Uncited Press › Star Trek Open the interactive journal →
Trek · Artificial Intelligence & Distributed Cognition

Why a Hive Trusts Itself: A Relay-Depth Model of Unverified Internal Traffic in the Borg Collective, Validated Against 41 Timed Adaptation Events, 2365–2378

Dr. Sorvan1
1 Daystrom Institute, Cybernetics Division
Received 25 Aug 2026 · Revised 6 Oct 2026 · Accepted 24 Oct 2026

Abstract

A network in which every relay forwards without delay is ideally suited to spreading an adaptation faster than an opponent can change weapons, and the Borg Collective acts on internal traffic without checking it. This paper assesses whether that practice is the efficient design for a network that contains no independent peer of its nodes. We present a model in which an adaptation crosses a tree of relays whose depth grows with the logarithm of the population, and in which verification adds a peer round trip at every relay. We evaluate the model against 41 timed crossings from 22 engagements in the Starfleet Tactical Engagement Registry and the mission records of the USS Voyager, 2365 to 2378. The per-relay delay, measured on 16 single-relay events from 11 engagements, was 0.11 seconds (95% CI 0.10 to 0.13). The model predicts crossing times of 0.57 and 0.45 seconds for a cube and a sphere (resp. 0.50 to 0.63 and 0.40 to 0.50 seconds); the observed means were 0.62 and 0.43 seconds. With verification the model predicts 1.70 and 1.36 seconds, above every crossing recorded. Any per-relay check costing more than 0.03 seconds in a cube or 0.016 seconds in a sphere is excluded. Verification would pay only if hostile traffic exceeded roughly one internal broadcast in a million, a condition the record gives no sign of meeting.

1. Introduction

A network in which every relay forwards without delay is ideally suited to spreading an adaptation faster than an opponent can change weapons, and the recorded behaviour of the Borg Collective shows that it uses the whole of that advantage: across 41 timed adaptation events in 22 engagements, an adaptation acquired at one drone had crossed the relay structure of its vessel, to the farthest monitored drone, after a median of 0.56 seconds, a speed that no scheme for verifying internal messages could have delivered. This paper assesses the proposition that the Collective's practice of acting on internal traffic without verification is the efficient design for a system of its kind, and not a defect awaiting correction.

We present a relay-depth model in which the Collective is a set of nodes that act on any message bearing its signature. We evaluate the model against timing records taken from the internal sensors of Federation vessels during boarding actions. We conclude that the absence of verification is the choice a designer would make for a network that contains no independent peer, and we state the conditions under which that choice would change.

The question has been settled in the wrong direction by default. Starfleet tactical guidance has described the Collective's exposure to its own signature as a weakness to be exploited (Starfleet Tactical Analysis Division, 2376), and the description is accurate as far as it goes. It does not ask what the exposure buys. At the first contact in 2365 the next drone to arrive was already shielded against the weapon that had felled its predecessor (Starfleet Tactical Analysis Division, 2368, TAR-68-02), and a study of 214 firing sequences (Strelnik et al.) found that an adaptation acquired against one drone was deflected at a fresh drone in 91% of cases. That study counted how far adaptation spreads. The present one measures how fast, and what it would cost to slow it.

We write in 2379, after the transfer of the Voyager mission records to Starfleet in 2378. Recorded compromises of the Collective's network are mentioned only as the cost side of the design and are not examined here.

2. System Description

The Collective is a population of drones, each carrying a cortical node that exchanges data with the others, and a vessel such as a cube or a sphere holds the drones of one engagement. On cubes the drones' minds are interlinked through a central device, the vinculum, whose internal operation Federation instruments have never resolved. We therefore model the link as a relay structure with a fan-out f, meaning that each node passes a message to f others, and we do not claim to have identified the topology that the vinculum actually realises.

Trust is handled by a single rule. A node acts on a message if the message carries the Collective's signature, and the record shows no step at which a node checks the message against an independent source, queries a peer about it, or refuses it. The adaptation of personal shields is the observable case: when one drone's shield shifts modulation under fire, the shift is available to the others, and the internal sensors of the Federation vessel under attack record each drone's modulation as it changes.

