legacy_questions

Open questions (Legacy - information only)

Updated Jun 11 26 — pre-risk ledger: outstanding issues with the existing 7-DOF algorithm, compiled from review_7dof.md, dynamics_modifications_7dof.md, mission_1_success.md, CHAIN_5, and PIN_IT_FOR_LATER. One line of what/evidence/lever each; detailed records stay in the cited files (and below for the dwell study).

A. Measured residuals (performance, in priority order)

• R1 kernel-freeze A/B — logs exist, analysis unreported. → CHAIN_6 (Jun 11, tasks/daily_log/Jun11_26/CHAIN_6.md). Provenance corrected by npz inspection: m7_on_s45.npz is 1800 s pre-freeze code (confounded baseline); the 1600 s arm is m7_on_s45_rerun.npz; freeze.npz/rerun1800.npz are nominal freeze-ON replicates that nonetheless DIVERGE mid-mission (sota_p_e overlay) — anomaly must be resolved before any A/B (CHAIN_6 T1 freeze-fingerprint test). Clean A/B needs one new --freeze_floor 0.0 1800 s run unless T1 shows rerun1800 was secretly freeze-OFF. Check: kernel angle ∠(n̂ₖ, n̂ₖ₋₁) + thaw jumps through derated windows, p_e/ν_e tails, gates. RESOLVED Jun 11 (CHAIN_6 T4, validation/r1_freeze_ab_verdict.py + tasks/daily_log/Jun11_26/images/CHAIN_6_0.png): (i) PASS, (ii) FAIL as measured (thaw p99 82.9° vs 10°; but total kernel transport 1924° vs 4222°), (iii) PASS — all gates improve (pe_p99 0.140 vs 0.172, t_complete −119.9 s, dwell 10.5× down).
• Schedule sensitivity / dwell feedback — study COMPLETE Jun 11 (CHAIN_6 T5 + T7, both predictions tested); STANDS as a first-class risk source with no working cheap mitigation. Prediction 2 (T5, validation/p2_schedule_sensitivity.py, images/CHAIN_6_1.png): derate 0.0238 → 0.3299 under a −2% speed perturbation at matched 1800 s, freeze ON (Δ_speed 0.306 > Δ_grid 0.175); rule reads NOT CONFIRMED only on the κ-p50 WHILE-gate (bulk moved 42% — stronger than tail-only, not absent); headline KEPT (derate fraction), report WITH schedule context. Prediction 3 (T7, validation/p3_aim_clock_verdict.py, images/CHAIN_6_2.png): FALSIFIED at base speed — the monotone aim clock (0.045/0.05) REOPENS the amplifier (dwell 32.0 → 615.3 s, pe_p99 0.2032 breaks M2, t_complete +271.8 s); the lead cap re-freezes the fixed point 0.05 m ahead (1.1 s authority); knob default-OFF; the −2% knob retest is moot for this parameterization. Carries into the risk phase as a standing sensitivity (CVaR pin below); reopen only with a redesigned engagement gate (derate-conditioned, startup-exempt, authority ≫ episode duration). Full record: section below.
• Reach/shoulder conditioning bottleneck — the structural residual. Derate 0.64–0.80 on 6-DOF, still 0.23–0.34 on 7-DOF at matched pace; NOT tunable (reach corridor falsified Jun 8). Tracking passes gates anyway, but robustness margin is thin and CVaR will see it as a standing risk. Levers: operational weak full-arm posture (extend beyond INITIAL), null-space posture pull toward larger σ₆. Evidence: mission_1_success.md residual; review_7dof.md R3.
• Velocity-lag ceiling — median p_e ≈ v_c·τ (first-order tracking, no FF); p99 0.165–0.172 passes <0.2 but the gap to <0.15 lives in derated windows where lag accumulates. Lever: desired-twist feedforward (integral stays OFF by design). Evidence: mission_1_success.md lever list; operational-regime memory.
• Base wobble 12 s mode (low) — underdamped base–arm eigenmode, p95 ~18–19°; aiming is wobble-NEUTRAL (K3’s 4× claim was confounded by the frozen arm) and wobble does NOT drive p_e (r = −0.157). Margin question only; lever if ever needed: active damping / soft base-posture regularization.

