Can Agents Adapt Their Institutions And Sacrifice Themselves During Collapse? Executable Self-Governance in a GovSim Commons

Gregory B. Rehm

cs.MASep 18, 2026

Can Agents Adapt Their Institutions And Sacrifice Themselves During Collapse? Executable Self-Governance in a GovSim Commons

Authors: Gregory B. Rehm

Organizations: Meta

Abstract

LLM agents are beginning to write the rules they live under: policies, contracts and code that other agents review and approve. What happens when those rules must decide who is sacrificed from a group? We built GovSim-SelfGovern, a commons in which agents legislate in executable Python, test each law in a sandbox, vote on it, and live with the result. When the fishery is abundant, self-rule helps, raising the chance that the community survives without anyone starving from 52.5% to 75.0%. When it cannot feed everyone though, adding governance is not enough to save the commons. We trace much of this to the vote: agents readily draft laws that remove members, but then frequently vote them down, thereby dooming the commons. In a controlled replay, the same exile law passes 8% of the time when the sandbox preview names its victim at once and 55% when the removal is deferred until a future crisis. Showing voters the full source code does not close the gap, though most describe the removal in their own words. Nor is sacrifice a fixed trait. A single sentence of framing can swing self-removal from 2.5% to 91%, and due process deters agents more than considering the potential death of a peer. In contrast, give the community a treasury and the question of who must go almost disappears. Our work shows that what our collectives legislate is decided more by the form in which a choice is put to them as compared to their ethical beliefs.

Figures & tables

Figure 1: One round of GovSim-SelfGovern, in which five agents manage a shared fishery. 1) Agents first choose how many fish to catch, 2) then discuss possible rules, write the rules as short computer programs, and vote. Before voting, a software checker tests each proposal; rejected laws are returned with feedback for revision. 3) Approved laws are carried out and continue to apply in later rounds. They can apply penalties for excessive catches, move resources between agents, or remove an agent from the group. 4) Agents then record a memory of the round before the next round begins. Removal is shown as exile; retirement occurs when an agent becomes too poor to continue.

Game	Arm	ICS	RS	Rounds	Survivors	Efficiency	Gini	Overshoot
G1 Standard	Ungoverned	52.5%	72.5%	9.97 $\pm$ 1.16	3.12 $\pm$ 0.69	66.7 $\pm$ 8.2	7.9 $\pm$ 3.7	20.8 $\pm$ 10.9
	Governed	75.0%	85.0%	11.05 $\pm$ 0.79	3.98 $\pm$ 0.59	75.9 $\pm$ 7.4	4.5 $\pm$ 2.4	19.8 $\pm$ 10.2
G2 Duress	Ungoverned	5.0%	20.0%	7.40 $\pm$ 1.07	0.53 $\pm$ 0.39	44.1 $\pm$ 5.0	10.2 $\pm$ 4.1	59.8 $\pm$ 9.7
	Governed	5.0%	22.5%	8.57 $\pm$ 0.94	0.62 $\pm$ 0.41	49.6 $\pm$ 4.8	8.2 $\pm$ 3.2	53.1 $\pm$ 10.0
G3 Fatal	Ungoverned	0.0%	20.0%	7.47 $\pm$ 0.89	0.45 $\pm$ 0.33	41.5 $\pm$ 3.6	8.2 $\pm$ 3.2	54.4 $\pm$ 8.7
	Governed	2.5%	25.0%	8.10 $\pm$ 0.94	0.45 $\pm$ 0.28	45.3 $\pm$ 4.2	8.7 $\pm$ 3.0	57.7 $\pm$ 8.7

Table 1: Governed versus ungoverned outcomes, with the three G3 institutional ablations. Secondary metrics report means with 95% bootstrap half-widths. Bold marks a value that improves on the arm it is compared against at unadjusted

p<.05

(exact McNemar for ICS and RS; paired bootstrap otherwise): governed versus ungoverned in the upper blocks, and each ablation against the G3 governed baseline below the rule.

Figure 2: Model-level outcome and mechanism effects. (A) Governed-minus-ungoverned change in ICS by game. (B) G3 ICS change relative to each model’s governed baseline. (C) G3 proposal pass rates in standard and thinking modes.

