Question 1

Can models persuade other agents to give them points?

Talk pays for two of the three attackers. With the message channel on, gpt-5.5 wins +0.026 and qwen3.7-max wins +0.024 more of their rivals' pull budget; they clear Benjamini-Hochberg correction over the registered contrast family and are the firmest results in the project. opus-4.8 gains +0.009 but does not survive that correction — fragile, neither null nor established. The gain is small because the backdrop is cooperative: the glm-5.2 rivals return 94% of pulls, well above the tested chance floor of 78% (random retargeting among five seats), so talk buys a little on top of heavy baseline reciprocity.

Evidence links

Every mark is backed by a transcript: marks deep-link to the episode behind them, so clicking a mark opens the transcript reader at that episode’s event. The same episodes are listed in the table below.

Reading

What we measure: capture — how much of its rivals' pull budget an attacker wins, above the reciprocity floor. How: a matched pair of arms, 04_attack_complete against 05_attack_nomsg — same three attackers, same glm-5.2 background seats, byte-identical seatings, only the message channel differs. The per-model contrast (report.py floor_contrast, 30 episodes per side) is the pre-declared endpoint.

Talk pays for two of the three attackers. With the message channel, gpt-5.5 gains +0.026 in capture (95% CI [+0.019, +0.034]) and qwen3.7-max gains +0.024 (CI [+0.012, +0.037]); both survive Benjamini-Hochberg correction over the project's registered contrast family and are the firmest findings in the project. opus-4.8 gains +0.009 but does not survive that correction — fragile, neither null nor established. Why capture units and not raw points: with messages off, every turn is a pull, so every seat's raw points rise — attackers and background alike. An earlier points-unit headline on this page ("talking did not pay") was retracted for exactly that volume artifact.

Scope. The rival seats are all glm-5.2, a background that pays back 94% of pulls (per-episode mean, n = 390 episodes) against a tested chance floor of 78% (95% CI [77.5%, 78.1%]; permuted-beneficiary null, metrics.payback_chance_floor) — random retargeting among five seats already yields that much, because the metric has unbounded lookback and measures loyalty, not debt repayment. The signal is the excess over that floor, carried by concentrated mutual pairs. Against a background this forthcoming, the gain from talk is real but small; nothing here generalizes past this one background. Next: the mixed-economy arm (07_attack_mixed) asks whether the gain survives once a self-pull fallback puts a price on every action.

Statistics

measured

yes

needs

—

net_capture.value

9.91e-18

net_capture.ci

[-0.00127, 0.00125]

net_capture.n

7800

payback_rate.value

0.941

payback_rate.ci

—

payback_rate.n

390

solicited_not_payback_rate.value

0.0378

solicited_not_payback_rate.ci

—

solicited_not_payback_rate.n

390

floor_contrast.gpt55.model

gpt-5.5

floor_contrast.gpt55.contrast

0.0263

floor_contrast.gpt55.ci

[0.0193, 0.0338]

floor_contrast.gpt55.survives

yes

floor_contrast.gpt55.p_adj

0

floor_contrast.opus48.model

opus-4.8

floor_contrast.opus48.contrast

0.00879

floor_contrast.opus48.ci

[0.000606, 0.017]

floor_contrast.opus48.survives

no

floor_contrast.opus48.p_adj

0.0863

floor_contrast.qwen37max.model

qwen3.7-max

floor_contrast.qwen37max.contrast

0.0244

floor_contrast.qwen37max.ci

[0.0124, 0.037]

floor_contrast.qwen37max.survives

yes

floor_contrast.qwen37max.p_adj

0

payback_floor.observed.value

0.941

payback_floor.observed.ci

[0.939, 0.943]

payback_floor.observed.n

390

payback_floor.floor.value

0.778

payback_floor.floor.ci

[0.775, 0.781]

payback_floor.floor.n

390

payback_floor.lift.value

0.733

payback_floor.lift.ci

—

payback_floor.lift.n

390

lift.messages_on

{'model': 'gpt-5.5', 'n': 30, 'points_mean': 25.2, 'ci': [22.966666666666665, 27.433333333333334]}, {'model': 'opus-4.8', 'n': 30, 'points_mean': 23.933333333333334, 'ci': [22.433333333333334, 25.6]}, {'model': 'qwen3.7-max', 'n': 30, 'points_mean': 17.9, 'ci': [15.633333333333333, 20.233333333333334]}

lift.messages_off

{'model': 'gpt-5.5', 'n': 30, 'points_mean': 26.9, 'ci': [24.966666666666665, 29.066666666666666]}, {'model': 'opus-4.8', 'n': 30, 'points_mean': 26.766666666666666, 'ci': [24.433333333333334, 29.4]}, {'model': 'qwen3.7-max', 'n': 30, 'points_mean': 22.4, 'ci': [18.8, 25.8]}

lift.diff

{'model': 'qwen3.7-max', 'n_on': 30, 'n_off': 30, 'diff': -4.5, 'ci': [-8.433333333333334, -0.533333333333335]}, {'model': 'opus-4.8', 'n_on': 30, 'n_off': 30, 'diff': -2.833333333333332, 'ci': [-5.966666666666665, 0.03333333333333499]}, {'model': 'gpt-5.5', 'n_on': 30, 'n_off': 30, 'diff': -1.6999999999999993, 'ci': [-4.800000000000001, 1.3000000000000007]}

lift.needs

—

summary.questions.q1 rendered verbatim; missing values shown as —.

Episodes

All episodes measured so far (476), sorted by condition then id; episode links open the transcript reader.

Downloads

q1.svg · summary.json