# Accepted Microduck regression baselines These nine trajectories were generated from original source commit `44d3681726468bdd00e88241894d5faa53e3c63e`, before changing physics / policy semantics. The reference tree was independently materialized under `/tmp` from `git archive`, not taken from the newly refactored runtime. Reference-only harness changes: 1. Stop the background async control loop and allow single-owner explicit steps. 2. Skip the title / BIOS render gates in test mode. 3. Disable automatic ball spawning in the test-mode ceremony callback and park the ball, isolating policy trajectories from cosmetic randomness and collisions. 4. Expose existing mode controls / observation / state for measurement. 5. For the original wall-clock sit timer, wait 850 ms before the recorded physics interval. The refactored timer receives an equivalent explicit virtual 850 ms advance; the recorded input/target sequence is unchanged. The first experimental harness did not fully isolate automatic ball spawning. Changing scene initialization then changed RNG consumption and revealed that problem. Those early fixtures were discarded. Accepted fixtures were regenerated from the independently archived **original** code with corrected isolation, then compared against the new implementation. No tolerances were widened to conceal that harness issue. Environment: MuJoCo WASM 3.11.0, onnxruntime-web WASM 1.27.0, one thread, Chromium headless (Playwright 1.64.1), fixed initial state / commands. Each case settles for 50 control ticks and records six snapshots spaced 25 ticks apart: qpos, qvel, ctrl, observation, lastAction, mode, recovery, simulation time. Cases: stand, walk, kick, sitstand, groundpick, roll, recovery, rollers, crouch. Comparisons are exact, including all recorded robot state; the parked ball is also present in each model state. They are representative trajectory windows, not proof that every possible long-running skill sequence is covered. After the user-requested expansion to a 4.5 m default arena, this original-code oracle explicitly pins the **original 3 m World** through a test-only World response override. The production default remains 4.5 m, tested independently in `bounded-world.spec.js`. No accepted trajectory or tolerance was replaced to hide the intentional World-size change. After replacing the production football with 5–7 random free-body geometries, the archived oracle additionally uses **`?test=1&oracle=legacy`**. `src/robots/microduck/test-oracle.js` supplies the original single parked sphere's dimensions, mass and contact parameters, with no football texture or gameplay. This keeps the oracle's model DOF count and floating-point accumulation unchanged; the nine accepted trajectories and exact comparisons remain untouched. `?test=1` alone still builds the real 5–7 toy scene; `prepareTest()` parks those toys for isolated stepping, and `scene-interaction.spec.js` explicitly restores them for real collision / drag tests. Production never selects the legacy probe unless both explicit test parameters are present. The Microduck composer fixes `jacobian="dense"` to retain the original solver path when the added toy DOFs cross MuJoCo's automatic sparse threshold. `npm run test:regression` reads these files. `npm run baseline:record` writes candidate files under ignored `test-results/baseline-candidates/`, never here. A baseline replacement needs an explicit reviewed cause and updated provenance. Node/native Python trajectories are not substituted for these WASM baselines.