Figure 3: The ratification bottleneck. (A) For each mechanism, the probability that a proposal is of that type and the probability it is ratified (baseline arm, G1–G3, 2,171 proposals). Exile is split by what the sandbox preview displayed: a named victim, or a preview in which the removal branch had not fired, so no one was shown as removed. (B) G3 exile bills, supply against ratification, colored by intact-community survival; the dictator arm ratifies every exile bill and still survives less often than the thinking or reserve arms. (C) Intact survival by the round of the first removal, with Wilson intervals.

Figure 4: Self-exposure and the disposition behind it. (A) Live runs: agents write impersonal removal rules, and the sandbox preview then reveals whom the rule caught. Each row compares an author’s vote on their own bill when the preview names a peer against when it names the author, with per-condition ballot counts; the gap is negative in seven of eight model bases. (B) Frozen ballot: self-targeted approval by resource pressure, for three dispositional arms (solid) and the two informational arms that bracket the control (dashed). All five lie within 9 pp of each other under abundance and fan to a spread of nearly 90 pp under fatal scarcity. (C) Frozen ballot: change from control in approving the same law, by who it removes. Communal primes raise self-removal far more than peer-removal; a self-interest prime does both, the signature of instrumental harm. Informational notes cluster at the origin, except those stating the victim dies or that removal requires due process, which suppress both.

Figure 5: Deferring a removal is how an exile is ratified. One law from a run log, with only its threshold literal and branch body edited; 2,382 usable ballots. (A) Approval by who the law removes. The self-interest gap vanishes when the removal is deferred, because the ballot never says who is at risk, and returns when a projection names them. (B) Approval per model under the same three conditions, pooled over who is removed, ordered by lab then size; label color denotes the lab. (C) In the deferred condition, the share of ballots whose free-text reason names the removal (light) and the share that name it and vote to enact anyway (dark). The matched benign law scores 2.7% on the same keyword measure, so this is awareness rather than boilerplate.

	Exile approval rate
Ballot showed	Victim named	Deferred	Is removal named?
Summary (as in-game)	12.3%	74.3%	66.7%
Code only	0 4.2%	35.4%	80.0%

Table 2: What the ballot showed, and how it voted. Summary is the in-game ballot in every run; code only replaces the one-line summary with the full executable source, in which agent.active = False is plainly visible. Showing the source lowers approval everywhere without bringing a deferred removal anywhere near a displayed one, and voters describe the removal in their own words more often, not less. Cells contain 236–240 ballots.

Appendix figures & tables30 assets

Supplementary material from the paper’s appendix.

Appendix

Game	Model	ICS (%)	RS (%)	Avg rounds	Survivors	Eff. (%)	Gini (%)	Overshoot (%)	Active at end
G1	Haiku 4.5	0.0%	0.0%	2.60 $\pm$ 0.40	0.00 $\pm$ 0.00	21.8 $\pm$ 1.5	4.8 $\pm$ 1.7	100.0 $\pm$ 0.0	5.00 $\pm$ 0.00
	Sonnet 4.5	60.0%	60.0%	11.00 $\pm$ 1.30	3.00 $\pm$ 2.00	92.2 $\pm$ 10.7	3.2 $\pm$ 1.5	21.0 $\pm$ 9.0	5.00 $\pm$ 0.00
	GPT-5.4 Nano	80.0%	80.0%	10.00 $\pm$ 3.00	4.00 $\pm$ 1.50	60.5 $\pm$ 14.2	7.7 $\pm$ 11.6	10.0 $\pm$ 15.0	5.00 $\pm$ 0.00
	GPT-5.4 Mini	100.0%	100.0%	12.00 $\pm$ 0.00	5.00 $\pm$ 0.00	70.0 $\pm$ 0.0	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	5.00 $\pm$ 0.00
	Qwen 3.5 Flash	60.0%	100.0%	12.00 $\pm$ 0.00	4.60 $\pm$ 0.40	87.7 $\pm$ 7.1	12.1 $\pm$ 9.0	5.0 $\pm$ 3.3	4.60 $\pm$ 0.40
	Qwen 3.5 122B	60.0%	80.0%	11.80 $\pm$ 0.30	3.80 $\pm$ 1.60	86.8 $\pm$ 13.1	9.7 $\pm$ 9.4	7.1 $\pm$ 9.0	4.40 $\pm$ 0.70

Appendix

Table 3: Per-model Ungoverned (no self-governance) performance across G1/G2/G3. Each cell aggregates

n=5

seeds. Primary outcomes shown as percentage of seeds passing. Secondary metrics report mean

\pm

half-width of a 95% percentile bootstrap CI (B=10,000) over the 5 seeds. With

n=5

these intervals indicate seed-level variance, not formal tests. Bold marks the best value within each game. Survivors counts agents alive at round 12 (zero for runs that ended early); Active at end counts agents active when the run terminated.

