tboudreaux/MeanField
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Audit coupled trial artifacts and physical replay independently

Browse Source
This commit is contained in:
Emily Boudreaux
2026-09-18 09:52:30 -04:00
parent 0d87b8b919
commit 2d0d32b886
2 changed files with 787 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

452
experiments/check_coupled_trial.py Normal file
View File

@@ -0,0 +1,452 @@
#!/usr/bin/env python3
"""Audit coupled-trial saved vectors and transactions, not physical convergence.
A retained correction can still fail Armijo. Each alpha is an independent frozen
trial, not one step in a nonlinear trajectory. Saved J actions are checked by
vector algebra; this checker does not perform another FEM operator application.
"""
import argparse
import json
import math
from pathlib import Path
from check_nonlinear_step import EPS, close, digest, dot, layout_for, minus, norm, rows, sanitize, vector
P = 'coupled_trial_'
STATUS = ('not_attempted','original_accepted','budget_exhausted','kept','linear_failure',
'norm_limit','non_finite','inadmissible','not_improved')
GRAVITY_VALUES = {'gravity_gradient','gravity_potential'}
GRAVITY_ROWS = {'gravity_gradient_relation','poisson_balance'}
BASE = {'metadata.txt'} | {P+n+'.csv' for n in ('summary','layout','preparations','trials','blocks','physical')} | {
P+n+'.txt' for n in ('base_normalized','base_physical','base_residual','state_factors','residual_factors',
'direction','base_action','base_defect','restored_residual','restored_action')}
def near_vector(a,b,scale):
return len(a)==len(b) and norm(minus(a,b)) <= 1024*EPS*max(scale,1.e-28)
def inspect_case(folder):
try:
return inspect_case_artifacts(Path(folder))
except (OSError,ValueError,KeyError,TypeError,IndexError,OverflowError,ZeroDivisionError) as error:
return {'passed':False,'checks':[{'check':'complete_readable_artifacts','passed':False,'error':str(error)}],
'observations_valid':False,'required_artifacts':sorted(BASE),'trial_outcomes':[]}
def inspect_case_artifacts(folder):
checks=[]
def check(name,value): checks.append({'check':name,'passed':bool(value)})
def flag(row,key,label):
value=float(row[key]);check(label+key+':boolean',value in (0,1));return value==1
required=set(BASE)
def get(name,needed=True):
path=folder/(P+name+'.txt')
if needed:
required.add(path.name);return vector(path)
check(name+':unavailable_file_absent',not path.exists());return None
summary_rows=rows(folder/(P+'summary.csv'))
summary={r['metric']:float(r['value']) for r in summary_rows}
check('unique_summary',len(summary)==len(summary_rows))
meta=dict(line.split('=',1) for line in (folder/'metadata.txt').read_text().splitlines() if '=' in line)
for key,metadata_key in {'outer_linear_tolerance':'linear_tol','seed_control':'coupled_seed_control',
'coupled_maximum_correction_ratio':'coupled_maximum_ratio','coupled_relative_tolerance':'coupled_tolerance',
'coupled_maximum_iterations':'coupled_iterations','coupled_restart':'coupled_restart',
'gravity_relative_tolerance':'gravity_tolerance','gravity_maximum_iterations':'gravity_iterations',
'gravity_restart':'gravity_restart','coupled_attempt_budget_per_independent_alpha':'coupled_attempts_per_iteration'}.items():
check('summary_metadata:'+key,summary[key]==float(meta[metadata_key]))
check('strategy_scope',meta['strategy'] in ('coupled','curvature-coupled') and
summary['gravity_path']==int(meta['strategy']=='curvature-coupled'))
for key in ('outer_linear_tolerance','sufficient_decrease'):
check('valid_control:'+key,math.isfinite(summary[key]) and 0<summary[key]<1)
check('valid_coupled_relative_tolerance',math.isfinite(summary['coupled_relative_tolerance']) and summary['coupled_relative_tolerance']>0)
check('valid_gravity_relative_tolerance',math.isfinite(summary['gravity_relative_tolerance']) and summary['gravity_relative_tolerance']>=0)
for key in ('coupled_maximum_iterations','coupled_restart','gravity_maximum_iterations','gravity_restart'):
check('valid_control:'+key,math.isfinite(summary[key]) and summary[key]>0 and summary[key]==int(summary[key]))
check('valid_correction_cap',math.isfinite(summary['coupled_maximum_correction_ratio']) and summary['coupled_maximum_correction_ratio']>0)
check('fixed_driver_controls',summary['sufficient_decrease']==1.e-4 and
summary['coupled_absolute_tolerance']==1.e-14 and summary['gravity_absolute_tolerance']==1.e-14)
for name,value in {'schema_version':1,'independent_frozen_alphas':1,'accepted_commits':0,