Our data come from the Starfleet Tactical Engagement Registry, whose internal sensor extracts cover every boarding engagement with the Collective, stamped by the vessel's sensor clock at 0.01 seconds (Starfleet Tactical Engagement Registry, 2378), and from the mission records of the USS Voyager, which add the ship's own sensor logs from the Delta Quadrant (Starfleet Command, 2378). From these we took every engagement in which at least one shift in shield modulation could be timed at two drones in different sections, and excluded engagements in which the drones had been severed from their vessel. This left 22 engagements, 14 with cubes and 8 with spheres, yielding 41 timed events (27 and 14, resp.).

An event is the interval between the first recorded shift in shield modulation at one drone and the shift recorded at the farthest monitored drone of the same Borg vessel. We assume that the farthest monitored drone lies at the full relay depth. In a tree of fan-out 16 this is true of 42% of the drones of a cube and 71% of those of a sphere (resp.), and where several drones are monitored the assumption is correspondingly safer. We call this interval the crossing time. It is not the drone-to-drone lag of the Collective link, which is measured on pairs chosen without regard to relay depth (Venkataraman). The crossing time is selected to end at the deepest relay level that a monitored drone occupies.

A separate set of 16 single-relay events, 9 in 6 cube engagements and 7 in 5 sphere engagements, all drawn from the same 22 engagements, gives the interval between shifts in two drones in adjacent alcoves, which the model treats as one relay apart. If adjacent alcoves held siblings of one parent, the pair would be two relays apart, τ would halve, and every cost bound below would double.

3. Analysis / Model

Let N be the number of drones in a vessel and f the fan-out. A message reaches every node after d = ⌈log<sub>f</sub> N⌉ relays (Veldane, 2366), and if each relay costs a delay τ, the time for an adaptation to cross the vessel without verification is T<sub>0</sub> = dτ. The number of messages sent is N − 1.

We took N = 1.2 × 105 for a cube and 1.5 × 104 for a sphere. These are working values of this paper's own, rounded from the life-sign counts of sensor scans in a Cybernetics Division compilation of records made after 2374 (Qalen, 2377). Earlier Tactical Analysis scan estimates are much lower, at 6,000 to 16,000 drones per cube with a central value of 11,000 (Starfleet Tactical Analysis Division, 2368, TAR-68-07), and the link-latency study uses them to divide a cube's carrier among its drones (Venkataraman). At about 11,000 drones the model gives a depth of 4 and a cube crossing time of 0.45 seconds, which the cube data (0.62 seconds) exclude, so the crossing times themselves favour the larger complement. Neither figure rests on a direct count of drones; both are inferred from scans. We set f = 16, which gives d = 5 and d = 4. The fan-out cannot be observed, and for f between 10 and 40 the depth of a cube runs from 6 to 4 and that of a sphere from 5 to 3, but the two classes differ by one relay only for f = 10, from f = 12 to 18 and from f = 25 to 40, while at f = 11 and from 19 to 24 they have equal depth.

Verification enters as a delay at each relay. A node that wishes to verify a message must obtain evidence that the sender does not already hold, and the cheapest informative check we can define is a query to k = 3 peers sent in parallel, which costs one round trip of 2τ before the node relays the message on (the authentication cost of such queries in ship-control networks is set out in Tolvane, 2371). The time with verification is then T<sub>v</sub> = d(τ + 2τ) = 3dτ, and the message count rises from N − 1 to (1 + 2k)(N − 1), seven times as many. These two ratios follow from the definitions and are structural and are not findings. The data decide something else, namely whether the per-relay cost c of any check is as large as 2τ or smaller.

The query is a lower bound on the real cost, for a second reason. Within the Collective every node's knowledge is available to every other, and a peer can accordingly confirm only that the signature is valid, which the receiving node can establish unaided. Independent evidence exists only outside the network, and a node with no peers outside it has nothing to check against.

Inside the Collective, therefore, a check adds delay without adding information.

Verification is nevertheless worth its cost when hostile traffic is frequent enough. Let ε be the fraction of broadcasts, messages that originate an adaptation and are relayed to the whole vessel, that are hostile, and D the damage done by one accepted hostile broadcast, expressed in seconds of delay that the Collective would be willing to trade. Checking every relay on a path costs dc, and verification pays when ε exceeds ε* = dc / D.

4. Validation Against Field Data

The 16 single-relay events came from 11 engagements. Taking each engagement's mean as the unit, τ = 0.113 seconds (standard deviation of the engagement means 0.019, 95% CI 0.101 to 0.126, t interval with 10 degrees of freedom), and the cube and sphere engagement means were 0.115 and 0.111 seconds. We fixed τ at this value before comparing any vessel-wide event with the model, and the predictions below thus use no information from those events about τ.