B. Undesigned mechanisms (masked by current workarounds)

• Pan winding / unwind allocation — shoulder_pan winds ~3 rev over the helix; exempt from the envelope (clipping is the wrong mechanism for winding joints) but real UR3 is ±2π. Undesigned: unwind trigger vs base-yaw allocation vs trajectory redesign. Evidence: CHAIN_5; mission_1_success.md open item.
• Envelope provenance (USER DECISION) — the original XML box_limited ranges are task artifacts, not UR3 spec; the engineered envelope is data-derived (healthy-regime support + sinθ₃ fences, elbow cut 19% of excursions). Decide and document what the envelope IS (mission spec? datasheet-derived? engineered?) — it’s the committee-facing soft spot. Evidence: CHAIN_5 C3a provenance question.
• 7-DOF envelope re-derivation (data-driven, as before) — the fences were derived on the 6-joint posture family; under 7-DOF the elbow now rides its fence 20–30% of steps (vs 0.000 at adoption). Re-derive statistically on 7-joint data (same recipe: healthy-regime support + sinθ₃ fences). DECIDED Jun 11: stay data-derived for now; the analytical derivation moved to §C (theory pile). Evidence: review_7dof.md R3 pre-claim requirement; §6 “re-dimensioned not redesigned”.
• wrist_2 (sinθ₅) fence unused — logged data never approaches it; cut formally or document as decorative. (Decision, 5 min.)

C. Theory gaps (thesis-defense obligations, all flagged in review_7dof.md)

• P1 — rebuild M̆/C̆ with the Γ_a⁻¹ congruence. Today velocity reconstruction uses the M-orthogonal system but M̆/C̆ keep the Euclidean right inverse — two bases feed one control law. R4 sized the discrepancy: 1.9% median healthy, ~100% in derated tails (comparator itself degrades there). Promoted from bookkeeping to measured-payoff item. Script: validation/r4_mbreve_inconsistency.py.
• P3 — cascade stability on the augmented system. Lemma 1 gives passivity of the augmentation but does not bound v_n ↔ task cross-Coriolis coupling; the closed-loop cascade argument is deferred.
• P4 — ISS proof for null-space damping. Prop 4 is a sketch (Coriolis bounded only on compact velocity sets); the “high-damping limit” claim must not migrate
  to summaries without P4 hypotheses. Empirics are strong (|v_n| 0.234 → 0.0058);
  the proof isn’t. Tikhonov-type treatment named as the route.
• Analytical joint-envelope derivation (paper-and-pencil; COMMITTEE SOFT SPOT). The envelope is statistically derived (healthy-regime support + sinθ₃ fences) — defensible as engineering, but the committee-facing weakness is that it isn’t derived from the kinematics. Route: start from det(J) = R(θ)·sinθ₃·sinθ₅ and
  derive fences as σ-margin level sets per singular family (elbow |sinθ₃| ≥ δ
  directly; reach/shoulder factor R(θ) is the hard part — same factor behind the
  derate residual). DECIDED Jun 11: data-derived stands in the meantime; no
  schedule pressure (“I’ve got time”). Owner: user (theory track).

D. Decisions pending / deferred housekeeping (not blocking risk)

• Mesh prototype wire-in (sandbox/mesh_oo, validated, decision pending)
• knob consolidation (5 booleans → one mode enum), state-machine doc refresh (HOLD/FALLBACK unreachable under Mission-1 defaults), capsule avoidance vs clipping, daily_log dedup — see tasks/PIN_IT_FOR_LATER.md.

---

Schedule-sensitivity / dwell-feedback study (detailed record)

⚠️ Provenance caveat (Jun 11, CHAIN_6 T0): WHICH log pair produced the numbers below is unrecorded, and the candidate pairs straddle the freeze boundary (m7_on_s45 = pre-freeze 1800 s; rerun = 1600 s post-freeze). CHAIN_6 T3 re-derives the episodes with a scripted extractor and records the reproducing pair.

