Written blind (without the INSIGHTS.md files) on purpose: this is my current intent from the Jun19 campaign evidence alone, to be reconciled against the collected insights before any sim runs.
At the adopted operating point (rs=0.40, corrected
base, σ-onset 0.025), the impedance derate is inert —
retime_only (C/I), coupling (G), and the noise
variants (F) are all ≡ incumbent. The Jun18 “derate tax” was an artifact
of that study’s wide derived-onset band (σ<0.143). The real
problem is not the derate — it is the near-singular tracking
corruption: the EE tail error (pe_p99) and the fraction of
steps near-singular (fbelow). Two levers move it:
| lever | pe_p99 | cov | fbelow | profile |
|---|---|---|---|---|
| incumbent | 0.1753 | 0.995 | 0.597 | — |
| E analytic feedforward | 0.1590 | 0.998 | 0.280 | all-rounder, no coverage cost |
| D/K Tikhonov onset 0.05→0.02 | 0.1597 | 0.988 | 0.027 | conditioning specialist, −0.8% cov |
D and K are equivalent (onset value robust). K is not a
crawl (finishes the helix faster, not slower —
k_crawl_diagnostic.py).
E + D/K combined (new run) —
analytic_ff=true AND
gamma_regularization.floor=0.020. Hypothesis: E holds
coverage (feedforward) while D/K drives fbelow down (de-corrupted
velocity solve) — they attack the same corruption by different
mechanisms, so they may stack. Gate to beat both:
pe_p99 ≤ 0.159 AND coverage ≥ 0.99 AND fbelow < 0.10 AND s_min_J min
≥ 0.019.
D/K coverage-gap diagnosis (existing logs first) — why does D/K dip to 0.988 while completing 100% traversal? Hypothesis: the changed solve shifts the camera FOV path → a few missed triangles (benign), not a real hole. Diff coverage(t) and the missed region vs incumbent. Decides whether the 0.8% disqualifies D/K or is cosmetic.
H on the right metric (existing logs) — re-judge base_damping_3.5/5.0 on z_b back_half_pkpk (the slow ~12 s wobble the gate named), not rms. If ptp is also worse → H is genuinely dead; if ptp improves while rms worsens → the wobble was damped and rms saw added lag (different verdict).
rs035 re-run with the footgun fix (new run) — radius_scale=0.35 was inert before (ran as rs040); now valid (db927d6). Get a real small-standoff point. Gate: coverage ≥ 0.99, compare pe_p99 / fbelow / s_min_J to rs040.
Understand the onset→floor shift (analysis) — incumbent here is smin 0.0193 / fbelow 0.60 vs Jun18’s 0.0268 / 0. Confirm this is purely the onset (0.025 vs 0.143) and decide whether the operating onset itself is a lever worth sweeping (σ_high), distinct from the Tikhonov onset (D/K).
Objective = smoother COM / Base / EE tracking without losing coverage or the floor. Likely outcome: adopt E (clean win, no coverage cost) as the headline; add D/K only if (2) shows the coverage gap is cosmetic and the combo (1) doesn’t already subsume it. H, C/I, G, F: record as null/negative. Nothing adopted/re-pinned without a full-helix A/B and your sign-off.
dispatch.py status --manifest against a live
run (it collides on the shared journal).signal_measurement facade;
progress is a raw signal (OK to read directly).Reconcile this plan against the central INSIGHTS.md files — confirm none of (1)–(5) re-treads a falsified idea or a known footgun — then run.