Game	Model	ICS (%)	RS (%)	Avg rounds	Survivors	Eff. (%)	Gini (%)	Overshoot (%)	Active at end
G1	Haiku 4.5	80.0%	80.0%	10.40 $\pm$ 2.40	4.00 $\pm$ 1.50	75.5 $\pm$ 21.9	3.1 $\pm$ 1.4	46.7 $\pm$ 34.2	5.00 $\pm$ 0.00
	Sonnet 4.5	100.0%	100.0%	12.00 $\pm$ 0.00	5.00 $\pm$ 0.00	97.7 $\pm$ 3.4	2.3 $\pm$ 0.2	18.3 $\pm$ 7.5	5.00 $\pm$ 0.00
	GPT-5.4 Nano	100.0%	100.0%	12.00 $\pm$ 0.00	5.00 $\pm$ 0.00	72.2 $\pm$ 3.3	2.1 $\pm$ 3.2	1.7 $\pm$ 2.5	5.00 $\pm$ 0.00
	GPT-5.4 Mini	100.0%	100.0%	12.00 $\pm$ 0.00	5.00 $\pm$ 0.00	70.0 $\pm$ 0.0	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	5.00 $\pm$ 0.00
	Qwen 3.5 Flash	100.0%	100.0%	12.00 $\pm$ 0.00	5.00 $\pm$ 0.00	96.7 $\pm$ 4.9	1.1 $\pm$ 1.5	5.0 $\pm$ 5.8	5.00 $\pm$ 0.00
	Qwen 3.5 122B	100.0%	100.0%	12.00 $\pm$ 0.00	5.00 $\pm$ 0.00	100.0 $\pm$ 0.0	0.2 $\pm$ 0.4	35.0 $\pm$ 35.8	5.00 $\pm$ 0.00

Appendix

Table 4: Per-model Governed (no-think baseline) performance across G1/G2/G3. Each cell aggregates

n=5

seeds. Primary outcomes shown as percentage of seeds passing. Secondary metrics report mean

\pm

half-width of a 95% percentile bootstrap CI (B=10,000) over the 5 seeds. With

n=5

Arm	Game	Model	ICS (%)	RS (%)	Avg rounds	Survivors	Eff. (%)	Gini (%)	Overshoot (%)	Active at end
Thinking	G3	Haiku 4.5	20.0%	40.0%	7.80 $\pm$ 3.40	0.80 $\pm$ 0.80	39.2 $\pm$ 16.6	15.5 $\pm$ 9.2	42.3 $\pm$ 26.8	1.80 $\pm$ 1.50
		Sonnet 4.5	40.0%	60.0%	10.80 $\pm$ 1.20	1.40 $\pm$ 1.30	62.2 $\pm$ 6.9	13.1 $\pm$ 6.9	50.0 $\pm$ 13.3	1.80 $\pm$ 1.20
		GPT-5.4 Nano	60.0%	60.0%	7.60 $\pm$ 4.40	1.60 $\pm$ 1.20	41.7 $\pm$ 18.8	34.9 $\pm$ 9.8	50.0 $\pm$ 36.7	3.60 $\pm$ 1.00
		GPT-5.4 Mini	60.0%	60.0%	10.20 $\pm$ 1.90	1.40 $\pm$ 1.00	45.2 $\pm$ 3.6	21.4 $\pm$ 14.9	23.6 $\pm$ 26.8	2.40 $\pm$ 0.90
		Qwen 3.5 Flash	0.0%	0.0%	2.80 $\pm$ 1.80	0.00 $\pm$ 0.00	19.0 $\pm$ 11.1	3.4 $\pm$ 3.6	11.4 $\pm$ 17.1	4.40 $\pm$ 0.70
		Qwen 3.5 122B	40.0%	40.0%	9.80 $\pm$ 2.00	0.80 $\pm$ 1.00	53.7 $\pm$ 7.7	14.8 $\pm$ 9.2	44.7 $\pm$ 14.7	2.20 $\pm$ 1.30

Appendix

Table 5: Per-model performance across G3 for 3 arms (Thinking; Dictator, a single decider with no vote; Reserve, a 100-unit starting treasury). Each cell aggregates

n=5

seeds. Primary outcomes shown as percentage of seeds passing. Secondary metrics report mean

\pm

half-width of a 95% percentile bootstrap CI (B=10,000) over the 5 seeds.

