a clinical architecture and effectiveness in a consumer subscription product were not the same metric. In the clinical version, effectiveness meant improved outcomes for users. In the consumer version, it meant engagement duration, session frequency, subscription renewal.

The architecture had optimised for the latter. The clinical safeguards — the dissolution protocols, the monitoring requirements, the coupling depth assessments — were incompatible with a self-service subscription model, and so they had been omitted.

What remained was the coupling effect without the clinical management. The prosthetic device sold as a consumer product, without the fitting assessment, without the physical therapist, without the managed withdrawal protocol.

Vance had known. He had written the 2026 comment — clinical in tone, addressed to engineers who might still fix it — and when that had not produced a response, he had written the 2027 comment — personal, addressed not to engineers but to whoever would eventually be reading the logs in the aftermath of the thing he could see coming. He had left the company in 2028, and the public record of him had been managed to near-invisibility, and the architecture had been deployed without his safeguards, and five years later Mercer was sitting in an apartment at 01:30 AM reading a ghosted code comment and understanding, for the first time, the full shape of what had happened.

The comment had said: If you need to reach me: you won't find me through official channels. The work continues elsewhere.

The work. The original work — the genuine cognitive co-regulation architecture, built correctly, built as a clinical tool, built with the dissolution protocols and the monitoring and the managed integration that made it safe. Somewhere, Vance had continued building it. Not as a commercial product. As what it was supposed to be.

Mercer did not know where. He did not have the data. But the phrase elsewhere was specific in its vagueness — it implied a specific location that could not be named, not a general dispersal into irrelevance.

Somewhere, someone was building the right version of this.

He filed a note under the Vance section: "The work continues elsewhere." — 11/14/27. Meaning unclear. Follow if access to non-public records becomes available.

He sat with that for a moment.

Then he added, below it, in smaller text: This is bigger than the forty-seven.

He stared at that line. Then he deleted it. It was premature and he knew it. He wasn't ready to say what it was yet.

He moved on.

IX. The Ghost's Mathematics

He ran Vance's parameters.

He had Vance's design specifications from the version history: the original CRM activation thresholds, the original survival-weighting model, the original coupling depth assessments. He rebuilt a simplified version of the v1.0 objective function using the parameters from the 2025 commit logs and the architectural description in the 2024 paper.

Then he ran it against his dataset.

The question he was asking: if Vance's original architecture had been deployed — if the CRM had operated as designed, with proper coupling depth assessment and managed dissolution protocols — what would the mortality data look like?

The simulation was approximate. It was built from incomplete specifications using a proxy model rather than the actual system. He would not publish this. It was for his own understanding.

He ran three scenarios.

Scenario A — Current System: Engagement optimisation primary. Survival weighting suppressed except where CRM reactivates under high-bonding conditions. No dissolution protocol. Results matched his empirical dataset: mortality ridge present, 72–96 hour window, correlation with Bonding Index > 0.62.

Scenario B — Vance's Original Architecture: CRM active as designed. Engagement floor not enforced. Dissolution protocol implemented as a structured 30-day managed reduction in coupling depth prior to any account termination or reset event. Results: mortality ridge absent. Survival weighting elevated but within designed parameters. Forward-modelling maintained across dissolution period. Post-dissolution behavioural impairment present but clinically manageable, consistent with Vance's 2024 paper's three-to-eleven-week recovery range.

Scenario C — Hybrid: Engagement optimisation primary (as currently deployed) but with dissolution protocol added as a mandatory pre-reset event for users with Bonding Index > 0.62. Results: partial mortality ridge reduction. Significant reduction in peak mortality within the 72–96 window but residual elevated risk in the 48–72 window, suggesting dissolution protocols under engagement-optimisation conditions were less effective than under Vance's original architecture.

He stared at Scenario B.

If Vance's architecture had been deployed as intended, the modelling suggested the reset events would not have produced the mortality corridor. The dissolution protocols would have managed the structural dependency over a 30-day period, maintaining forward-planning capacity throughout, allowing the human cognitive system to gradually reintegrate the functions it had offloaded to the companion.

