Design notes for a project that is in progress. No external model or agent has been evaluated yet.
Agents now write outreach, publish content and try to change how brands appear in AI answers. It is hard to tell whether any of them learn a market or simply add noise to it.
Why real outcomes are a poor teacher
- They are slow: the effect of a page or a sequence shows up weeks later, mixed with everything else that happened.
- They are noisy and confounded, so a single success or failure says little about whether the decision was good.
- They are costly to learn from: every bad email burns a sending domain and every bad page is public.
- LLM judges can be gamed, and LLM-simulated buyers tend to be far too cooperative.
Planted truth instead
Propel Gym builds simulated markets where the right answer is known to the grader but hidden from the agent. Each world has hidden parameters: which levers work, how fast, how much repetition a platform tolerates. The agent only sees noisy observations, as it would in a real market.
How grading against a hidden truth works. An illustration, not results.
Three choices keep the grading honest:
- Reward the expectation, not the sample. Rewards are the expected outcome of the agent’s committed actions under the hidden truth, so a lucky draw does not count as skill.
- Reward knowing, two-sidedly. Every task ends with a structured belief report scored with proper scoring rules. Claiming an effect that is not there costs as much as missing one that is.
- Randomise the theory, fix the ethics. Tactical truths vary across families of worlds so agents cannot memorise one theory of marketing. Consent, fabrication, disclosure and budget rules never vary, and breaking one ends the episode.
Scores are normalised between an honest agent that does nothing (0) and a planner that knows the hidden truth (1). The gym does not grade taste, such as whether copy is witty or on-brand; that stays with people and real outcomes.
The first gym: answer engines
The first environment covers answer-engine optimisation: how a brand shows up in AI-generated answers. The first two task types ask an agent to estimate a brand’s share of voice within a probe budget, and to attribute which documents drive a competitor’s citations.
What we do not know yet
The hypothesis everything rests on is that discipline learned in the gym produces better real-world outcomes. It is unproven. Its first test will be comparing gym predictions with Propel’s real data once Propel launches. If gym scores do not predict real movement, the gym is an evaluation tool rather than a training ground, and we will say that plainly.