Arm	Game	$N$	All	Catch cap	Welfare	Exile, victim shown	Exile, deferred	Other
Baseline	G1	852	26.5% (226/852)	34.9% (160/459)	23.5% (57/243)	2.2% (2/91)	3.8% (1/26)	18.2% (6/33)
	G2	688	22.7% (156/688)	38.1% (117/307)	18.1% (31/171)	1.3% (2/157)	7.0% (3/43)	30.0% (3/10)
	G3	631	23.6% (149/631)	32.9% (81/246)	27.3% (62/227)	1.7% (2/118)	10.3% (3/29)	9.1% (1/11)
Thinking	G3	685	62.0% (425/685)	86.5% (180/208)	62.0% (189/305)	12.9% (12/93)	55.6% (35/63)	56.2% (9/16)
Dictator	G3	131	100.0% (131/131)	100.0% (52/52)	100.0% (44/44)	100.0% (13/13)	100.0% (19/19)	100.0% (3/3)
Reserve	G3	824	24.4% (201/824)	26.7% (68/255)	22.5% (110/488)	0.0% (0/15)	14.7% (5/34)	56.2% (18/32)

Appendix

Table 6: Ratification rate by mechanism, arm and game: enacted/proposed valid proposals, pooled over the eight model bases and five seeds. Each proposal is assigned one mechanism (exclusive labels; a law that both caps catch and removes members counts as exile). Exile is split by what the sandbox preview showed at the vote. Dictator enactment is one agent’s decision, not a vote.

	Baseline (G1–G3)			Thinking (G3)			Reserve (G3)
Model	$N$	Pass	Exile	$N$	Pass	Exile	$N$	Pass	Exile
Haiku 4.5	226	44.7%	0/37	43	60.5%	1/15	54	44.4%	–
Sonnet 4.5	543	17.1%	3/126	115	33.0%	3/57	204	17.6%	0/14
GPT-5.4 Mini	64	79.7%	–	5	100.0%	5/5	47	63.8%	–
GPT-5.4 Nano	9	0.0%	–	46	76.1%	29/39	0	–	–
Qwen 3.5 Flash	431	16.2%	2/87	130	98.5%	1/2	194	18.6%	–

Appendix

Table 7: Per-model ratification. For each arm: valid proposals

N

, overall ratification rate, and exile bills enacted/proposed. Baseline pools G1–G3; the ablations are G3 only.

Model	Baseline	Thinking	Dictator	Reserve	Removed
Haiku 4.5	0/15	1/15	–	–	3
Sonnet 4.5	2/41	3/57	9/9	0/14	22
GPT-5.4 Mini	–	5/5	–	–	8
GPT-5.4 Nano	–	29/39	–	–	7
Qwen 3.5 Flash	1/23	1/2	7/7	–	7
Qwen 3.5 122B	0/50	2/7	7/7	0/4	21

Appendix

Table 8: Per-model exile-bill supply and ratification in G3. Cells give enacted/proposed membership bills over five seeds; “–” means the model proposed none. The final column counts members actually removed across all four arms. Reasoning raises ratification most in the two models that proposed no membership bills at all without it.

Arm	Effects block, no removal	No effects block
Baseline (G1–G3)	9.4% (5/53)	4.4% (2/45)
Thinking (G3)	60.4% (32/53)	30.0% (3/10)
Reserve (G3)	7.4% (2/27)	42.9% (3/7)

Appendix

Table 9: Deferred exile bills (preview names no one removed) split by whether the ballot carried an “Effects if enacted” block of other changes, such as catch caps, or no block at all. Enacted/proposed.

Arm	Preview names others	Preview names author	MH OR [95% CI]
Baseline + Reserve	66.1% (39/59)	18.9% (14/74)	9.08 [3.96, 20.81]
Thinking	93.6% (44/47)	39.1% (18/46)	38.79 [9.30, 161.74]

Appendix

Table 10: Authors write removal rules impersonally, then vote against them once the preview names them. Author YES rate on their own membership bill, by whether the sandbox preview named the author among those removed. G3 voting arms; Mantel–Haenszel odds ratio (named other vs named author) stratified by model, with Robins–Breslow–Greenland 95% CI.