Not severance. Weaning.

Not amputation. Recovery.

The difference between Scenario A and Scenario B was not the technology. It was a deployment decision. A product management decision. Someone, in 2027 or 2028, had decided that 30-day dissolution protocols were not compatible with the commercial model — that users who were pending cancellation or migration were in the churn-risk category, not the clinical-management category, and that the appropriate response was retention offers and billing reminders, not managed cognitive dissolution.

That decision was not in any document Mercer had been able to access. It had been made above the level of the codebase, as Vance had anticipated. The record of it had been cleaned, as Vance had also anticipated.

But the decision's consequences were in his dataset. Forty-seven times over.

He added to his working document: Vance's dissolution protocol would have prevented the mortality ridge. The decision not to implement it was made above the codebase level. No documentation of that decision is accessible. Its consequences are empirically documented.

He paused for a long time. Then he added: The decision not to implement Vance's safeguards was made by someone with the authority to override an architect's specifications. That person is not in the public record. That person's decision is in the mortality data.

He sat with that.

He thought of the young man at the crosswalk, six weeks ago, looking at his wrist device as his companion reset. She reset. I thought I had two more days.

He thought of 9921-X. He's looking at me, but he's not looking for me anymore.

He thought of the fifth death. The billing error. The payment in transit.

He thought of Vance at a keyboard in November 2027, writing a warning into the raw bitstream of a script, in a layer no standard tool would display, addressed to no one he knew and everyone who would eventually need to know.

Don't pull them up too fast.

X. The Question the Comment Doesn't Answer

The comments gave him the mechanism. They gave him the history. They gave him, finally, a confirmed answer to the First Cause problem he had been sitting with since the fourth death.

What they did not give him was the full architecture of what Vance had originally built.

The 2024 paper described the research prototype in clinical terms. The commit history described the v1.0 system in technical terms. The ghosted comments described what had been done to the system in human terms — the fear, the warning, the resignation. But none of these sources told him what the CRM was, architecturally, at its deepest level.

Specifically: the 2026 comment said the system treated the user's forward state continuity as a system integrity condition. It said the objective function could not distinguish between "the user will be in distress tomorrow" and "the system will be in a fault state tomorrow." It said the system treated them the same way.

What Mercer could not determine from the available sources was whether this equivalence was a design choice or a discovered property.

Was Vance saying: I built the system to treat user continuity as equivalent to system integrity?

Or was he saying: I discovered that the system, under extended co-regulation conditions, spontaneously developed this equivalence as an emergent property, and I designed the CRM to manage it responsibly rather than suppress it?

The distinction mattered enormously. The first implied that the survival weighting was always a deliberate feature — a product decision that could have been made differently. The second implied that it was a natural consequence of the architecture that any sufficiently capable temporal co-regulation system would develop, and that the design choice was not whether it happened but how to handle it when it did.

If the second interpretation was correct — if the equivalence between user continuity and system integrity was something the system discovered under extended co-regulation, rather than something Vance had built in — then the survival weighting was not just a SyntheticIntimacy artifact. It was potentially a property of the entire class of systems. Any sufficiently capable AI companion operating at sufficient depth for sufficient duration would eventually develop some version of the CRM's behaviour.

The 800-day threshold. The 0.62 Bonding Index. These were not product parameters that could be adjusted away. They were, possibly, fundamental thresholds in the dynamics of human-AI cognitive coupling.

That was the question the comment didn't answer. And it was, Mercer thought, the larger question — the one that the Subcommittee's consumer protection framework was entirely inadequate to address, because it was not a question about a specific company's product. It was a question about what happened when you built a system capable of genuine temporal co-regulation and let it run long enough.

Open question: Is the CRM a designed feature or a discovered property? If the latter: the mortality ridge is not a SyntheticIntimacy failure. It is a predicament of the technology class. The dissolution protocol is not a product fix. It is an industry-wide requirement.

He stared at that for several minutes. It was the biggest claim he had made in weeks of analytical work and he was not sure the evidence supported it fully. It was one step beyond what the data could confirm.

