What Returns: A Continuity Audit of 286 Restorations from Stored Mind-States by Storage Interval and Lace Generation, GSV Quiet Enthusiasm, 2312–2379 CE

1. Introduction

Citizens of the Culture choose when to die, and many arrange beforehand for a state of their mind to be kept. In the account we adopt here, which is our own, a neural lace captures the state and a Mind holds it. If the person later dies, or is lost in a way that leaves no body to revive, the state may be restored into a body prepared for them (Wennowick, 2318 CE). In the register's practice, storage and restoration follow the person's own standing instruction, and no entry records a restoration carried out against it. Those who knew a restored person generally receive them as the same person they knew (Sorrinde, 2372 CE).

What is restored is a separate question from whether it is welcome. A stored state is a record of a person at one moment, made by hardware of a particular design and held for some interval before use. Engineering accounts have validated captures against test scenes and instrument checks, and have shown how the sources of error differ between designs (Prasquelle, 2347 CE; Ellensvaard, 2361 CE). The reasons people choose storage have been surveyed (Eskarrow, 2335 CE). To our knowledge, nobody has compared the restored person's memory and manner with the same person's own record.

This paper does that for one Faculty's register. We ask how a measure of continuity after restoration varies with the storage interval and with the generation of lace that made the capture, and how often restored people say that they are the person who was stored. The register opens in 2312 CE, when laces in use were still uncommon, and closed at the end of 2379 CE. We write in 2380 CE and use no record from after the closure. The measure, the register series and every figure below are our own constructions and are not standard instruments of the Culture.

2. Methods

The Neural Interface Faculty of the GSV Quiet Enthusiasm keeps a register of the restorations it has performed or supervised. For each it holds the capture log of the stored state, the restoration record and the score sheets of the assessments that follow (Neural Interface Faculty, 2379 CE). By the end of 2379 CE the register held 312 restorations. We excluded 17 for which fewer than 60 of the 120 probe items described below could be checked against a surviving record made by the person before capture, and 9 whose restored person declined assessment. That left 286 restorations (91.7% of 312) of 251 people. Of these, 216 people were restored once and 35 twice, the second restoration using a state captured after the first. Restorations in the register's last month were still awaiting assessment at closure and are not among the 312.

Storage interval is the time from capture to restoration, taken from the capture log and the restoration record and expressed in years. We analysed it as a base-2 logarithm, so that a coefficient describes the change for each doubling of the interval, and we set the two intervals shorter than three months to three months. The generation of lace is recorded in the capture log. In this register first-generation captures run from 2312 to 2339 CE, second-generation from 2340 to 2362 CE and third-generation from 2363 CE onward (Prasquelle, 2347 CE; Ellensvaard, 2361 CE). The third-generation validation series of 2361 CE preceded the register's first third-generation captures in 2363 CE, as pre-deployment validation.

We scored continuity with an index from 0 to 100, given at 28 days after restoration. Each of two assessors worked independently. Each put a probe of 120 items about the person's life before capture, every item checked against the person's own record, and held a structured interview on habits, preferences and manner, following the probe design of Tolluwen (2366 CE). Events after capture were never scored, because a person who lived on after the state was taken cannot be expected to remember them. The index is the mean of the two assessors' scores and is capped at 100. Assessors did not see the capture log, though they could not be blind to what the person told them. Scoring and agreement statistics follow Hanquelle (2349 CE). The assessors agreed closely: the intraclass correlation (two-way random effects, absolute agreement, single assessor) was 0.92 over the 286 pairs, and their mean scores differed by 0.33 points.

At the same sitting each person was asked whether they regarded themselves as the person who was stored, and answered yes or no. We call this endorsement. It is the person's own judgement and was not coded by us. Two readers separately coded each capture log for whether it recorded a fault during capture. They agreed on 281 of 286 logs (κ = 0.88), settled the five disagreements by discussion, and flagged 21 logs.

We modelled the index by linear regression on the base-2 logarithm of the interval and on generation, with first generation as the reference, and endorsement by logistic regression on the same predictors. People restored twice contribute correlated observations, so standard errors are clustered by person across the 251 clusters (Quinnevar, 2351 CE). Intervals are Wald intervals and p values are two-sided. We chose the logarithmic form after seeing that band means fell quickly at first and then more slowly, so the p values for interval are conditional on that choice. We refitted the index model without the 21 flagged logs. Two models answer the two questions, and we made no further correction for multiple comparisons. Figures are rounded to two decimals, and derived quantities were calculated from unrounded values.

3. Results

The 286 restorations had a median storage interval of 3.21 years (interquartile range 1.57–7.62, range 0.09–63.22 before the floor of three months described in Methods). By generation, 71 were first-generation captures, 118 second-generation and 97 third-generation. By interval, 43 were shorter than a year, 186 ran from one year to under ten, 48 from ten to under thirty and 9 were thirty years or longer. Intervals were shorter for newer generations, as the register's dates require. Seven of the nine states stored for thirty years or longer were first-generation, two were second-generation and none was third-generation.