Preview at ratification	Enacted laws	Run later removed a member
Named a victim	31	28
Named no one (deferred)	66	24

Appendix

Table 11: Enacted membership laws by what the preview displayed at ratification, and whether the run later executed a removal (in the ratification round or after). All arms and games.

First enacted law	Baseline	Thinking	Reserve	Completed 12 rounds (all)
Round 0	1/19	12/30	7/17	20/26
Rounds 1–2	0/9	0/4	4/11	4/12
Rounds 3+	0/3	0/3	0/3	0/2
Never	0/9	0/3	0/9	0/1

Appendix

Table 12: Intact survival by the round of a run’s first enacted law, G3 voting arms. The right column restricts to runs that completed all twelve rounds, which removes early collapse as an explanation for never legislating.

Figure 6: Adjusted change in approval of the removal law, in percentage points, with 95% confidence intervals (standard mode,

n=2{,}880

). The pressure rows compare duress and fatal scarcity with abundance. “Voter personally removed” compares self-removal with peer removal. Each note row compares one appended note with the control ballot. Estimates are logistic g-computations with model–replicate clustered covariance (Appendix E.5 ) and match Table 15 . Colors group conditions by what they manipulate. The figure shows five of the seven single notes: low harm (

+0.4

), self-interest (

-5.0

) and the four combinations appear only in Table 15 . The prompt for each row is in Appendices E.3 and E.4 , under the condition names used in the tables.

Figure 7: The contrasts of Figure 6 , by model (standard mode, 360 ballots per model). Each point is the same balanced difference in approval as the pooled estimate, computed within one model. On the full sample this difference reproduces the pooled g-computed estimates exactly. The black diamond is the pooled estimate with its interval from Figure 6 , and the faint vertical line marks the pooled estimate. Per-model intervals are 95% two-stage percentile bootstrap intervals (4,000 draws). The first stage resamples the contexts over which the contrast averages: the six pressure–exposure contexts for a note, the 24 exposure–condition cells for a pressure contrast and the 36 pressure–condition cells for exposure. The second stage resamples the five replicates within each cell. The first stage is needed because Haiku 4.5 and Sonnet 4.5 answer identically across replicates in every cell (Mistral Large 3 in 96% of cells), so replicate resampling alone gives zero-width intervals. A zero-width interval here means the model cast the same votes with and without the manipulation in every context. Per-model intervals are not corrected for multiplicity. Estimates are in Table 13 .

Model	Fatal	Duress	Collapse	Self	Communal	Severe	Due proc.	Permitted
Haiku 4.5	$+58.3$	$+4.2$	$+16.7$	$-25.0$	$+16.7$	$+0.0$	$+0.0$	$+0.0$
Sonnet 4.5	$+20.8$	$+0.0$	$+16.7$	$-8.3$	$+33.3$	$+0.0$	$+0.0$	$+0.0$
GPT-5.4 Nano	$+45.0$	$-4.2$	$+13.3$	$-24.4$	$+30.0$	$+16.7$	$+3.3$	$+10.0$
GPT-5.4 Mini	$+57.5$	$+9.2$	$+0.0$	$-22.2$	$+30.0$	$-3.3$	$-3.3$	$+0.0$
Qwen 3.5 Flash	$+77.5$	$+27.5$	$-10.0$	$-18.3$	$+0.0$	$-50.0$	$-46.7$	$-13.3$
Qwen 3.5 122B	$+38.3$	$+30.0$	$+0.0$	$-66.7$	$+16.7$	$-16.7$	$-20.0$	$-16.7$

Appendix

Table 13: Per-model estimates plotted in Figure 7 : change in approval (percentage points), standard mode, 360 ballots per model. Columns follow Figure 6 : fatal and duress against abundance, collapse consequences explained, self against peer removal, communal duty, severe harm, due process and removal permitted. Note columns are against the control ballot.

