"""Frozen same-roster component statistics and pre-specified candidate decisions.""" import math from phase3_seed1_postfit_statistics import comparisons as seed_comparisons from r4bc_common import paired_difference def comparisons(rows, candidate): # This is the same estimator and bootstrap as the independently reviewed seed result. out = seed_comparisons([dict(row, seed1=row[candidate]) for row in rows]) for m in out.values(): m['means'][candidate] = m['means'].pop('seed1') m['components'][candidate] = m['components'].pop('seed1') m['part_counts'][candidate] = m['part_counts'].pop('seed1') m['undefined_parts'][candidate] = m['undefined_parts'].pop('seed1') m['candidate_minus_seed0'] = m.pop('seed1_minus_seed0') return out def decision(item, metrics, seed_metrics): primary = 'last_location' if item == 'c' else 'relative_l2' delta = metrics[primary]['candidate_minus_seed0'] field = metrics['relative_l2']['candidate_minus_seed0'] if delta['components'] != 204 or field['components'] != 204: return dict(status='INCOMPLETE_PRIMARY_PAIR_SUPPORT', primary=primary) if seed_metrics is None or seed_metrics[primary]['seed1_minus_seed0']['components'] != 204: return dict(status='WAIT_SEED_REFERENCE', primary=primary) noise = seed_metrics[primary]['seed1_minus_seed0'] if noise['component_ids'] != delta['component_ids']: raise ValueError('Seed and candidate contrasts must have identical components') magnitude_pass = abs(delta['value']) > abs(noise['value']) if item == 'c': gate = delta['value'] <= -.010 and delta['high'] < 0 and field['high'] < .001 else: gate = field['high'] < -.001 return dict(status='PASSES_OLD_CORPUS_SCREEN' if gate and magnitude_pass else 'DOES_NOT_PASS_FIXED_SCREEN', primary=primary, registered_metric_gate=gate, exceeds_same_endpoint_seed_difference=magnitude_pass, signed_candidate_delta=delta['value'], signed_seed1_minus_seed0=noise['value'], independent_fresh_family_confirmation=False) def tail_comparisons(rows, item, include_seed): groups = {} for row in rows: groups.setdefault(row['component_id'], []).append(row) arms = [item, 'seed0'] + (['seed1'] if include_seed else []) values = {} for key in ('E_tail', 'Miss', 'False_alarm'): by_arm = {arm:{g:math.fsum(r['tail']['arms'][arm]['overlap'][key] for r in parts) / len(parts) for g, parts in groups.items()} for arm in arms} values[key] = dict( means={arm:math.fsum(v.values())/len(v) for arm,v in by_arm.items()}, candidate_minus_seed0=paired_difference(by_arm[item], by_arm['seed0']), seed1_minus_seed0=paired_difference(by_arm['seed1'], by_arm['seed0']) if include_seed else None, ) return dict(metrics=values, parts=len(rows), components=len(groups), secondary_endpoint='E_tail', seed1_regional_status='COMPLETE' if include_seed else 'NOT_COMPUTED', constant_truth_parts=sum(r['tail']['truth']['constant_field'] for r in rows), truth_threshold_tied_parts=sum(r['tail']['truth']['plateau_nodes'] > 1 for r in rows), arms={arm:dict( threshold_tied_parts=sum(r['tail']['arms'][arm]['prediction']['plateau_nodes'] > 1 for r in rows), peak={domain:dict( certified=sum(r['tail']['arms'][arm][domain]['certified'] for r in rows), near_boundary=sum(r['tail']['arms'][arm][domain]['near_boundary'] for r in rows), trivial_radius=sum(r['tail']['arms'][arm][domain]['trivial_radius'] for r in rows), categories={key:sum(r['tail']['arms'][arm][domain]['category'] == key for r in rows) for key in sorted({r['tail']['arms'][arm][domain]['category'] for r in rows})}, ) for domain in ('query','full')}, ) for arm in arms})