CHAIN_13 D2 — attribution audit: which mitigation layer owns the in-window error?

Pre-registered in tasks/chain13_criteria.md (D2); band/edge definitions and the sj-stands-in-for-s_min_G note are in the script header (validation/chain13_attribution.py). All masses below are INSIDE the κ-peak windows (committed CHAIN_10 extractor, unchanged). ‘steps’ = share of in-window steps spent in that band; ‘ze/pe’ = share of in-window error mass accrued there. A band owning much more mass than time is a suspect.

r0_6dof_path (windows cover 0.0% of the run; ze mass in windows 0.0% of total, pe 0.0%)

edges: tik<0.0490, derate/freeze<0.025, J+damp<0.02, hold<0.005; thaw events 0, hold-last crossings 11891, switch-proximate ze mass 0.0% of in-window

r0_6dof_minus2pct (windows cover 0.0% of the run; ze mass in windows 0.0% of total, pe 0.0%)

edges: tik<0.0490, derate/freeze<0.025, J+damp<0.02, hold<0.005; thaw events 0, hold-last crossings 349, switch-proximate ze mass 0.0% of in-window

r0_6dof_plus2pct (windows cover 0.0% of the run; ze mass in windows 0.0% of total, pe 0.0%)

edges: tik<0.0490, derate/freeze<0.025, J+damp<0.02, hold<0.005; thaw events 0, hold-last crossings 4338, switch-proximate ze mass 0.0% of in-window

r1_7dof_path (windows cover 0.6% of the run; ze mass in windows 1.6% of total, pe 0.7%)

edges: tik<0.0490, derate/freeze<0.025, J+damp<0.02, hold<0.005; thaw events 47, hold-last crossings 30, switch-proximate ze mass 7.4% of in-window

r1_7dof_minus2pct (windows cover 0.5% of the run; ze mass in windows 1.2% of total, pe 0.4%)

edges: tik<0.0490, derate/freeze<0.025, J+damp<0.02, hold<0.005; thaw events 37, hold-last crossings 6, switch-proximate ze mass 19.8% of in-window

r1_7dof_plus2pct (windows cover 0.6% of the run; ze mass in windows 3.1% of total, pe 0.6%)

edges: tik<0.0490, derate/freeze<0.025, J+damp<0.02, hold<0.005; thaw events 8, hold-last crossings 16, switch-proximate ze mass 21.9% of in-window

r2_7dof_reactive (windows cover 1.2% of the run; ze mass in windows 1.4% of total, pe 2.2%)

edges: tik<0.0490, derate/freeze<0.025, J+damp<0.02, hold<0.005; thaw events 62, hold-last crossings 38, switch-proximate ze mass 13.5% of in-window

r3_7dof_anchor (windows cover 0.5% of the run; ze mass in windows 0.9% of total, pe 0.6%)

edges: tik<0.0490, derate/freeze<0.025, J+damp<0.02, hold<0.005; thaw events 59, hold-last crossings 26, switch-proximate ze mass 41.3% of in-window

Lag-vs-event split (absorbs the old C1)

• r0 (6-DOF): median-pe fit pe ≈ v·τ with τ = 0.0795 s (worst relative residual 5.2%); medians {‘r0_6dof_path’: 0.0699, ‘r0_6dof_minus2pct’: 0.0739, ‘r0_6dof_plus2pct’: 0.0708}
• r1 (7-DOF): median-pe fit pe ≈ v·τ with τ = 0.1644 s (worst relative residual 35.9%); medians {‘r1_7dof_path’: 0.1089, ‘r1_7dof_minus2pct’: 0.155, ‘r1_7dof_plus2pct’: 0.1795}

Event term: the per-run ‘ze/pe mass in windows’ percentages above. Nuance: the windows hold a small share of TOTAL error mass yet CHAIN_10 measured the error ELEVATED there (Cliff’s δ on window means) — both are true. Events set the p99 gate metrics; cruise accumulates the bulk. L1’s A/B must therefore report event-window error alongside the cruise lag, as AMENDMENT 1 + criteria require.

Observations outside the registered rules (recorded, not adjudicated)

• Hard-floor chattering (r0): the 6-DOF crosses the hold-last floor (σ=0.005) path: 11891×, minus2pct: 349×, plus2pct: 4338× — the threshold sits INSIDE r0’s operating distribution (Holt & Desrochers’ chattering regime). Feeds D4’s threshold-spread discussion and lever R’s case.
• The marginal bands are empty at the peaks: B1 (Γ-Tikhonov-only) and B2 (derate ramp) carry ~0% of in-window mass on every run — the worst windows sit in B3/B4. The study’s layer-ordering flag (§2) is exonerated within κ windows; any cost of the marginal layers accrues in cruise, not at the peaks.
• L4-rule instrument caveat: thaw events sit at window EXITS by construction (σ rising through the freeze floor ends a κ peak), so the in-window ∩ ±5-step mask structurally under-counts thaw-driven mass. The rule fired NO-GO as registered; a v2 rule (windows extended past thaw re-entry) would need new pre-registration — checkpoint decision.

Pre-registered implication rules

• L4 (switch smoothing) — switch-proximate ze ≥ 50% of in-window ze on ≥ 3/5 7-DOF runs: does NOT fire → L4 NO-GO (0/5; r1_7dof_path 7.4%, r1_7dof_minus2pct 19.8%, r1_7dof_plus2pct 21.9%, r2_7dof_reactive 13.5%, r3_7dof_anchor 41.3%)
• Damping implicated (directional-filtering follow-up, recorded not built) — B1 ze mass ≥ 2× its step share on ≥ 5/8 runs: does NOT fire (0/8; r0_6dof_path n, r0_6dof_minus2pct n, r0_6dof_plus2pct n, r1_7dof_path n, r1_7dof_minus2pct n, r1_7dof_plus2pct n, r2_7dof_reactive n, r3_7dof_anchor n)
• L3 (r0 pointing retune) — r0_6dof_path in-window ze mass < 50% of total: PASSES → L3 GO (0.0% in windows)

regen: /Users/antoniahoffman/miniforge3/envs/new-pin-env/bin/python validation/chain13_attribution.py