# Reproduce the physical feedback assay Kuber Mehta ยท FLM / ChatFLM The records archive contains all 27 physical conditions, the exact repeat, the four fixed sensory-choice checkpoints, their numerical parity fixtures, source files, the declared protocol and component licenses. It does not require access to the private repository. These checkpoints take four sensory inputs and predict two actions; they are separate from the language models. ## Audit existing observations Extract `closed-loop-records.zip` to a new directory and open a terminal there. Use Python with NumPy installed. The recorded simulator environment used Python 3.12.14 and NumPy 2.5.3; `requirements-embodied-lock.txt` lists the full physical environment. The audit itself needs NumPy and the included Python files, without FlyGym, MuJoCo, PyTorch, a GPU or a display: ```sh python scripts/audit_feedback_release.py ``` The audit first verifies every manifest payload and the four-model parity fixture. It then independently derives the sampled sensory cue from the recorded body pose, replays every neural frame, verifies the delayed query decision and reconstructs the actual motor-command timeline. It recomputes angular error, distance, phase-specific progress and the other metrics, then checks the repeated physical and neural arrays for exact equality. The complete cohort contains 54,027 physical observations, 10,800 control frames and 972 delayed decisions. The additional repeat contains 2,001 observations, 400 frames and 36 decisions. The three scripted cases have sensory and command records but no invented neural state or probabilities. State/logit parity uses absolute tolerance 2e-6 and relative tolerance 2e-5 together. A different NumPy runtime can produce small arithmetic differences; a failed tolerance check must be investigated rather than hidden by relabeling the record. The manifest establishes file identity, not a third-party signature. An audit of recorded data is distinct from rerunning MuJoCo physics. The archive also includes the original predeclared video for the eligibility/live/switch case. Two simulated seconds play in eight seconds. The target is a virtual coordinate; the body camera does not depict it as a visible object. ## Run fresh physics Use the pinned physical environment and an **empty, separate experiment directory**. Copy these paths from the archive, preserving their layout: - `experiments/embodiment/` for the simulator, controller and verifier; - `data/controllers/choice-v1/` for all four models and parity fixtures; - `docs/CLOSED-LOOP-PROTOCOL.md` and `requirements-embodied-lock.txt`; - `scripts/audit_feedback_release.py` for a subsequent numerical audit. Do not copy the existing `runs/` directory or its root manifest into a fresh experiment. From the new directory run: ```sh python experiments/embodiment/closed_loop.py python scripts/audit_feedback_release.py ``` The runner creates an identity from its sources, inputs and current environment, runs all 27 cases and repeats the predeclared case. On an interrupted run, use the same command after confirming that its previous process stopped. Completed trials resume only when their identities and artifact hashes agree. An OS-held lock prevents two simultaneous physical writers. A failed trial is recorded explicitly and requires investigation; the runner will not silently replace it. Do not overwrite the downloaded observations with a new run. Exact physical equality is established for the repeated seed in the recorded environment. Differences in platform, simulation libraries or floating-point behavior can change a fresh trajectory. Preserve its new identity and report those differences. One repeated seed is not a population uncertainty estimate. ## Read the results The website's summary JSON and CSV report all conditions. The complete records include the 1 ms body traces and 5 ms control frames. Compare live and frozen pose at the same model and scenario. Negative live-minus-frozen mean angular error favors live feedback. Target-switch progress is computed separately against each phase's fixed target, avoiding an artificial distance jump when the target changes. The scripted reference receives the same cue, 50 ms delay, straight-ahead deadband and gait interface. Identical generalized-position arrays are reported as identical physical paths even if their neural states or logits differ. Neither a successful learned choice nor a shared path demonstrates that learned recurrent dynamics, anatomical wiring, language knowledge or learned gait were necessary. The hypothesis about anatomical wiring is tested in the separate matched language topology study. See the included [declared protocol](CLOSED-LOOP-PROTOCOL.md) and [original learning protocol](LOCAL-LEARNING-PROTOCOL.md). The physical body, hybrid gait and contact feedback follow the official [FlyGym turning tutorial](https://neuromechfly.org/tutorials/4d_turning_controller/).