Polar — final delivery note, 2 November 2026

Polar runs Qwen3.5-4B locally with 4-bit weights. It measures the actual 2,560-dimensional residual stream and displays an approximate 3D projection. The exact geometric coordinate change is not a claim of dimensionality reduction or a new computational law.

Delivered: streamed text and measured trajectories; up to 1,024 answer tokens; up to four runs sharing one fixed projection; prompt and neuron-intervention comparisons; full answer cards; answer-token and token-through-layers views; play/pause/restart/token scrubbing; a common fitted camera; faded older segments. Neuron controls support interpretability-direction addition, activation scaling, and activation replacement.

Playback reuses retained measured events. Live-session text replay uses exact tokenizer-decoded prefixes; older saved captures without prefix history replay the path while leaving their original answer visible. Fitted view changes only the common camera, not the captured coordinates. Low 3D retained energy or a flat-looking path does not establish high-dimensional representational collapse. Joining final-layer states across tokens is a visualization, not a closed autonomous dynamical system.

Limits: all projected paths and full-dimensional scalar metrics are retained per live run, while downloadable full vectors cover the latest 64 positions. Four live runs survive page refresh but not a backend restart; starting a fifth replaces the oldest. The live quantized runtime is not an exact reproduction of the interpretability's saved BF16 experiments. The separate polar-forward equivalence check belongs to the saved calibration capture and is not rerun on every live prompt.

Validation: the preceding streaming release passed 53 Python tests and actual local 256-, 16- and 9-token browser checks. The replay release was checked on four retained user runs: play/pause/resume, exact prefix text, token cursor, scrub, completion, restart, full view, layer playback, fitted camera/reset, and mobile overflow. No new model inference, backend restart or POST request was used for replay QA.

The mathematical foundations, data contract, capture notes, delivered README and research plan are included. A research plan describes proposed work, not completed findings.