The “couldn’t answer a question” wound. Each answer is grounded in
notes/talk_reservoir.md and the papers. Numbers are current
gold (703 s). The user owns these — refine to your own
words.
Q: Is 0.104 m a fundamental floor, or just where your tuning landed? A: Both are addressed. The theoretical floor is the Coriolis-only limit \(\tilde{\boldsymbol x}_{ss}=-(\boldsymbol J_{\tilde x}^\top\breve{\boldsymbol K})^{-1}\breve{\boldsymbol C}\,\breve{\boldsymbol v}_d\) — a standing pose error set by the moving reference, not tuning. The measured operational floor is ~0.104 m. The gap between them is the open agenda (§8), not an unknown.
Q: Why is the EE error ~100× the CoM error? A: Separation of concerns. The CoM loop is decoupled (3×3, inertially free, \(m\dot{\boldsymbol v}_c=\boldsymbol f_c\)) and holds ~1 mm. The EE rides the coupled 9×9 circumcentroidal block, where the cruising arm’s Coriolis force and the CoM coupling inject the standing pose error of the \(\tilde{\boldsymbol x}_{ss}\) balance. The floor is the coupling, not the controller.
Q: Can you drive it to zero? A: Not with perfect-CoM-tracking alone — the Coriolis-only term survives even then. Driving it lower means attacking the coupling term (a forward-plan item), not retuning gains.
Q: Did decoupling the CoM improve performance? A: No — on accuracy, pointing, and singularity clearance. We tested the hypothesis and falsified it: the EE floor is invariant (median flat, p99 −4.4%), \(s_{\min,G}\) unchanged. We keep the decouple because it makes the CoM channel exact (fixes a 108× coupling defect) — a correctness win, not a performance lever. A falsified hypothesis, honestly reported, is stronger than a fragile positive.
Q: Then why keep it? A: Correctness/exactness near singularity, and it removes a confound for the risk-aware phase.
Q: How close to singular does the system get? A: It operates at the conditioning floor — \(s_{\min,G}\) median 0.0245, just below the 0.025 derate floor; the arm spends ~half its working time near singular. Four layers keep it safe (Tikhonov Γ-reg, speed derate, damped \(\boldsymbol J^\oplus\) three-tier, hold-last freeze). \(\sigma_{\mathrm{crit,1}}=v/x_{\max}=0.143\).
Q: \(s_{\min,G}\) vs \(s_{\min,J}\)? A: Near-identical — Γ is singular iff \(\boldsymbol J^\oplus_{\nu_e}\) is (Spearman ≥ 0.9969). We report \(s_{\min,G}\) (Γ conditioning). [Internal: a label swap is being cleaned up; physics unaffected.]
Q: How did you set \(\sigma_{\mathrm{crit,2}},\sigma_{\mathrm{crit,3}}\)?
A: One velocity budget threaded through all three floors (reproduces the
code’s 0.02) — the cleanest to defend. [Confirm this is the version on
the slide; see singularities/final.tex:291.]
Q: Why did the mission shrink 36%? A: Orbit-synced stop — end on achieved orbit progress, not a fixed clock — removes a dead idle tail. Coverage held at 0.9972.
Q: Is POSE mode enough, or do you need the scorer? A: POSE reaches complete coverage on simple targets. ANCHOR + the information-aware scorer are retained (inert) for risk-aware planning.
Q: Sign convention on the desired EE position? A:
\(\boldsymbol p_{ed}=\boldsymbol p_c -
r_{cam}\hat{\boldsymbol r}\) — camera placed at standoff
back along the aim direction. [Internal: reconcile against
current_sota eq (5.7), which writes +.]
Q: Are you ready for the risk-aware phase? A: Yes. Guidance is deterministic with complete coverage; control is characterised against Giordano 2019; the error floor is understood analytically and empirically and shown architectural; singularity is handled; the scorer is ready. The nominal baseline’s limits are now known and defensible — the prerequisite for introducing uncertainty.
current_sota; \(s_{\min,G}/s_{\min,J}\) label swap; \(\sigma_{\mathrm{crit}}\) variants. All four
are logged in notes/talk_reservoir.md for the M2
cleanup.