fragility_criteria

Phase 0 fragility study — pre-registered criteria (written before any run)

Design

• System: unmodified nominal, Mission-1 adopted config (orbit_path ON), full 1100 s helix.
• Sources injected SEPARATELY (one source per cell), magnitudes per source below, seeds 0-4 (N=5).
• Nominal reference cell: uncertainty disabled, seeds vacuous (deterministic) -> single run.

Cells

source	knobs	lo	mid	hi
camera	pos_std (m) / angle_std (deg)	0.01 / 0.5	0.03 / 2.0	0.10 / 5.0
state	pos_std (m) / vel_std (m/s)	0.005/0.005	0.02 / 0.02	0.05 / 0.05
actuation	force_std (N)	0.5	2.0	5.0

Metrics (per run; aggregated mean +/- 95% CI across seeds)

• area_coverage_frac (final) - the realized coverage under perturbed marking
• p_e error p99 (m) - ||p_e_actual - p_e_desired|| 99th percentile
• cumulative |f_c| - control effort
• completion: progress at t_end (or time of orbit completion if reached)

Pre-registered expectations (the gate)

• E1: each metric degrades monotonically with magnitude within a source (non-monotone => injection or metric is wrong; STOP and investigate).
• E2: nominal cell reproduces the adopted headline (coverage 0.999, pe_p99 0.165). NOTE (pre-registered Jun 11, before the campaign): Phase 00’s verdict re-pins the frozen baseline (phase00.md T9); E2’s numbers and the Design config line update to the winner BEFORE any fragility run — the campaign measures the post-verdict baseline.
• E3 (scope): state noise = COM localization only; EE estimated-state and arm actuation noise are Phase B/C scope and get fragility columns then.

Outputs

• generated_reports/analysis/fragility_table.md (the yardstick table)
• saved NPZ log per run under logs/fragility/ (provenance: regen via validation/fragility_study.py)