Phase 00 round-1 verdict — configuration shootout + ±2% sensitivity probe

Date: Jun 12, 2026. Protocol: tasks/phase00_protocol.md (pre-registered before any run). Chain: tasks/daily_log/Jun11_26/CHAIN_8.md (task IDs T0–T9 decoded there; T8 = the user-released round-2 scope decision referenced at the end of this report).

Legend

Runs (full inspection path — a pole-to-pole spiral around the target, 30 revolutions — simulation capped at 1800 s; matched knobs, no per-config tuning). Speed variants below are written in the fixed order nominal (0.90 m/s) / −2% (0.882) / +2% (0.918). “Arm” in this report always means the robot manipulator, never an experiment branch.

Metrics (direction and gate/stretch for each; EE = end effector, the camera-carrying hand; twist = linear + angular velocity stacked, m/s with rad/s — hence “mixed units”):

Results

Headline numbers (full console tables: generated_reports/GNC/phase00_round1_table.txt, generated_reports/GNC/phase00_round1_with_probe_table.txt; regen commands at the end):

Verdict (by the pre-registered ranking rule, applied verbatim)

Rule: 1) gates passed (pe_p99 < 0.2, coverage ≥ 0.99, completed) 2) pe_p99 3) completion 4) dwell. Fragility rule: a winner that loses its gates under ±2% speed is FRAGILE → prefer a robust runner-up.

1. WINNER: r0_6dof_path — and it is ROBUST. The only configuration passing all three gates, and it keeps all three at both ±2% speed variants (pe_p99 nominal/−2%/+2%: 0.1651 / 0.1703 / 0.1660; coverage ≥ 0.9981; completed in every run). The fragility rule does not fire.
2. Runner-up r1_7dof_path fails the pe gate at every speed (nominal/−2%/+2%: 0.2294 / 0.2804 / 0.3092) — robustly failing, not marginally: no ±2% nudge rescues it.
3. r3_7dof_anchor third by the rule; r2_7dof_reactive last (pe_p99 0.567).

Tie-break by ze_p99 was not needed.

Who wins what, stated symmetrically: r0 wins the gate discriminator (pe_p99) plus pointing (ze_p99 5–30× lower than every 7-DOF run) and command smoothness (nu_jump_p99 0.059 vs 0.064–0.134). The 7-DOF family wins coverage (1.000, the stretch target), completion (−19% at matched nominal speed, −37% at +2%), and every conditioning metric (dwell, throttled, stalled).

Provenance sanity (the fragility study’s E2 check applied early: the nominal run should reproduce the adopted headline): the r0 nominal-speed run reproduces the adopted Mission-1 numbers to three digits on the gate metric — pe_p99 0.1651 vs the adopted 0.1654, completion 1070.8 s vs 1064.5 s, coverage 0.9981 vs 0.9990. Residuals are attributed (not isolated) to the 1800 s cap vs the adopted run’s 1100 s and the Phase 00 runner vs the Jun-10 c4_chain5_verdict script; bit-for-bit determinism holds only for identical config + script, which this deliberately is not.

Mechanism story — why the ranking looks like this

r0’s conditioning picture is structurally saturated — and that is also why it is robust. Throttled ~64% of steps, dwell ~630 s, stalled ~14% — at every speed. This is the known pre-existing reach/shoulder singular factor of the 6-DOF arm (singular factor = the arm geometry, here shoulder near full reach, that drives σ₆ toward zero; reference runs measured ~32% derated, and the adopted Mission-1 config has run 0.64–0.80 throttled all along — high_level_plans/decisions.md, CHAIN_5 entry). Saturation means ±2% cannot meaningfully worsen a regime already pinned at its conditioning floor. The cost of crawling: completion ~1071 s vs the 7-DOF’s 676–869 s.

r1’s gate failure is consistent with the velocity-lag mechanism, with an honest gap. Prior art (decisions.md CHAIN_6 entry; trusted log logs/m7_on_s45_freeze.npz): 7-DOF matched-pace runs at 0.45 m/s hit pe_p99 0.140, and the v_c sweep found pe median scaling ~linearly with speed. Strict linear scaling from 0.140 at 0.45 m/s predicts ~0.28 at 0.90 m/s; the observed 0.229 is better than that, and the probe arms are non-monotone in speed (−2% 0.2804 > nominal 0.2294 < +2% 0.3092) where pure lag predicts monotone. So lag is the dominant term but not the whole p99 story — event-driven contributions (dwell/aim transients) are plausible and are handed to CHAIN_10’s peak analysis, not explained away here. (r0 shows the same non-monotone ordering at only ~3% amplitude, 0.1651–0.1703; no replicate-pair noise band exists for these schedule-confounded runs, so that swing is noted as unquantified rather than excused. r1’s 35% swing is far outside any plausible band.)