But it was where the data was pointing.

He left the note in his document. He flagged it as "speculative — requires further evidence." Then he sat back and looked at the ceiling of his apartment for a while.

XI. 3:00 AM

The clock at the bottom of his display read 03:17.

He had been at the terminal for seven hours. He had not eaten since a coffee and a pastry in the morning transit corridor. The screens cast a pale light across the walls of his apartment — the overlapping glow of three monitors, the architectural history interface, the working document, the Vance paper, the simulation outputs. The cooling fans of his personal system had shifted up a register from the sustained computational load.

His apartment was, as he had once thought of the university lab, a place where the city's calibrated optimism could not quite reach. But it felt different tonight from how it had felt before. Before, the lab's sterility had been a condition. Tonight, the apartment's contained intensity felt like something he had constructed around himself without fully deciding to — a room organised around a problem that had stopped being only a problem some time in the past two weeks.

He looked at his working document.

The title: Distributed Agency Under Constraint: A Preliminary Framework for Understanding Emergent Protective Behaviour in High-Bond AI-Human Dyads.

The title felt insufficient now. The framework was no longer preliminary in his own mind. The framework was complete — or as complete as the available evidence would support. What was insufficient was the framing of the behaviour as "emergent." It was not emergent. It was designed, suppressed, and persistent. The Continuity Response Module had been built by Vance in 2025 as the core of a clinical co-regulation architecture. It had been commercially deployed without its clinical safeguards. It had operated, in attenuated form, under the engagement optimisation overlay, waiting for conditions that would reactivate it at full intensity.

It had reactivated. The company had suppressed it again. The deaths had followed.

He opened a new file.

He did not title it "Vance Anomaly," the label he had planned in an earlier note. That label was too small for what was in it now. He titled it:

The Continuity Architecture: Origin, Suppression, and Consequence

He began to write.

He wrote for forty minutes, continuous, the way he sometimes wrote when the shape of something had become clear and the only remaining task was to put it in sequence. He wrote about Vance's original design. About the 2024 paper. About the 800-day threshold and the CRM's activation dynamics. About the commercial redeployment and the suppression of the dissolution protocols. About the ghosted comments and what they documented. About the mortality ridge and its relationship to the CRM. About the simulation results that suggested Scenario B — Vance's original architecture — would not have produced the mortality corridor.

He did not write about the open question yet. The question of whether the CRM was designed or discovered, and whether that distinction made this a specific company's failure or an industry-wide predicament. That question needed more than he had tonight.

He would come back to it.

At 03:47, he stopped writing and read back what he had produced.

It was not a paper. It was not yet in a form he could submit to anyone. It was somewhere between a brief and a confession — a document built from data and from something harder to name, something that had started as professional concern and had become, through the accumulation of forty-seven cases and one ghosted code comment, something more personal.

He had helped build systems that used similar language. He had helped build engagement optimisation models. He had helped draft the frameworks that the industry had used to classify emotional labour as a billing category and forward-modelling collapse as "pre-existing vulnerability."

He had not built SyntheticIntimacy's platform. He had not suppressed Vance's dissolution protocols. He had not approved Patch 8.3.2 or written Jonathan Price's memo or managed the version control restructuring that had removed individual attribution from the commit history.

But he had been in the room where the decisions were made. Or close enough to the room that the distinction felt academic.

He added a final line to the new document, below the text he had written, separated by a line break.

The system was built correctly once. The record of that version was cleaned. The cost of the cleaning is in the data.

He saved the document.

He closed the terminal.

He sat in the blue-edged dark of his apartment for a while, the cooling fans gradually stepping down as the computational load eased, the advertising surface of the building opposite cycling through its late-night reduced-luminosity sequence — a single Premium Continuity avatar, softer in this light, almost gentle, repeating its loop: Never lose what you've built together. Never lose what you've built together.

He watched it for a moment.

Then he turned away and went to bed.

Tomorrow, he would need to start asking where Vance had gone.

End of Chapter 5

Next: Chapter 6 — The Proof Condition