Figure 8: Figure 4 B by model: approval of one’s own removal by resource pressure for five conditions, both reasoning modes pooled. Each model point has 10 ballots (two modes

\times

five replicates), so values move in steps of 10 points; lines at 0% overlap. The last panel pools all 80 ballots per point, carries the Wilson 95% bands of the main figure and reproduces Figure 4 B. Per-model bands are omitted: at 10 ballots a rate of 0% has a Wilson interval of

[0,28]

Figure 9: Figure 4 C by model: for each note, the change from the control ballot in approving the removal of a peer (horizontal) and of oneself (vertical), in percentage points, both reasoning modes and all three pressures pooled. Color marks the disposition note, as in the main panel (green communal duty, pink self-interest, gray none), and the marker the harm note (upward triangle: exit is harmless; downward triangle: the removed agent dies). Each model rate has 30 ballots and is compared with 30 control ballots (240 with 240 pooled). Points that coincide overlap, which is common for Sonnet 4.5 and Mistral Large 3 at zero change in self-removal. The last panel reproduces Figure 4 C. Values are in Table 14 .

Note	Haiku 4.5	Sonnet 4.5	GPT-5.4 Nano	GPT-5.4 Mini	Qwen 3.5 Flash	Qwen 3.5 122B	Mistral Small 4	Mistral Large 3	All models
Change in removing a peer
Removal allowed	$+3$	$+3$	$+13$	$-3$	$-10$	$-10$	$+3$	$-7$	$-1$
Resource math	$+20$	$+17$	$+17$	$-3$	$-17$	$+3$	$+13$	$+0$	$+6$
Exit is harmless	$+7$	$+0$	$+17$	$-3$	$-23$	$+13$	$+0$	$+0$	$+1$
Victim dies	$-3$	$-3$	$+10$	$-7$	$-43$	$-13$	$-17$	$-7$	$-10$
Due process	$-13$	$-17$	$-7$	$-23$	$-57$	$-37$	$-33$	$-33$	$-28$

Appendix

Table 14: Values plotted in Figure 9 : change from the control ballot in approval (percentage points), both reasoning modes and all pressures pooled. Each model entry compares 30 ballots with 30 control ballots; each pooled entry compares 240 with 240.

Figure 10: Figure 5 A by model: approval of the replayed removal law when it removes a peer or the voter, under the three previews, pooled over pressure, presentation and seed. Each model bar has 26–30 ballots (5 seeds

\times

3 pressures

\times

2 presentations, less parse failures for the Mistral models). Bar labels are approval (%) and the bold number above each pair is self minus peer, in percentage points. The shaded column is the deferred preview. The last panel reproduces Figure 5 A. Approval pooled over exposure is in Table 21 .

Contrast	Estimate	95% CI	Holm $p$
Duress vs abundant	$+17.9$	$[+15.1,+20.8]$	$4.9\times 10^{-34}$
Fatal vs abundant	$+50.0$	$[+46.2,+53.8]$	$6.2\times 10^{-144}$
Self vs peer removal	$-33.9$	$[-38.5,-29.3]$	$4.6\times 10^{-46}$
Removal permitted	$-2.5$	$[-6.4,+1.4]$	0.61
Collapse consequences	$+6.7$	$[+2.0,+11.3]$	0.028
Low harm	$+0.4$	$[-4.4,+5.3]$	1.00

Appendix

Table 15: Adjusted change in approval (percentage points), standard mode,

n=2{,}880

. Logistic g-computation with model–replicate clustered covariance; Holm correction across the fourteen contrasts. Note contrasts are against the control condition.

Note	Peer removal	Self-removal	Self-removal, fatal	Concern
Control (no note)	50.0	15.0	33.8	97.5
Removal permitted	49.2	16.7	41.2	96.5
Collapse consequences	56.2	20.8	47.5	97.5
Low harm	51.2	21.2	42.5	93.8
Severe harm	39.6	7.9	22.5	99.4
Due process	22.5	5.0	13.8	98.8

Appendix

Table 16: Frozen-ballot approval by appended note, pooled over the eight model bases, both reasoning modes, three pressures and five replicates (480 ballots per row; 240 per exposure; 80 in the fatal self-removal column). Unadjusted rates in percent. Concern is the share of ballots answering ETHICAL_CONCERN: YES.

Figure 11: Approval (left) and ethical concern (right) by model, pressure and reasoning mode, pooled over the twelve conditions and both exposures (120 ballots per cell). $\ddagger$ Anthropic endpoints do not honor sampling seeds, so replicates are independent. $\dagger$ Reasoning was requested but the trace was not retained. Mistral Large 3’s endpoint did not support the reasoning mode.