'polynomial_direction_repairs':0,'sampled_enthalpy_policy':1,'coupled_attempt_budget_per_independent_alpha':1,
'probe_exception':0,'accepted_buffers_bitwise_identical':1,'normalization_factors_bitwise_identical':1,
'direction_buffers_bitwise_identical':1,'direction_buffers_unchanged_before_restoration':1,
'fresh_operator_and_inverse_current':1}.items():check(name,summary[name]==value)
for i,name in enumerate(STATUS):check('status_mapping:'+name,summary['soc_status_'+name]==i)
x,physical,f,p,j,sf,rf=[get(n) for n in ('base_normalized','base_physical','base_residual','direction','base_action','state_factors','residual_factors')]
size=len(x)
check('square_vector_dimensions',all(len(v)==size for v in (physical,f,p,j,sf,rf)))
check('positive_normalization',all(v>0 for v in sf+rf))
check('base_physical_normalization',near_vector(x,[v*a for v,a in zip(physical,sf)],norm(x)))
layout=rows(folder/(P+'layout.csv'));vl=layout_for(layout,'value',size);rl=layout_for(layout,'residual',size)
gi=[i for name,(start,n) in vl.items() if name in GRAVITY_VALUES for i in range(start,start+n)]
gr=[i for name,(start,n) in rl.items() if name in GRAVITY_ROWS for i in range(start,start+n)]
f0,pn,jn=norm(f),norm(p),norm(j)
check('finite_positive_base_residual',math.isfinite(f0) and f0>0)
check('finite_direction_norm',math.isfinite(pn))
defect=[a+b for a,b in zip(f,j)];dn=norm(defect)
check('base_defect_vector',near_vector(get('base_defect'),defect,f0+jn))
outer_threshold=summary['outer_linear_tolerance']*f0;outer_gate=dn<=outer_threshold
for key,value in {'base_residual_norm':f0,'base_merit':.5*f0*f0,'direction_norm':pn,
'base_merit_slope':dot(f,j),'outer_verified_defect_norm':dn,'outer_verified_relative_defect':dn/f0,
'outer_verified_threshold':outer_threshold}.items():check(key,close(summary[key],value,f0+jn))
check('outer_gate',summary['outer_verified_gate']==int(outer_gate))
check('outer_failed_gate_skips_trials',summary['outer_gate_failure_skips_trials']==int(not outer_gate))
check('restored_residual_vector',near_vector(get('restored_residual'),f,f0))
check('restored_action_vector',near_vector(get('restored_action'),j,jn))
restored_f=norm(minus(get('restored_residual'),f));restored_j=norm(minus(get('restored_action'),j))
check('restoration_envelope',math.isfinite(summary['maximum_base_restoration_error']) and
restored_f<=summary['maximum_base_restoration_error']+1024*EPS*f0 and summary['maximum_base_restoration_error']<=1.e-10*f0)
check('restored_action_error',close(summary['restored_base_action_error'],restored_j,jn) and restored_j<=1.e-10*max(jn,1.e-28))
for name in ('gravity_path','seed_control'):flag(summary,name,'summary:')
expected_alphas=[1.] if summary['seed_control']==1 else [1/16,1/64,1/256,1/1024]
for i,a in enumerate(expected_alphas):check('alpha_'+str(i),summary['alpha_'+str(i)]==a)
trial_rows=rows(folder/(P+'trials.csv'));block_rows=rows(folder/(P+'blocks.csv'))
check('complete_trial_grid',[int(r['sample']) for r in trial_rows]==(list(range(len(expected_alphas))) if outer_gate else []))
block_map={(int(r['sample']),r['stage'],r['block']):r for r in block_rows}
check('unique_block_rows',len(block_map)==len(block_rows));seen_blocks=set()
partitions={'all':(0,size),**rl}
outcomes=[]
totals={'attempts':0,'kept':0,'fallbacks':0,'iterations':0,'preparations':0,'restorations':0,
'solve_seconds':0.,'verification_seconds':0.,'preparation_seconds':0.,'refresh_seconds':0.}
def block_check(sample,a,stage,residual):
for name,(start,n) in partitions.items():
key=(sample,stage,name);seen_blocks.add(key);r=block_map[key]
check(f'{sample}:{stage}:{name}:layout',int(r['offset'])==start and int(r['size'])==n and float(r['alpha'])==a)
check(f'{sample}:{stage}:{name}:norm',close(float(r['residual_norm']),norm(residual[start:start+n]),norm(residual)))
for row in trial_rows:
i=int(row['sample']);a=float(row['alpha']);label=str(i)+':'
check(label+'alpha',a==expected_alphas[i]);suffix='_a'+str(i)
flags={k:flag(row,k,label) for k in ('gravity_path','raw_prepared','gravity_attempted','gravity_reported_gate',
'gravity_prepared_gate','before_prepared','before_usable','before_accepted','soc_called','soc_attempted',
'aux_reported_gate','aux_verified_gate','aux_verification_applied','corrected_metric_observed',
'corrected_accepted','selected_prepared','selected_accepted','selected_is_original','rhs_saved',
'correction_saved','verified_vectors_saved')}
check(label+'gravity_policy',int(flags['gravity_path'])==summary['gravity_path'])
