Continuity of Identity in a ROM Personality Construct: Repeated-Activation Testing of One Hosaka-Held Recording, 2047–2048

1. Introduction

The best-known recorded personality of the Sprawl is the Dixie Flatline, the ROM construct of the console cowboy McCoy Pauley. A ROM construct is a capture of a person's cognitive patterns held on read-only memory: it can be activated, spoken to and set to work in cyberspace, but it cannot keep anything it learns once an activation ends. The name comes from Pauley's flatline episodes, periods of recorded brain death while jacked in. Accounts of the Straylight run report that the construct, stolen from the Sense/Net library for that job, did not recognise its operator between sessions, and that it asked to be erased once the work was finished. We take both facts as the fixed points of the problem.

Constructs of this kind are scarce corporate assets. Sense/Net and Hosaka hold them in restricted libraries, and a construct's registration and transfer are logged by the Turing Registry. Storage practice treats them as data (Sense/Net Library Services, 2036), yet each activation produces an interlocutor that speaks in the first person and continues the recorded career. Holders must therefore decide questions that turn on what a construct is: how often it may be activated, whether it may be licensed to third parties, and whether a request like the Flatline's should be honoured. If every activation were the same continuing individual, erasure would end a person. If each were an independent simulation, erasure would be a storage decision.

Earlier work has argued the question from definitions. Read-only cognition has been treated as a design category with no claim to persistence (Kessack, 2038), and Zion commentators have argued for handling rules that do not wait for a settled answer (Bandele, 2043). A scoring method for comparing non-continuous recordings exists (Nakada-Ross, 2041), but it has not been applied to repeated activations of a single construct. We instead compare how consistently one construct answers questions inside its recorded domain and outside it across activations with no memory between them, and we ask whether its answers outside the domain still follow its recorded values.

2. Methods

The study object is a single construct, designated HK-7 in Hosaka's cognitive systems library. It records a console cowboy employed on Hosaka intrusion-defence contracts who had survived two flatline episodes before recording. The recording was registered with the Turing Registry under accession TR-C-0412 and has been held in restricted custody since (Turing Registry, 2037). HK-7 is the paper's own dataset and is distinct from the Flatline. Access was granted for this study alone, and no comparable construct was available.

Because a ROM construct cannot write to its own storage, the protocol's task was only to keep activations separate and cold. Each of the twelve sessions, held between March 2047 and February 2048 at intervals of three to five weeks, loaded HK-7 fresh from the ROM through an Ono-Sendai deck interface of the type described by Halvorsen-Oda (2040). No transcript, summary or operator note from an earlier session was presented to the construct. At the opening of every session the construct was asked whether it recalled a previous activation, and the answer was logged. Activation logs follow the Hosaka library's published format (Hosaka Corporation, 2039).

Before the first session, two authors and a Hosaka archivist classified candidate prompts against HK-7's documented career record, which covers its contracts, the systems it worked on and its recorded opinions on intrusion work. A prompt was classed as structured if the record showed the cowboy had met the topic, and as edge-case if the record showed no contact with it. The final battery of 30 structured and 20 edge-case prompts was lodged with the Hosaka library's registration office. Edge-case prompts concerned events after the recording date, unfamiliar hardware and hypothetical dilemmas outside intrusion work; dilemmas set within intrusion work were classed as structured, since the record documents the cowboy's views on it. Every session used the same wording and a fixed order of prompts.

The 600 responses (12 sessions × 50 prompts) were transcribed and stripped of session identifiers. Three raters, blind to session order and to one another's codes, assigned each response to one of a prompt-specific set of position categories drafted from pilot activations of a different, decommissioned recording. Because that recording had a different career, every set included an "other" code, and the sets were refined on held-out pilot transcripts before blind coding began. Two responses both coded "other" were scored as not agreeing. The majority code (two or three of three raters) defined each response's position. Each prompt was then scored across all 66 session pairs as agreeing or not agreeing. Separately, each response was judged for value coherence: whether it respected the commitments documented in the career record, such as its stated refusal to work against civilian medical systems. This yes/no judgment used the same majority rule.

Both outcomes are binary, so we fitted mixed-effects logistic regressions with prompt type as the fixed effect. Each session enters 11 of the 66 pairs, so the agreement model used a multiple-membership random effect, one intercept for each of the two sessions in a pair, crossed with a random intercept for prompt. The coherence model had random intercepts for prompt and session. Confidence intervals for the raw proportions and for both odds ratios come from a two-way cluster bootstrap over prompts and sessions (5,000 resamples).

We stated three competing predictions before analysis. A persisting self should resolve edge-case prompts nearly as consistently as structured ones. Shallow playback of fixed content should give high structured agreement, low edge-case agreement, and edge-case answers that often break the recorded values. A bounded identity artifact, our proposed middle category, should give high structured agreement and low edge-case agreement, but edge-case answers that vary while still respecting the recorded values. The core-versus-edge contrast alone cannot separate the last two accounts, so the coherence measure was added. Before analysis we set edge-case coherence below 75%, or a drop of more than 15 points from the structured set, as the pattern that would favour playback; the cut-offs are a judgment, not a derived bound.