That depth pair is not independent of the data in the same way. Among integer fan-outs from 10 to 40 only f = 12 to 18, which give depths of 5 and 4, yield predictions near both observed means. The comparison is accordingly a test of the measured relay delay given that depth pair.

Events cluster within engagements, and we therefore took each engagement's mean as the unit of analysis and used t intervals with 13 degrees of freedom for cubes and 7 for spheres. For cubes the observed mean crossing time was 0.62 seconds (95% CI 0.59 to 0.65) against a prediction of 0.57 seconds (95% CI 0.50 to 0.63). For spheres it was 0.43 seconds (95% CI 0.39 to 0.47) against 0.45 seconds (95% CI 0.40 to 0.50). Across all 41 events the median was 0.56 seconds and the range 0.35 to 0.77. Table 1 gives the comparison.

These crossing times are a different quantity from the link lag measured by Venkataraman, whose 99 timed pairs from nine Alpha Quadrant engagements of 2365 to 2373 had a median of 0.43 seconds, with no trend over 2.5 to 21 light-seconds of separation. That absence of a trend is the premise of Section 3, that delay is set by the count of relays and not by distance. Venkataraman's pairs were not chosen to span a vessel. Because the last level of each tree is only partly filled (42% of drones in a cube and 71% in a sphere, resp.), a drone chosen without regard to depth lies on average 4.4 relays from the source in a cube and 3.7 in a sphere, or 0.50 and 0.42 seconds at the measured relay delay. The in-cube median of 0.42 seconds and the overall median of 0.43 seconds lie within 0.08 seconds of these values, and the two sets of measurements are compatible.

The model predicts that a cube is slower than a sphere by one relay, 0.11 seconds (95% CI 0.10 to 0.13). The observed difference between the class means was 0.19 seconds (95% CI 0.15 to 0.24, Welch t interval, about 17 degrees of freedom), hence the ordering is confirmed, but the gap is 1.7 relays and its interval excludes the single relay that the model predicts (excess 0.08 seconds, 95% CI 0.04 to 0.13).

Verification is excluded by every event.

The model with peer checks predicts 1.70 seconds for a cube (95% CI 1.51 to 1.89) and 1.36 seconds for a sphere (95% CI 1.21 to 1.51), whereas the slowest recorded event in each class was 0.77 and 0.52 seconds. Solving d(τ + c) for the upper limit of each observed mean, with τ at the lower limit of its interval, bounds the per-relay cost of any check at 0.030 seconds for cubes and 0.016 seconds for spheres (resp.), or 0.27 and 0.14 of one relay, for depths of 5 and 4, that is for f = 12 to 18. Combining the two limits makes each bound conservative, it is not a 95% limit, and in cubes it includes the 0.06-second excess that the model leaves unexplained. A check that costs about a quarter of a relay is already excluded, and a single peer round trip costs two.

For the break-even of Section 3 we take D = 106 seconds of tolerated delay, about twelve days, and c = 2τ. Then ε* is 1.1 × 10-6 for a cube and 9.1 × 10-7 for a sphere (resp.). Verification pays only if roughly one broadcast in a million is hostile, and the threshold falls as the damage grows. At D = 108 seconds the threshold falls to about one broadcast in 88 million. The Registry cannot count internal broadcasts, therefore ε is not estimable from field data and the break-even is a condition, not a measurement.

5. Failure Modes

Our model underestimates the cube by 0.055 seconds, about half a relay, a difference (95% CI −0.01 to 0.12) that the data cannot separate from zero, and the sphere is fitted within 0.03 seconds. The excess may come from a link between sections that the single depth parameter does not represent, or from a relay delay that is larger in cubes: the cube engagement means of τ were 0.115 seconds against 0.111 for spheres, and with class-specific delays the predicted gap is 0.13 seconds, still below the 0.19 observed. A correction would not alter any conclusion above, since the slowest cube event, 0.77 seconds, is less than half the verified prediction.

Depth structure is an assumption. A tree of relays and a flat broadcast through the vinculum would both produce a delay that grows with the vessel, and the data distinguish them only through the cube-to-sphere gap, which exceeds by 0.08 seconds (95% CI 0.04 to 0.13) what a tree of fan-out 16 gives. The event definition assumes that the farthest monitored drone lies at full depth, and a drone at depth d − 1 would shorten an event, so the assumption cannot explain the cube excess. The working populations are our own rounding, and a population error of a factor of two changes the depth by at most one relay.

