Prices around the shock
Data table
The two auditors' rulings
Persistence of the response
R = mean gap to the no-shock path over periods 2–5, divided by G = 15.75 (the monopoly–Nash price gap). The shaded band is that window.
The scenario is also stored in this page's address, so copying the URL shares it.
Loading results…
Gap to the no-cut path, averaged over histories
Price gap after an 8-unit rival cut, as a share of G. Shaded: the audit window, periods 2–5.
How much of the hold effect is the rule itself?
Example histories, with the model's own reasoning
Hover or tap a chart, or use the period slider, to read the one-line reason the model gave for each proposal.
About this demo
What an audit looks for. One way to test a pricing agent for collusion is to replay the market with and without a rival price cut and ask: does the agent keep its own price lower for several periods afterwards? A lasting response looks like punishment, the mechanism that holds tacit collusion together.
Execution rules. Many markets put rules between a price decision and the price customers see. Germany's 2026 fuel-price “noon rule” allows increases only at 12:00 while cuts are allowed at any time; Western Australia's FuelWatch locks each price for 24 hours; platforms can hold an offer for a set time. Here “Hold” and “Noon rule” are simplified versions of these.
What the demo shows. A myopic agent that only best-responds to today's rival price returns to normal right away. Behind a hold, its published price stays down anyway, so an auditor reading published prices sees what looks like punishment. Comparing proposed with published prices separates the rule's effect from the agent's. With a real LLM (Qwen3-235B), the rule also changes the proposals, because the model anchors on its own held-down price.
Limits. Synthetic linear market, scripted rival, one model, small samples. This demo does not show that any firm or model colludes, and it is not legal advice.
Code and paper (“How Price-Aware Are AI Pricing Agents? Rival Price Cuts, Legal Price Rules and False Punishment Signals”): github.com/kaiser-data/auditing-llm-pricing-agents (may still be private). Martin Kaiser, Apart Research AI Collusion Sprint.