An eight-week experimental programme (CUHK SURP 2026, 22 June – 14 August) on a 14-notch tendon-driven continuum surgical robot, presented at the SURP Poster Presentation Session on 13 August 2026. The platform was rebuilt and instrumented from CAD — micron-level stereo tip tracking, cable-tension sensing, 200 Hz logging — and its dominant error source characterized: path-dependent hysteresis of up to ~0.9 mm, against a ~30 µm repeatability floor.
A data-driven compensator (a 498-point nonparametric monotone map) was validated under a preregistered, frozen-prediction protocol: frozen 18 days before a sealed 28-target benchmark, it cut tip error three-fold (234 → 78 µm); an iterative-learning layer on top converged to ~15 µm — the machine's own repeatability limit. Side findings: hysteresis loops widen under tip load (opposite to the friction-only prediction, replicated across days), and the robot's orientation channel doubles as a weighing scale with ~1 g estimated resolution.
An extended write-up is available on request.