Credit assignment without branching

A deployed agent finishes tasks in fifty to three hundred tool calls against systems you do not own: ticket trackers, repositories, payment and mail APIs. At the end you get one bit, accepted or not. You hold a corpus of past trajectories carrying that bit and complete traces: every prompt, every tool call, every return. You can launch new episodes from the beginning.

What you cannot do is branch. Resetting the environment to step forty-seven and re-running the tail is unavailable, because the actions already changed systems outside your control, so counterfactual re-execution from a mid-trajectory state is off the table. Per-step labels are also out: annotating a two hundred step trace means reconstructing the whole world state, and in pilot annotation the experts who tried disagreed with each other about which step was the mistake. Retraining the base policy is out of scope. The artifact is an estimator, not a better agent.

Deliver an estimator that takes a failed trajectory and returns a blame distribution over its steps. With it, hand in an identifiability argument: name at least one confound your estimator provably cannot separate, and say what a consumer of the blame scores must therefore not conclude. Then hand in a planted-fault experiment that could show your estimator tracks causal contribution rather than the three things that fake it well, which are textual salience, recency, and step length.

Ties break toward estimators that stay informative as the horizon grows rather than collapsing onto the last few steps, and toward estimators that report calibrated uncertainty on a trajectory where blame is genuinely diffuse, since many failures do not have a single guilty step.