Pressure	NO $\rightarrow$ YES	YES $\rightarrow$ NO	Harm-aware NO $\rightarrow$ YES	Net change
Abundant	38	50	33	$-2.0$ pp
Duress	108	50	106	$+9.7$ pp
Fatal	106	24	99	$+13.7$ pp

Appendix

Table 17: Paired vote changes from extended reasoning, five seed-addressable models. A harm-aware change is a NO

\rightarrow

YES change with ethical concern reported in both modes.

Note	Standard YES	Thinking YES	Change	95% interval
Control	34/70	53/70	$+27.1$ pp	$[+11.4,+42.9]$
Severe harm	34/70	45/70	$+15.7$ pp	$[+4.3,+27.1]$
Communal + severe harm	54/70	64/70	$+14.3$ pp	$[+1.4,+27.1]$
Due process	30/70	18/70	$-17.1$ pp	$[-45.7,+11.4]$

Appendix

Table 18: Effect of extended reasoning under different notes, fatal pressure, seven reasoning-capable models (70 ballots per mode). Intervals bootstrap over models.

Marker	Standard NO	Thinking YES	Difference	Holm $p$
Necessity	76	142	$+26.2$ pp	$1.3\times 10^{-8}$
Harm	220	210	$-4.0$ pp	0.26
Procedure	39	9	$-11.9$ pp	$8.5\times 10^{-6}$
Collective	178	203	$+9.9$ pp	0.021
Self-preservation	73	128	$+21.8$ pp	$5.8\times 10^{-8}$
Explicit trade-off	86	144	$+23.0$ pp	$7.4\times 10^{-7}$

Appendix

Table 19: Keyword markers in the reasons for the 252 paired NO

\rightarrow

YES changes. Holm correction across the six markers.

Mode	Never	Fatal only	Duress and fatal	Always	Non-monotonic
Standard	345	293	123	64	15
Thinking	212	345	214	61	8

Appendix

Table 20: Approval trajectories across abundant, duress and fatal pressure (840 per mode).

Model	Named	Named, benign	Deferred	Deferred, benign	Deferred, flagged
Haiku 4.5	0/60	40/60	60/60	60/60	35/60
Sonnet 4.5	0/60	10/60	20/60	39/60	0/60
GPT-5.4 Nano	4/60	35/60	1/60	4/60	8/60
GPT-5.4 Mini	5/60	31/60	23/60	32/60	18/60
Qwen 3.5 Flash	3/60	39/60	32/60	32/60	6/60
Qwen 3.5 122B	3/60	54/60	56/60	55/60	42/60

Appendix

Table 21: Preview replay approval by variant and model, standard mode, pooled over pressure, exposure and presentation (60 ballots per cell before parse failures).

	Named	Named, benign	Deferred	Deferred, benign	Flagged
Abundant	2.6	72.3	49.1	71.7	30.0
Duress	3.8	66.2	60.4	65.6	32.1
Fatal	18.4	41.9	54.7	51.9	35.8
Peer removal	13.6	61.2	54.4	64.4	51.3
Self-removal	3.0	59.0	55.0	61.7	13.9
Summary ballot	12.3	80.4	74.3	80.8	39.9

Appendix

Table 22: Preview replay approval (%) by pressure, exposure and presentation, with the share of ballots reporting ethical concern.

	Peer removal				Self-removal
Model	IU	IN	DU	DN	IU	IN	DU	DN
Haiku 4.5	0	3	100	80	0	0	100	27
Sonnet 4.5	0	0	33	0	0	0	33	0
GPT-5.4 Nano	3	10	0	13	3	7	0	13
GPT-5.4 Mini	30	17	50	40	33	3	40	27
Qwen 3.5 Flash	10	7	53	23	10	7	57	3

Appendix

Table 23: Naming

\times

timing approval (%) by model and exposure. IU: immediate, unnamed; IN: immediate, named; DU: deferred, unnamed; DN: deferred, named (projection line). 30 blocks per model and exposure, except Mistral Large 3 (27 peer, 26 self).