The 7-DOF dwell cliff — the sharpest schedule-sensitivity signature yet recorded, and it is NOT a clean monotone story. r1’s dwell at nominal/−2%/+2%: 431 s / 231 s / 3.27 s. A +2% timetable nudge all but eliminates near-singular dwell (throttled 0.6%, stalled exactly 0.0000, ze_p99 3× better at 0.2815) — consistent in direction with CHAIN_6 P2 (slower = worse). But the −2% variant (231 s) is better than nominal (431 s), which CHAIN_6’s direction does not predict: dwell is non-monotone in speed at this operating point, peaking at nominal. Both facts go to CHAIN_10 as open questions (tasks/daily_log/Jun12_26/chain10_kappa_forensics_design.md). The 6-DOF, by contrast, is schedule-INSENSITIVE (dwell 614–655 s across variants) because it sits on a saturated floor. Schedule sensitivity is therefore a property of the 7-DOF’s marginal regime — the timetable decides how closely commanded reach approaches the singular manifold (the set of arm poses where the arm locks up), and the contrast pair r1 nominal (431 s) vs r1 +2% (3.27 s) — differing in ONE knob by 2% — is CHAIN_10’s primary specimen.

Round-2 lever assignments (winner = 6-DOF ⇒ the lever list narrows)

Pre-registered levers (protocol): L1 velocity-lag feedforward, L2 σ-gradient null-space pull (σ-gradient pull = steer the spare 7th-joint motion uphill in σ₆), L3 pointing smoothing retune, L4 kernel-freeze thaw policy.

• Applicable to the r0 winner: L1 and L3 only. L2 and L4 act on the null space, and the 6-DOF arm is not redundant — it has no always-available task-free direction (its null space is zero-dimensional except exactly at singular poses). L1 targets pe_p99 0.165 → < 0.1 (the stretch); L3 targets ze_p99 0.055 → toward 0.001 (a long way; the floor may be architectural).
• Decision point for the user (protocol task T8, the round-2 scope release). The protocol scopes round 2 to the winner, and the winner is the slow-but-gated 6-DOF. The trade is genuine, not one-sided: the 7-DOF family is far ahead on coverage, completion, and conditioning, but behind on pointing and smoothness, and its single failing gate is precisely L1’s target. Two defensible paths: (a) per protocol — harden r0 with L1+L3; (b) a recorded protocol amendment (logged before any round-2 run) — apply L1 to BOTH r0 and r1 and let the gate decide. Note the bar honestly: an L1-hardened r1 wins outright only if its pe_p99 drops below r0’s 0.1651 m (rule criterion 2), not merely below the 0.2 m gate; at, say, 0.18 m it is gated yet still ranks second. Path (b) deviates from the protocol’s “winner only” scope and therefore requires an explicit, recorded user decision before any run.

Regen commands (provenance, copy-paste complete)

/Users/antoniahoffman/miniforge3/envs/new-pin-env/bin/python validation/run_mission.py run --robot UR3   --orbit pose   --suppress off --cap 1800 --npz logs/phase00/r0_6dof_path.npz
/Users/antoniahoffman/miniforge3/envs/new-pin-env/bin/python validation/run_mission.py run --robot UR3   --orbit pose   --suppress off --cap 1800 --speed 0.882 --npz logs/phase00/r0_6dof_minus2pct.npz
/Users/antoniahoffman/miniforge3/envs/new-pin-env/bin/python validation/run_mission.py run --robot UR3   --orbit pose   --suppress off --cap 1800 --speed 0.918 --npz logs/phase00/r0_6dof_plus2pct.npz
/Users/antoniahoffman/miniforge3/envs/new-pin-env/bin/python validation/run_mission.py run --robot UR3_7 --orbit pose   --suppress on  --cap 1800 --npz logs/phase00/r1_7dof_path.npz
/Users/antoniahoffman/miniforge3/envs/new-pin-env/bin/python validation/run_mission.py run --robot UR3_7 --orbit pose   --suppress on  --cap 1800 --speed 0.882 --npz logs/phase00/r1_7dof_minus2pct.npz
/Users/antoniahoffman/miniforge3/envs/new-pin-env/bin/python validation/run_mission.py run --robot UR3_7 --orbit pose   --suppress on  --cap 1800 --speed 0.918 --npz logs/phase00/r1_7dof_plus2pct.npz
/Users/antoniahoffman/miniforge3/envs/new-pin-env/bin/python validation/run_mission.py run --robot UR3_7 --orbit off    --suppress on  --cap 1800 --npz logs/phase00/r2_7dof_reactive.npz
/Users/antoniahoffman/miniforge3/envs/new-pin-env/bin/python validation/run_mission.py run --robot UR3_7 --orbit anchor --suppress on  --cap 1800 --npz logs/phase00/r3_7dof_anchor.npz

/Users/antoniahoffman/miniforge3/envs/new-pin-env/bin/python validation/run_scoreboard.py logs/phase00/r0_6dof_path.npz logs/phase00/r1_7dof_path.npz logs/phase00/r2_7dof_reactive.npz logs/phase00/r3_7dof_anchor.npz > generated_reports/GNC/phase00_round1_table.txt
/Users/antoniahoffman/miniforge3/envs/new-pin-env/bin/python validation/run_scoreboard.py logs/phase00/r0_6dof_path.npz logs/phase00/r0_6dof_minus2pct.npz logs/phase00/r0_6dof_plus2pct.npz logs/phase00/r1_7dof_path.npz logs/phase00/r1_7dof_minus2pct.npz logs/phase00/r1_7dof_plus2pct.npz logs/phase00/r2_7dof_reactive.npz logs/phase00/r3_7dof_anchor.npz > generated_reports/GNC/phase00_round1_with_probe_table.txt

npz files (~25–27 MB each) are NOT committed — provenance via the commands above (deterministic: same config + seed + script reproduces bit-for-bit).