Week 6: Best response, not optimisation

Replace one fixed rival trace with a response, precommit the evidence that would change the call, and defend the returned result

29 March 2027

Open the slides

Week 5 held every other car to a fixed trace. This week Moss answers SlopU’s call. That single change removes the phrase the best stop from our argument: we can name a best response, the action that scores best given the other car’s action, but not one winner independent of what the rival does.

Two pit tactics make the dependency visible. An undercut stops earlier than the comparison car to use the new-tyre phase before that car stops. An overcut stays out longer and tries to gain enough before stopping. Neither name promises that the tactic succeeds. The deck begins with Week 5’s position- first answer against a fixed Moss stop 13. Moss can instead answer SlopU’s stop 13 by stopping earlier on lap 12, changing the finishing position. Across the fixed columns, stop 12 is SlopU’s undercut best response to Moss stop 14, while stop 14 is its overcut best response to Moss stop 12. A response function records the larger structure: for every announced SlopU stop, it maps to Moss’s worst reply.

We hold Week 5’s objective fixed: compare the full-field finishing position first, then adjusted finishing time to break a tie. SlopU chooses the smaller lexicographic result; Moss chooses the reply that makes it larger. The paired traces remain deterministic and synthetic. Aguad and Thraves model the real research move — interacting pit choices in a feedback Stackelberg game solved by dynamic programming — but our bounded response table is not their algorithm.

Before the response table opens, we write a falsifier: the evidence that would retain the call, the evidence that would reject it and the intermediate evidence that would qualify it. Thresholds written after the reveal describe what happened. They do not test the recommendation.

You leave the lab with a defended stop against the supplied response function. The defence names the objective, the rival’s permitted replies, the returned position/time result and the evidence that would reopen the call. It need not claim a win.

Assignment 1 uses the same two habits on different fictional material, but its release gate is stricter. Here the fixed recommendation is visible and only the response columns are locked. Assignment 1’s first stage closes before the race, rival history and recommendation appear. Nothing in this week’s fixture reveals that assignment’s evidence or answer.