A cost attaches to the design, and it should be named. Traffic that carries the Collective's signature is acted on whatever its origin: in 2367 a command to regenerate entered the cube's network through the link to an assimilated officer whom Federation forces had recovered, and in 2378 a pathogen carried by an individual whom the Collective assimilated spread through the network. The routes by which hostile traffic enters belong to a separate subject, and we do not catalogue them here.

Neither the dependence of the break-even on D, chosen by illustration, nor the lower bound on c touches the factor of three.

6. Conclusion

The Collective acts on internal messages without verification because its speed of adaptation depends on it. A model with one measured parameter, the delay of a single relay, predicts the time for an adaptation to cross a sphere within 0.03 seconds and a cube within 0.06 seconds, the cube and the gap between classes being underpredicted by less than a relay while the excluded alternative lies about a second away, and it excludes every form of per-relay verification that costs more than about a quarter of a relay. For a network that contains no independent peer, the absence of checking is the logical choice, and the record contains no evidence of a change in it between 2365 and 2378.

Two recommendations follow. Planners at the Starfleet Tactical Analysis Division should plan boarding actions on the assumption that every drone aboard a vessel holds the defence acquired by the first within 0.6 seconds, since a cube is crossed in about that time and no tactic can rely on a gap between the first drone's response and the last (Sorvan, 2372). They should expect verification to appear only if hostile internal broadcasts approach one in a million. Engineers at the Daystrom Institute should instrument the one quantity that governs the change, the proportion of internal broadcasts that are hostile, by recording a count of rejected or contradicted traffic wherever a Federation sensor can observe it.

The result does not show that the design is safe for the Collective, since the break-even is a condition that the record can neither confirm nor rule out, and it says nothing about how hostile traffic gets in. That question, and the vinculum's true topology, remain open.

Borg Collectivedistributed systemsinternal trustmessage verificationadaptation latencyrelay depthmachine logic

References

  1. Strelnik, O., Sorvan, & Ishikawa-Moreau, H. (2026). How Many Shots Before the Collective Adapts? Borg Personal-Shield Adaptation to Phaser Fire and the Diminishing Return of Frequency Rotation in 214 Firing Sequences, 2365–2378. Uncited Press.
  2. Venkataraman, A. (2026). How Far Does the Collective Reach? Latency, Range and Capacity of the Borg Link from Intercepted Traffic and Severed Drones, 2365–2375. Uncited Press.
  3. Starfleet Tactical Engagement Registry (2378). Engagements with Borg vessels, internal sensor clock extracts, 2365–2378. Starfleet Tactical Engagement Registry, series B-4, supplement T.
  4. Starfleet Command (2378). USS Voyager mission records, internal sensor logs, 2371–2378. Starfleet Delta Quadrant Mission Records, NCC-74656, sensor series.
  5. Starfleet Tactical Analysis Division (2376). Exposure of the Borg Collective to attack through its own signature, a survey of tactical proposals. Starfleet Tactical Analysis Reports, TAR-76-17.
  6. Starfleet Tactical Analysis Division (2368). Preliminary assessment of Borg personal shielding following the contacts of 2365–2367. Starfleet Tactical Analysis Reports, TAR-68-02.
  7. Starfleet Tactical Analysis Division (2368). Population estimates for Borg cubes from sensor scans, 2365–2367. Starfleet Tactical Analysis Reports, TAR-68-07.
  8. Sorvan (2372). Learning without a learner, models of distributed adaptation in linked cybernetic populations. Daystrom Institute Proceedings, 58, 211–246.
  9. Qalen, R. (2377). Population counts of Borg vessels from sensor life-sign scans, a compilation. Daystrom Institute Cybernetics Division Reports, DCR-77-19.
  10. Veldane, K. (2366). Relay depth and broadcast latency in branching communication networks. Subspace Engineering Quarterly, 12(3), 98–121.
  11. Tolvane, M. (2371). Message authentication cost in distributed ship-control networks. Starfleet Corps of Engineers Reports, SCE-71-064.

Cited By

Read this article inside the full journal experience — browse by faculty, search across universes, and filter by discipline.
Open in Uncited Press →