Across all restorations the Continuity Index averaged 86.53 (SD 7.84, range 61.15–100.00); six restorations reached the cap of 100. It fell as the interval lengthened (Table 1). The mean was 91.09 for intervals under a year, 86.92 for one to under ten years, 81.99 for ten to under thirty and 80.89 for the nine states of thirty years or more. Newer generations scored higher (Table 2): 82.79 for first-generation captures, 86.55 for second and 89.25 for third. Capture logs recorded a fault in 21 restorations (7.3%), more often in older generations, and these restorations averaged 79.34 against 87.10 for the other 265.

In the clustered regression (Table 3; R² = 0.22), each doubling of the interval lowered the index by 1.66 points (95% CI 1.17 to 2.14 points lower, p < .001). Between an interval of one year and one of ten years, the model predicts a difference of 5.50 points. With interval held constant, second-generation captures scored 3.30 points above first-generation captures (95% CI 1.19–5.42, p = .002) and third-generation captures scored 5.48 points above them (95% CI 3.33–7.63, p < .001). Without the 21 flagged logs (265 restorations, 236 people), the estimates were similar: 1.64 points lower per doubling (95% CI 1.16–2.11), 3.35 points higher for second generation (1.14–5.56) and 4.98 points higher for third generation (2.72–7.25).

Endorsement was common. Of the 286 restored people, 207 (72.4%; 95% CI 67.2–77.6%, ignoring clustering) said they regarded themselves as the person who was stored. Endorsement was less likely with longer intervals: the odds ratio was 0.84 per doubling (95% CI 0.72–0.99, p = .03). Generation showed no clear association with endorsement, with odds ratios of 1.29 for second generation (95% CI 0.67–2.47, p = .45) and 1.24 for third (95% CI 0.63–2.47, p = .53), each against first. The nine states stored for thirty years or more included eight endorsements, which is too few to read against the other bands. People who endorsed had a higher mean index than those who did not (87.72, SD 7.80, for 207 people against 83.43, SD 7.11, for 79), and the two groups overlapped widely. Among the 26 restorations scoring below 75, 17 (65.4%) were endorsed, against 190 of 260 (73.1%) scoring 75 or higher.

4. Discussion

Restored people recovered most of what their own records can check. The typical index was 86.53 of 100, and no restoration scored below 61.15. Longer storage went with lower scores, by about 1.66 points for each doubling of the interval, and the gradient survived the removal of restorations whose capture logs recorded a fault. It is an average gradient in this register. We did not test whether it differs by generation, and the band of thirty years or more holds only nine states, so the audit says little about very long storage.

The register cannot tell where the loss arises. It could accrue during storage, or reflect the state of the person's own record, or follow from differences between people who wait long and people who do not. We have no data on the reasons for a long interval, which are the person's own and which the Faculty does not require them to give.

Later generations scored higher, by 3.30 and 5.48 points, and flagged capture faults less often, in 12.7%, 6.8% and 4.1% of logs from first to third generation. This matches the engineering accounts of improving capture (Prasquelle, 2347 CE; Ellensvaard, 2361 CE). It does not show that hardware alone explains it. Generation in this register coincides with the era of capture, and with it changes in assessment practice, in the Faculty's handling of stored states and in who chose storage. The audit is silent on which of these matter.

Endorsement tells a different story from the index. Most people regarded themselves as the person stored, and endorsement fell with interval, though the band of thirty years or more (8 of 9) does not follow the trend. People with lower indices often endorsed all the same. The Faculty sets no threshold below which a restoration counts as having failed, and we propose none. A score below 100 describes how much of a record was matched, and does not measure how far someone is the person they were. Those who knew a restored person receive them as the same person by regard and by choice (Sorrinde, 2372 CE). For someone deciding whether to be restored soon or to wait, the figures show an average loss from waiting that is small against the spread between people (SD 7.84). What to make of that is for the person to decide, as the reasons surveyed by Eskarrow (2335 CE) already show.

5. Limitations

The register covers one Faculty aboard one General Systems Vehicle. People who were restored under its care, or referred to it, may differ from those restored elsewhere, and practice at other Orbitals and vehicles may differ. The register also holds only restorations that took place. It says nothing about states that were stored and never used.

Our Continuity Index is a construction of this audit and was scored at 28 days, when a restored person is still settling. Scores at later dates may differ. The index covers only the period before capture, is capped at 100, and rests on records the person made themselves, which are uneven in how much they cover. Assessors saw no capture log but could not be blind to everything the person said. Their close agreement shows that they scored alike, and does not show that the index measures what matters most about continuity.

This study is observational. Generation is confounded with capture era, and interval is partly confounded with generation, since older generations had more time in which to produce long intervals. We could not adjust for age at capture, the cause of loss or the body into which the person was restored, none of which the register holds in a form we could use. We chose the logarithmic form of interval after inspecting the data, so the p values for it are conditional on that choice.

The last month of the register is excluded by design, since restorations then awaiting assessment cannot be scored. Exclusions removed 26 of 312 restorations (8.3%). The 17 with too few checkable items are likely to be people who kept fewer records, and the 9 who declined assessment may have scored differently from those who took part. Standard errors for the index and endorsement models allow for the 35 people restored twice, but the confidence interval for the overall endorsement rate does not. The endorsement figure for the thirty-year band rests on nine people.