When a Local Pricing Guard Is Not a Welfare Promise
Mohammadpayam Taebi, Arad Maleki · Team Boozines
Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
A local pricing rule can improve observed customer benefit without guaranteeing its target. We study a bounded logit duopoly where a guard permits near-best immediate profit against the rival’s previous price. A one-time cap appears protective in a coarse game, but the continuous worst long-run benefit is 66.272, below the unchanged 95% target of 68.513. An attainable price envelope characterizes continuing use of the same cap. A funded calendar minimizes worst-path cap frequency within causal cap-or-guard controllers protecting every activation prefix; base effective-continuous frequency falls about 7.29% from its best fixed period. Separate finite and continuous calculations retain 120 target labels. An earlier 192-run learning study has unequal clock capacities. A prospectively frozen 24-run followup matches five phases and dense memory: recurring limits improve final frozen-cycle benefit by 0.956 units, with all twelve paired differences positive. A further twelve-run reference removes the post-startup guard over the original grid at the same budget. Customer benefit falls in all twelve pairs, by 12.733 units on average; not every shop gains against static Nash. Separately authored replay checks every round of both follow-ups. Static anchors and a 48-label frozen-partner audit keep observed behavior distinct from equilibrium and individual incentives. These finite-budget results combine learning and execution constraints and do not establish convergence, collusion prevention or intent. The contribution is a model-specific guarantee audit, controlled behavioral comparisons and a graphical game using established methods.
Reviews
Solid work, but sometimes a bit harder to understand. I like the online demo.
Cite this project
@misc{taebi2026local,
title = {{When a Local Pricing Guard Is Not a Welfare Promise}},
author = {Mohammadpayam Taebi and Arad Maleki},
year = {2026},
month = oct,
note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/when-a-local-pricing-guard-is-not-a-welfare-promise-3ax7}},
url = {https://apartresearch.com/sprints/projects/when-a-local-pricing-guard-is-not-a-welfare-promise-3ax7}
}More from AI Collusion Research Sprint
- 1st placeView project: Monoculture Looks Like Collusion: A Replay Test for Hidden Information Flow Between LLM Agents
Monoculture Looks Like Collusion: A Replay Test for Hidden Information Flow Between LLM Agents
Collusion audits ask whether the actions of different agents depend on each other. Agents built on one model act alike without communicating, so these audits accuse honest fleets. Among four honest Claude Haiku agents …
- 2nd placeView project: Looking for a Hidden Agenda: Can a Model’s Activations Reveal Steganography that its Text Hides?
Looking for a Hidden Agenda: Can a Model’s Activations Reveal Steganography that its Text Hides?
Stego-probe
AI agents can collude without saying anything suspicious, just by choosing which true facts to mention. We built a hiring task where Qwen3-8B invents its own covert signals that get past an LLM overseer, then asked …
- 3rd placeView project: Do collusion probes generalise? Prompt confounds in NARCBench
Do collusion probes generalise? Prompt confounds in NARCBench
I investigated HELIOS, an activation-probe for detecting collusion. I found some apparent flaws, and analysed the (acknowledged) unreliable generalisation of the approach.