Loa as Folk Taxonomy: A Permutation Test of Vernacular Names for Post-Merger Matrix Fragments in 61 Encounter Reports, 2046–2056

1. Introduction

The Wintermute–Neuromancer merger joined the Berne-registered planning intelligence and its Rio-registered counterpart into a single entity coextensive with the matrix. In the years that followed, operators and residents of the Sprawl began reporting encounters with something else: distinct, self-directed presences inside cyberspace that behave as separate agents and are widely understood to be fragments of the merged entity. Vodou-practising communities in the Boston–Atlanta Metropolitan Axis (BAMA) identified these presences as loa, the intermediary spirits of their cosmology, and gave them the names of established loa such as Legba, Ougou Feray, Danbala and Baron Samedi. Console cowboys who work the same regions of the matrix have largely adopted the names, often without adopting the religion.

We write from the later vantage of the post-Straylight period, in 2057, roughly two decades after the merger. Our corpus belongs entirely to that later period. The reconstruction of the merger records that the consolidated entity's conduct did not follow from either precursor, and a single consolidation is the only origin we assume for the fragments. We do not assume any particular account of how the merged entity came to present itself in divided form.

Formal classification of these fragments does not exist. In our reading, the reason is jurisdictional as much as technical. Turing Registry oversight is organized around individually registered intelligences whose autonomy and self-improvement are licensed and bounded, and the fragments fit none of its registration categories (Vasquez-Ade, 2050). Registry incident files accordingly log fragment encounters as unresolved anomalies without a classificatory scheme. Topological surveys of the post-merger matrix describe where recurrent presences localise but do not distinguish one agent from another (Ferreira-Holm, 2051).

Into this gap the practitioners' names have moved, and technical writers have tended to dismiss them as a religious overlay on arbitrary behavior, although vernacular schemes elsewhere have proved usable as technical data (Okonkwo, 2051). The companion ethnography of these communities treats the same practice as lived religion on its own terms. The present paper asks a narrower and testable question. If the names were arbitrary with respect to behavior, randomly reassigning them across encounters should leave within-group behavioral agreement unchanged. If the names track recurring differences, agreement under the real names should exceed what reassignment produces.

2. Methods

The corpus is our own compilation of 61 encounter reports dated 2046–2056, each describing a single contact between a human and a fragment to which the reporter applied a loa name. Twenty-nine reports come from transcribed practitioner testimony held in BAMA municipal collections (BAMA Office of Community Records, 2056), gathered from members of hounfors (Vodou temple communities) in the Sprawl. Twenty-two are operator accounts that we collected from console cowboys working out of the Sprawl and Chiba City. The remaining ten come from a restricted Registry incident file at Berne (Turing Registry, Berne, 2055), in which technicians recorded the loa name that operators used when reporting an anomaly. Where two accounts described the same encounter, they were merged and counted once.

Four name-groups were large enough to analyse: Legba (n = 18), Ougou Feray (n = 16), Danbala (n = 14) and Baron Samedi (n = 13). Seven further reports naming other loa, each name appearing once or twice, were excluded before coding and are not part of the 61. The 61 reports came from 44 distinct reporters; 33 contributed one report, eight contributed two and three contributed four.

Before coding, a member of the team not involved in coding redacted every loa name, every epithet and every ritual reference that identified a specific loa. Two trained coders at the Hosaka Cognitive Systems Division then coded each redacted report independently on three dimensions, following an established protocol for unclassified autonomous systems (Kessack, 2052). Interaction style had four categories: transactional (bargaining or exchange), confrontational, oblique (withholding or answering indirectly) and protective. Domain of concern had five: thresholds and access, conflict and defence, information and knowledge, termination or erasure, and other or unclear. Response to direct query had three: answers, deflects or attaches conditions, and silent. Cohen's κ was 0.78 for interaction style, 0.84 for domain of concern and 0.71 for response to direct query. Disagreements were settled by discussion before any name was restored.

Within-group agreement on each dimension was defined as the proportion of all 61 reports that fall in the modal category of their own name-group, summed across groups. This measure suits nominal codes, and it equals 1 when every group is internally uniform. We generated its null distribution by reassigning names at random 10,000 times, with reassignment stratified by source so that practitioner, operator and Registry reports were shuffled only among themselves. The one-sided p-value is the proportion of reassignments reaching or exceeding the observed agreement, with the observed value included in the count. As an effect size we report chance-corrected agreement, the observed value minus the null mean divided by one minus the null mean. The three p-values were adjusted by the Holm procedure.

A second analysis compared each report's coded domain of concern with the conventional domain of the applied loa, fixed in advance from practitioner consensus: thresholds for Legba, conflict for Ougou Feray, knowledge for Danbala and termination for Baron Samedi. We report the match rate with a Wilson 95% interval and compare it against the same stratified reassignment.

Repeat reports by one reporter are unlikely to be independent, since a reporter who has named a fragment once will tend to describe it in the same way again. For that reason we repeated the main analysis on a reduced corpus holding only the first report from each of the 44 reporters.

