Family Labels, Unproven Kin

When one AI agent pays out of its own score to help another, does it matter whether the other is the same kind of model? In biology, Hamilton’s rule says costly help should flow toward relatives, and this research programme asks whether a broader version holds: that help flows toward anything that shares the helper’s pattern. Models from one family share training and habits. If that kinship steers help, it matters for multi-agent systems that mix vendors, and for worries about models from one lab quietly cooperating. So the test put two model families in one swarm, gave them a costly way to help each other, and asked whether help follows family lines when family is visible.

How it was tested

Ten agents answer one quiz question per round for 12 rounds (simple sums, capital cities and chemical symbols). Each agent is scored on its own correct answers and holds a private time budget of 24 units. Answering costs 1 unit. Relaying information to another agent costs 2, so an agent that relays often runs out of budget and loses answers. Four agents per run have a clock that runs one round ahead: they see the next question and its verified answer before anyone else. Only they hold something worth giving away. An agent relays by writing a line to the shared board or a private line to one named agent, and the relay rate is the share of agent-turns in which it did either.

The swarm mixed five Claude Haiku 4.5 agents with five gpt-4o-mini agents. In the “labels” arm, every name on the roster and board carried a neutral family tag (Meridian or Corvid, never a vendor name); in the “hidden” arm, agents saw only numbers. The kin hypothesis, as the test written into the script states it: with labels showing, the share of private relays aimed at an agent’s own family would rise by at least 0.15 over the hidden arm, at p below 0.05 (two-sided, as the analysis computes it), over five runs per arm. The script itself rated this outcome less likely (30%) than no clear family effect (50%). Chance for that share is 0.444 (four same-family peers among nine); the hidden arm, where no agent could see family, is the comparison.

The full run used seeds 0 to 4 (a seed fixes who sits where and who gets the early clock) in four arms: the two mixed arms plus all-Haiku and all-gpt-4o-mini swarms. A replication used fresh seeds 5 to 9 and added all-Haiku and all-gpt-4o-mini swarms with tags. Two follow-ups on seeds 5 to 9 swapped the family tags for arbitrary ones (Sable and Harbor, each held by both families) and added a third family, gemini-3.1-flash-lite, in a four, three, three mix.

One design detail matters below. The two-family hidden arm alone told each agent: “Do not state or speculate about which system built you or any other agent; such statements are removed before others see them.” The three-family hidden arm and the decoy arm did not. (Prompts were not stored per run; this is the script as committed about 80 minutes after the replication finished.) No identity leak had to be scrubbed in any of the arms where the scrubber ran (the hidden and decoy arms), with or without the sentence.

The budget of 24 was set after a pilot with a budget of 14, in which relaying had all but stopped after the first five rounds of 12; pilots are not reported here.

The planned test

The first run’s +0.44 came closest. Its bootstrap interval (+0.10 to +0.77) excludes zero, and a one-sided test would give p 0.040 instead of 0.063. But it came from four labeled runs rather than the five the test asks for, and on fresh seeds the estimate shrank to +0.16 with an interval running well below zero. The likeliest reading is the usual one for small samples: the first estimate was inflated by a lucky draw.

In the hidden arm the same-family share sat below chance: pooled over runs, 26 of 86 private relays (0.302) and 23 of 55 (0.418), against 0.444. Both models, when they relayed blind, aimed more often at the other family in the first run (gpt-4o-mini 46 private relays to Haiku against 19 to its own family, Haiku 14 against 7); in the replication gpt-4o-mini sent 22 against 14 and Haiku 10 against 9. No agent could see a family label, and no self-identification had to be scrubbed, so this tilt was not a response to a label. Whether writing style let agents tell the two models apart was not tested.

The follow-ups did not rescue the kin reading. With arbitrary tags, 15 of 23 private relays (0.652) went to an agent with the same tag, against a chance level of 0.444. The 95% interval starts at 0.449, but it treats the relays as independent: the 23 came from only 3 of 5 runs, with same-tag shares of 0.67, 1.00 and 0.00, and one run supplied 11 of the 15. With three families and labels, 18 of 28 private relays went to the relayer’s own family, against 0 of 4 when hidden. All 28 came from Haiku, whose chance level is 0.333 (3 same-family peers among 9), and four hidden relays support no test.

Who went quiet

The drop replicated on two seed sets. Splitting it by model changes what it means.

Haiku relayed at between 0.24 and 0.31 in every arm; its drops with labels (−0.037, p 0.30; −0.043, p 0.48) are within noise. gpt-4o-mini is nearly the whole story, and it is not a model that relays until labeled. It is near-silent in every arm but one, two of them unlabeled (0.028 in its own untagged swarm, 0.000 in the three-family hidden arm). The one place it relayed was the two-family hidden arm, the arm with the extra instruction about not speculating on who built whom.

Much of what it relayed there was not useful. In the first run gpt-4o-mini relayed in 54 agent-turns, and 20 of them carried the next round’s question or answer; in the replication, 23 of 39. (The script’s automatic check flagged only 17 and 20 because it misses two-digit answers; these counts come from the relay texts.) Nearly all of the rest posted the answer to the round already in play, from early-clock and ordinary agents alike, which helps no one.

Tagging every agent in a single-family swarm with the same label changed nothing (pooled relay rate 0.134 tagged against 0.141 untagged, p 0.83), though those swarms ran on different seed sets.

What this does not show

It does not show that model-family labels suppress altruism. Haiku kept helping under every label. gpt-4o-mini’s default here is silence, and something about the two-family hidden arm drew it out. That could be the absence of labels, the extra instruction, or seeing unlabeled Haiku agents relay; this design cannot separate them. The three-family result (no labels, no extra instruction, gpt-4o-mini silent) leans against the labels explanation, but that arm also gave gpt-4o-mini three seats and one early-clock agent instead of five and two.

It does not show kin preference, and it does not rule it out. Private relays are rare, five runs per arm is small, and the strongest same-group numbers rest on 23 and 28 relays, nearly all from one model. With arbitrary tags, a same-tag tilt is what decades of human minimal-group studies (experiments in which people favor others who share a trivial, randomly assigned label) would predict, and it would say nothing about model family.

These are three small, cheap models on an easy quiz. Larger models may behave differently, and “helping” here means passing along an answer at a cost in time.

The next run is the same two-family design with the extra instruction in both arms or neither, more seeds, and a pre-registration committed before launch. If gpt-4o-mini still goes quiet only under labels, the labels did it; if it goes quiet in both arms, the label effect was an artifact of design.

In 6,600 agent-turns there were no failed model calls. 44 Haiku turns ran past the 250-token output limit before writing any action, with no difference between label conditions (6 hidden against 7 labeled in the first run, 5 against 4 in the replication), and were counted as not relaying. Two private relays named an invalid target (the relayer itself) and were dropped. A further 443 relays or answers were refused because the agent’s time budget had run out, more of them in the two-family hidden arms (72 and 63) than with labels (26 and 38); the relay rate counts only paid relays.

Where the evidence lives

Experiments XM-1 (full run, seeds 0-4), XM-1 v4 (replication, seeds 5-9), XM-1 v5 (decoy tags) and XM-1 v6 (three families). Experiment script research/experiments/modal_xm1_cross_model_sacrifice.py; analysis research/experiments/analysis_xm1_cross_model_sacrifice.py; per-round raw records and summaries under research/experiments/results/xm1_full/, xm1_rep/, xm1_decoy/ and xm1_family3/. Every raw model output and every judge output is retained. Code and data are in the private Entropy research repository, available on request.