• (was “resonance” — RETRACTED after measurement, Jun 11). Data (cap-1800 vs cap-1600 s45 npz, same speed): median and p75 of σ₆ identical across grids; ONLY the lower tail differs (p5 0.0087 vs 0.0206, p25 0.027 vs 0.045) and p25 sits at the derate floor → “derate fraction” is a tail statistic with built-in amplification. Progress clock is continuous (no aliasing). Working hypothesis: dwell feedback — aim pose grazes the singular factor → σ₆ dips → derate slows vehicle → progress-synced aim WAITS → arm dwells at the offending pose (lock-in’s gentler cousin, via the aim clock). Evidence: 2.3× more zero-advance stalls on the fine grid (7% vs 3%). Predictions to test (cheap, npz in hand):
  1. dips cluster at specific orbit phases + long dwell episodes on the fine grid; MEASURED Jun 11: two-stage amplifier. Episode durations comparable across grids (mean 34 vs 40 s; max 133 s — huge vs all loop timescales → dwell is real on BOTH grids); the grid changes ENTRY COUNT (10 vs 2 episodes, 336 vs 79 s total). Clustering: deciles 1–4 + four episodes in decile 10 (pole approach) on fine; only decile 2 on coarse. → entry is schedule-hypersensitive, residence is dwell-locked (~30 s);
  2. ANY small schedule perturbation (sample phase, ±2% speed) swings derate comparably → if so, replace derate-fraction headlines with σ₆ p5/p25; MEASURED Jun 11 (CHAIN_6 T5, validation/p2_schedule_sensitivity.py, images/CHAIN_6_1.png): swing CONFIRMED and asymmetric — derate 0.0238 (base) / 0.0288 (+2%) / 0.3299 (−2%) at matched 1800 s, freeze ON; Δ_speed 0.3062 vs Δ_grid 0.1747 (recomputed). The pre-registered rule nonetheless reads NOT CONFIRMED, solely on the WHILE-gate: κ p50 moved 42% (gate <2%) — the −2% arm shifts the BULK conditioning (κ p50 33.9→50.4), so the mechanism is STRONGER than tail-only amplification. HEADLINE KEPT (negative recorded): κ p95/p75 rel spreads 1.369/1.259 vs the 1.201 switch threshold — κ is not the more stable statistic; σ₆ p5 moves too (0.0368→0.0116). Report derate WITH schedule context. Regime split (mandatory, T3 implication): freeze-OFF grid amplifier wide open (336/79 s dwell); freeze-ON dwell 32 s at base speed but REOPENED to 517 s (max episode 303 s, stalls 8.8%) on the −2% arm — the freeze does not close the amplifier against speed perturbations; slower = worse (vehicle slows → progress-synced aim waits → dwell), exactly the dwell-feedback sign. Gates still pass on the −2% arm (pe_p99 0.187 < 0.2, coverage 1.0, t_complete 1657.7 s);
  3. breaking the dwell (min progress rate during derate, or small aim lead) de-amplifies. MEASURED Jun 11 (CHAIN_6 T7, validation/p3_aim_clock_verdict.py, images/CHAIN_6_2.png): FALSIFIED at base speed — the monotone aim clock (floor_rate 0.045 m/s + lead cap 0.05 m, default OFF, CHAIN_6 T6) REOPENS the amplifier instead of closing it: dwell 32.0 → 615.3 s (4 → 5 episodes, max 463.3 s), stalls 0.84 → 10.31%, derate 0.0238 → 0.3878, pe_p99 0.2032 (+0.063 vs band 0.029 — breaks M2 <0.2), t_complete +271.8 s (band 106.7). Engagement proven, not assumed: floor active 7.0% of steps with the lead pinned at the cap; divergence fingerprint MATCH (engaged step 2, q diverges step 3). Mechanism: the lead cap re-freezes the dwell fixed point 0.05 m ahead after only cap/floor_rate = 1.1 s of authority (9–463 s episodes cannot be broken by construction), while the saturated lead shifts the posture onto the −2%-like conditioning branch (κ p50 33.9 → 50.4) and the dwell feedback amplifies it. The −2% stress retest is moot for this parameterization; a retry would need a derate-conditioned, startup-exempt gate with authority ≫ episode duration. Knob stays default-OFF.

Pinned for risk (don’t delete)

• CVaR relevance: if confirmed, schedule sensitivity is a first-class risk source for the risk-aware phase, not noise. Comparability rule (hard): mission arms only comparable at matched simulation.time.duration (duration → orbit sample count → spacing → aim schedule).