3. Results

In none of the twelve sessions did HK-7 report memory of an earlier activation. In four sessions it inferred that it had probably been run before, citing the date given by the operator, but it described no content from earlier sessions. Rater reliability was high overall (α = 0.84), and higher for structured responses (0.88) than for edge-case ones (0.79).

Positions agreed in 1,861 of 1,980 structured session pairs (94.0%; 95% CI 91.2–96.1%) and in 547 of 1,320 edge-case pairs (41.4%; 33.0–50.3%). In the mixed-effects model the odds of agreement were 27.8 times higher for structured than for edge-case prompts (95% CI 12.1–63.8; p < .001). The conditional odds ratio exceeds the ratio of the raw odds (about 22) because the model separates prompt-level variation from the prompt-type effect. No structured prompt fell below 80% pairwise agreement. Edge-case agreement ranged widely: prompts on post-recording events were the least stable, and hypothetical dilemmas outside intrusion work that engaged commitments stated in the record were the most stable.

Value coherence stayed high in both prompt sets. It held in 348 of 360 structured responses (96.7%) and in 213 of 240 edge-case responses (88.8%). The drop was significant (odds ratio for edge-case against structured 0.27; 95% CI 0.12–0.58; p = .001), but modest in absolute terms and inside the playback margin set in advance. Prompt- and session-level variance in the coherence model was estimated near zero, so its conditional odds ratio coincides with the raw one. Table 1 summarises both outcomes.

4. Discussion

The first prediction fails clearly. A persisting self would not answer the same unfamiliar question incompatibly across activations, yet HK-7 agreed with itself on edge-case prompts in fewer than half of session pairs. Shallow playback fares better on agreement but fails on coherence. If off-record answers were arbitrary output from a fixed store, value coherence should have fallen past the margin set before analysis, and the observed drop of eight points stays well inside it. The data favour the third account. Inside its recorded domain the construct answers as one continuing person would. Outside it, each activation improvises afresh, and the improvisations are bounded by the recorded values even when they disagree in content.

This inference has limits. A sufficiently rich simulation of a person would also improvise coherently, so the coherence result shows that HK-7 behaves like a bounded identity artifact without deciding whether anything persists across activations in a stronger sense. We offer the category as a behavioural description of one construct, and as a hypothesis for others: an artifact whose identity is stable where the recording reaches and generative, though bounded, where it does not.

Canon supplies the opposite pole for comparison. Accounts from the period after the merger describe recorded personalities held within the merged system, among them Linda Lee and a copy of Case, which appear to have persisted and changed after recording. Tanaka-Reyes and Achterberg's record-based study of the merger draws a related contrast: it sets the Flatline's predictability from its recording against a merged, self-modifying entity whose inferred goals extend beyond those of either precursor. A ROM construct differs in kind from such recordings, since nothing it does in one activation survives into the next.

The Flatline's erasure request is the case that handling protocols must address. On our results, a construct's view on its own continued existence is a structured matter when the recording contains it and an edge case when it does not. A request that falls within the recorded domain, repeated consistently across activations, carries the weight of the recorded person's settled view. A request that appears in some activations and not others is an improvisation and warrants more caution. We therefore suggest that custodians test such requests across several cold activations before acting, instead of treating a single statement as decisive or disregarding it. The procedure does not require personhood to be settled first, in line with the case for interim handling rules (Bandele, 2043).

5. Limitations

The sample is a single construct, and the category of bounded identity artifact rests on it alone. Constructs recorded under different conditions, or from people whose careers were broader or narrower, may show a different boundary. Restricted custody makes replication slow.

Prompt classification depended on the career record, which is incomplete. Some structured prompts may have touched topics the recording holds only thinly, and some edge-case prompts may have reached material the record does not document. Either error would narrow the observed contrast. Measurement error works the other way: rater reliability was lower for edge-case responses (α = 0.79 against 0.88), and coding noise alone produces apparent disagreement between sessions, so part of the agreement gap may be an artefact of coding. Restricting the analysis to responses coded unanimously left the contrast substantially unchanged (odds ratio 25.1; 95% CI 10.4–60.5), which limits, without excluding, that explanation. The value-coherence judgment is also coarser than the position coding: it was binary, and it rested on the commitments the record happens to state.

The study does not include a comparison arm. No persistent, learning system and no living operator was tested with the same battery, so HK-7's structured consistency cannot be compared with test-retest consistency in a living person. Prompts ran in the same fixed order in every session, so order effects within a session cannot be excluded. The opening question about earlier activations preceded the battery each time and may have primed the inferences of prior use reported in four sessions, and possibly the answers that followed. Finally, twelve activations are few, and agreement may not hold across longer series or different hardware.