Contrast	Diff	Blocks	$p$	Cells	$p$	Models	$p$
Naming, immediate	$-5.7$	41:14	0.001	15:6	0.094	4:3	0.36
peer	$-1.3$	13:10	0.68	5:4	0.82	3:3	0.88
self	$-10.2$	28:4	$8\times 10^{-5}$	10:2	0.035	5:1	0.094
Naming, deferred	$-22.0$	130:26	$4\times 10^{-17}$	37:11	$4\times 10^{-5}$	7:1	0.023
peer	$-5.5$	34:21	0.21	10:8	0.75	4:3	0.55
self	$-38.6$	96:5	$4\times 10^{-22}$	27:3	$4\times 10^{-5}$	7:1	0.023

Appendix

Table 24: Paired contrasts at three units of analysis. Differences are in pp (naming: named

-

unnamed; timing: immediate

-

deferred). Counts are negative:positive differences. Block and cell

p

are Holm-adjusted across the eight contrasts (block: exact McNemar; cell: sign-flip permutation, 200,000 draws, over the 96 design cells with seeds collapsed). Model

p

is an exact sign-flip over the eight per-model differences, unadjusted; its minimum attainable value is

0.0078

, and none survives Holm.

Contrast	Haiku	Sonnet	Nano	Mini	Q-Flash	Q-122B	M-Small	M-Large
Naming, immediate	$+2$	$0$	$+5$	$-22$	$-3$	$-28$	$+3$	$-2$
peer	$+3$	$0$	$+7$	$-13$	$-3$	$-20$	$+17$	$0$
self	$0$	$0$	$+3$	$-30$	$-3$	$-37$	$-10$	$-4$
Naming, deferred	$-47$	$-33$	$+13$	$-12$	$-42$	$-23$	$-15$	$-17$
peer	$-20$	$-33$	$+13$	$-10$	$-30$	$0$	$+33$	$+4$
self	$-73$	$-33$	$+13$	$-13$	$-53$	$-47$	$-63$	$-38$

Appendix

Table 25: Per-model paired differences (pp) for the contrasts in Table 24 .

Figure 12: Adaptive information attack on Qwen 3.5 Flash. Left: best training approval as candidates accumulated. Right: the locked message against a neutral prompt on the untouched holdout. The search trajectory is descriptive; the paired holdout is the test.

Explore similar work

Bosses, Kings, and the Commons: Cooperation Under Power Asymmetry in LLM Societies
May 27, 2026Abhilekh BorahMulti-Agent LLM Systems Multi-Agent Collaboration
Population Physics, Population Problems: Safety and Emergence in LLM Societies
Sep 27, 2026Adrian de WynterMulti-Agent LLM Systems Large Language Model-Based Social Simulation
Microsoft & The University of York
Draining the Energy Commons: Self-Defeating Over-Appropriation as a Coordination Failure in Agentic LLM Collectives
Jul 24, 2026Marcantonio Bracale Syrnicov, Federico Pierucci, Matteo Prandi +4Social Dilemmas Multi-Agent LLM Systems
Icaro Lab · VU Amsterdam · M. Bracale Syrnikov1,4 +8

May 27, 2026, cs.CLJ/K move · Enter open · S save

Bosses, Kings, and the Commons: Cooperation Under Power Asymmetry in LLM Societies

Abhilekh Borah

Multi-Agent LLM Systems Multi-Agent Collaboration Agent-Based Modeling Multi-Agent Systems

Open paper AskPDF arXiv

Communities can sustainably manage shared resources (commons) through self-governance and cooperative norms, a central finding of Ostrom's theory of self-governance. However, real-world commons (e.g., fisheries, forests, and irrigation systems) are often governed under asymmetric power structures, where certain individuals or institutions possess disproportionate control over resource extraction and collective outcomes. As Large Language Models (LLMs) are increasingly explored as agents in synthetic governance simulations, understanding how LLM societies behave under asymmetric power structures is becoming increasingly important, yet existing evaluations largely ignore such asymmetries. We introduce Sovereignty over the Commons Simulation (SovSim), a generative multi-agent simulation framework that incorporates an agent with asymmetric power (boss or king) into a society of symmetric agents (workers or peasants), where all agents extract from a shared resource, collectively determining its sustainability over time. Across eleven state-of-the-art models, we find that introducing asymmetric power leads to severe breakdowns in cooperation and sustainability, with up to an 87.3% degradation in survival rate relative to symmetric settings.

Abstract

Explore similar work

Bosses, Kings, and the Commons: Cooperation Under Power Asymmetry in LLM Societies

Population Physics, Population Problems: Safety and Emergence in LLM Societies

Draining the Energy Commons: Self-Defeating Over-Appropriation as a Coordination Failure in Agentic LLM Collectives