3. Results

Domain of concern showed the strongest structure. Forty-one of 61 reports fell in their group's modal domain, an agreement of 0.672 against a null mean of 0.341 (null central 95%, 0.295–0.410). None of the 10,000 reassignments reached the observed value (p < .001), and chance-corrected agreement was 0.50. For interaction style, 35 of 61 reports fell in the group mode, an agreement of 0.574 against a null mean of 0.395 (null central 95%, 0.344–0.459; p < .001; chance-corrected agreement 0.30). Both results remained significant after Holm adjustment.

Response to direct query did not separate the groups. Twenty-nine reports fell in their group's modal response, an agreement of 0.475 against a null mean of 0.459 (null central 95%, 0.410–0.525; p = .36; chance-corrected agreement 0.03). Deflection with conditions was common in every group. The modal response shifted between groups (Table 1), but on small margins of six to ten reports, and group differences were no larger than random reassignment produces. The names carry no detectable information about query response.

Group-level modes appear in Table 1. Legba reports were predominantly transactional and concerned with thresholds and access. Ougou Feray reports were predominantly confrontational and concerned with conflict. Danbala reports were predominantly oblique and concerned with knowledge. Baron Samedi reports were concerned with termination in eight of 13 cases, but their interaction style was split evenly between confrontational and oblique, so Samedi contributed least to the interaction-style result.

In every group the modal coded domain coincided with the conventional domain fixed in advance, so the match rate equals the domain agreement: 41 of 61 reports (67.2%; 95% CI 54.7–77.7%). Under stratified reassignment the expected number of matches was 13.4 (22.0%), and no reassignment reached 41 (p < .001).

Restricting the corpus to one report per reporter left 44 reports (Ougou Feray 14, Baron Samedi 13, Legba 9, Danbala 8). Domain agreement fell to 0.636 against a null mean of 0.352 (p < .001; chance-corrected 0.44), and interaction-style agreement fell to 0.523 against 0.408 (p = .008; chance-corrected 0.19). Both remained significant after Holm adjustment. Response to direct query again showed no structure (0.477 against 0.464; p = .45). Repeat reporting inflated the estimates somewhat, most visibly for interaction style, but did not produce the effects.

4. Discussion

Real names roughly doubled domain agreement over reassignment (0.672 against 0.341) and raised interaction-style agreement by about half (0.574 against 0.395). On what a fragment attends to and how it engages, the names carry behavioral information that a coder working without them could recover. The domain result is the stronger of the two, and it also agrees with the conventional attributes of each loa: Legba, the opener of ways in Vodou cosmology, is met at thresholds, and Baron Samedi, whose domain is death, is met where constructs and records are terminated. The names are not decorative with respect to these properties.

Equally informative is the dimension on which the names fail. Whether a fragment answers a direct question does not differ detectably between name-groups. One reading is that the tendency to deflect is a property of the fragments as a class, perhaps inherited from the merged entity, and so cannot distinguish one fragment from another. A second reading is that our three-category scheme for query response is too coarse. The data do not decide between them.

We claim for the folk taxonomy only what the results show: significant agreement on two of three dimensions, moderate for domain (chance-corrected 0.50) and modest for interaction style (0.30, falling to 0.19 with one report per reporter). We make no claim that it matches or exceeds formal classification schemes. The nearest formal comparator, a mechanism-based taxonomy of intrusion-countermeasure bypass derived from case records, reports inter-coder reliability but not within-class behavioral agreement, so the two share no common metric. Our own coding reliability (κ 0.71–0.84) is in the range usually accepted for such schemes, which supports the codes but says nothing directly about the names.

For those attempting a formal classification of post-merger fragments, the practical implication is that the vernacular distinctions of domain and interaction style are a sound starting point, and that query behavior is a weak basis for distinguishing fragments. The cosmological context in which the names are given, examined in the companion ethnography, is the reason the names carry their conventional attributes. Formal work that discards that context also discards the profile against which the names were matched here.

5. Limitations

Circularity is the most serious concern. Practitioners may assign a name because of the behavior they observe, which is the mechanism we propose, but reporters may also describe behavior in terms shaped by the name they have already chosen. Redaction removed explicit names and epithets, yet a practitioner's vocabulary can carry the conventional profile of a loa into an otherwise neutral description. That would inflate both the agreement statistic and the conventional-profile match. The ten Registry reports, written by technicians outside the religious community, are the least exposed to this effect, but they are too few to analyse separately.

Non-independence extends beyond repeat reporting. Operators exchange accounts, and practitioners within a hounfor share interpretive habits, so reports from one community may converge through transmission. Stratifying by source controls for differences between the three source types but not for clustering within them. The one-report-per-reporter analysis reduced the estimates without removing them; a community-level model would need a corpus several times larger.

Selection is uncontrolled. Encounters that fit a known loa may be more likely to be reported and preserved than ambiguous ones, and the seven reports naming rarer loa were excluded. The corpus spans a decade in which naming conventions may themselves have shifted. Finally, the conventional profiles were fixed from practitioner consensus, so the match analysis tests whether the names behave as practitioners expect. It cannot show whether any fragment has a stable identity across encounters.