Scope you set: shovel-ready levers now (autonomous); 7-DOF posture lever + thesis loose ends in the morning with you. Shovel-ready bucket = the Block-2 levers minus R (already adopted rs 0.40 earlier tonight). Nothing was adopted — every re-pin is your call. All runs local on the Mac (fleet left idle: it still has the old code/rs 0.55, and syncing at 3am risked a blocking prompt).
velocity_ff) is now inert (was catastrophic at rs
0.55). Genuine un-confounded cruise lag remains; L1 v2 is a clean
feedforward derivation — a morning math discussion, not a fishing
trip.velocity_ff stays default-OFF. Both knobs
remain in the code, byte-identical OFF.| stage | result |
|---|---|
| time_scale probe (200 s) | WEAK — ze_p99 −1.7%, pe worse +1.7%. Broadband ze is NOT a low-pass-lag artifact. |
| slew-cap diagnostic | cap binds 5.7% of steps (slew p99 = the cap exactly) → ze tail is
rate-limited; rate_max knob triggered. |
| rate_max probe (200 s) | rate_max 1.0 Pareto-dominant: ze −41%, nu_jump −28%, pe −26%. Looked like a clear win. |
| full-helix A/B | INVERTED. Tracking still −20-29% AND robust ±2%, BUT derate 0.637→0.971, t_complete +19%, sj_med (conditioning) 0.0229→0.0187 (worse). |
Mechanism (the keeper finding): the aim-axis slew cap was doing double duty — smoothness and implicit singularity protection. Capping how fast the camera could re-aim kept the arm out of extension. Loosen it → the arm chases aggressive re-aims into near-singular postures → derate throttles → vehicle crawls → tracking looks better because it’s slow. The tracking gain is commensurate with the +19% slowdown (crawl confound); the sj_med drop is the independent tell that conditioning genuinely worsened. This is Chiaverini’s “hard limits make their own transients” (study §3.4) on the output-smoothing side — and it says: if better pointing is wanted, the lever is controller pointing authority, not output smoothing.
Honest provenance: by the letter of the pre-registered L3 bars (ze↓, pe≤0.1586, cov≥0.99, completed, nu_jump≤OFF) rate_max 1.0 passes — I guarded tracking but not derate/completion. That’s an incomplete-criteria lesson, recorded, NOT back-fitted to reject. My recommendation is NO-ADOPT on the chain’s own terms (conditioning); your call. One adversarial check if you want it: a matched-completion- time comparison would nail the crawl confound beyond the sj_med evidence.
Cruise (t 90–200) RHS term decomposition (EE-row norm medians), rs 0.55 → rs 0.40:
| regime | C | D | K (stiffness) | accel-FF | pe_med | sj_med |
|---|---|---|---|---|---|---|
| rs 0.55 (deep-derate) | 0.030 | 0.004 | 0.051 | 0.221 | 0.075 | 0.0063 |
| rs 0.40 (post-R) | 0.073 | 0.051 | 0.299 | 0.136 | 0.148 | 0.0347 |
R lifted the cruise above the derate floor (sj_med 0.0063→0.0347), so the term balance flipped: accel-FF was dominant only because deep-derate scaled the gains ×0.25; at full gains, stiffness K dominates. Consequences:
velocity_ff OFF vs ON is now
nearly identical (pe_med 0.1475 vs 0.1537, progress 1921 vs 1920, derate
0 both) — v1’s rs-0.55 catastrophe (“re-energize past the derate”) is
gone because there’s no cruise derate to escape. v1 is
inert (not harmful, not helpful).48522d7 — L3 knobs (--axis_ts,
--axis_rate, LOAD-time, byte-identical OFF,
canary-verified) + dispatcher validation/l3_sweep.py +
pre-registered tasks/chain13_l3_criteria.md + probe
report.logs/chain13/l3_*.npz,
l1_refresh_*.npz (probes, not pinned).