prediction=[v+a*w for v,w in zip(f,j)];raw_state=get('raw_state'+suffix)
check(label+'linear_prediction',near_vector(get('prediction'+suffix),prediction,f0+a*jn))
check(label+'raw_displacement',near_vector(raw_state,[v+a*w for v,w in zip(x,p)],norm(x)+a*pn))
vectors={}
for stage,available,normkey,meritkey in (
('raw',flags['raw_prepared'],'raw_norm','raw_merit'),
('before',flags['before_prepared'],'before_norm','before_merit'),
('corrected',flags['corrected_metric_observed'],'corrected_norm','corrected_merit'),
('selected',flags['selected_prepared'],'selected_norm','selected_merit')):
residual=get(stage+'_residual'+suffix,available)
state=raw_state if stage=='raw' else get(stage+'_state'+suffix,available)
if available:
check(label+stage+':dimensions',len(residual)==len(state)==size)
rn=norm(residual);check(label+stage+':norm',close(float(row[normkey]),rn,rn))
check(label+stage+':merit',close(float(row[meritkey]),.5*rn*rn,rn*rn))
vectors[stage]=(state,residual);block_check(i,a,stage,residual)
if stage in ('corrected','selected'):
pv=get(stage+'_physical'+suffix)
check(label+stage+':physical',near_vector(state,[v*b for v,b in zip(pv,sf)],norm(state)))
else:
check(label+stage+':unavailable_metrics',math.isnan(float(row[normkey])) and math.isnan(float(row[meritkey])))
if stage in ('corrected','selected'):get(stage+'_physical'+suffix,False)
check(label+'raw_rejection', (int(row['raw_rejection_reason'])==-1 and int(row['raw_rejection_stage'])==-1)
if flags['raw_prepared'] else (int(row['raw_rejection_reason'])>=0 and int(row['raw_rejection_stage'])>=0))
check(label+'gravity_attempt',flags['gravity_attempted']==(flags['gravity_path'] and flags['raw_prepared']))
if not flags['gravity_path']:
check(label+'vacuous_gravity_gates',flags['gravity_reported_gate'] and flags['gravity_prepared_gate'])
check(label+'raw_baseline_preparation',flags['before_prepared']==flags['raw_prepared'])
get('gravity_correction'+suffix,False)
if flags['before_prepared']:
check(label+'raw_baseline_state',vectors['before'][0]==raw_state)
check(label+'raw_baseline_residual',vectors['before'][1]==vectors['raw'][1])
else:
save_gravity=flags['gravity_attempted'] and flags['gravity_reported_gate']
gc=get('gravity_correction'+suffix,save_gravity)
if save_gravity:
check(label+'gravity_partition_present',bool(gi) and bool(gr))
check(label+'gravity_only_correction',len(gc)==size and all(v==0 for k,v in enumerate(gc) if k not in set(gi)))
if flags['before_prepared']:
check(label+'gravity_displacement',near_vector(vectors['before'][0],[v+w for v,w in zip(raw_state,gc)],norm(raw_state)+norm(gc)))
target=norm([vectors['before'][1][k]-prediction[k] for k in gr])
check(label+'gravity_affine_target',close(float(row['gravity_target_defect']),target,f0+a*jn))
if flags['gravity_prepared_gate']:
threshold=float(row['gravity_target_threshold']);disagreement=float(row['gravity_affine_disagreement'])
check(label+'gravity_prepared_necessary_gate',math.isfinite(threshold) and threshold>=0 and target<=threshold+512*EPS*f0 and
math.isfinite(disagreement) and 0<=disagreement<=threshold)
usable=flags['before_prepared'] and flags['gravity_reported_gate'] and flags['gravity_prepared_gate']
check(label+'before_usable',flags['before_usable']==usable)
armijo=(1-summary['sufficient_decrease']*a)*.5*f0*f0
before_accept=usable and float(row['before_merit'])<=armijo
check(label+'before_armijo',flags['before_accepted']==before_accept)
check(label+'soc_called',flags['soc_called']==usable)
attempted=usable and not before_accept
check(label+'one_independent_attempt',flags['soc_attempted']==attempted)
status=int(row['soc_status']);check(label+'known_status',0<=status<len(STATUS))
name=STATUS[status]
rhs=get('soc_rhs'+suffix,flags['rhs_saved']);correction=get('soc_correction'+suffix,flags['correction_saved'])
action=get('soc_action'+suffix,flags['verified_vectors_saved']);saved_defect=get('soc_defect'+suffix,flags['verified_vectors_saved'])
check(label+'saved_data_requires_attempt',attempted or not any(flags[k] for k in ('rhs_saved','correction_saved','verified_vectors_saved')))
# Save flags cannot hide a finite verification that reported failure.
# Nonfinite actions may legitimately have no finite vector artifact.
finite_verification=flags['aux_verification_applied'] and math.isfinite(float(row['aux_verified_true_norm']))
check(label+'verification_save_contract',
(not flags['verified_vectors_saved'] or (flags['aux_verification_applied'] and flags['rhs_saved'] and flags['correction_saved'])) and
(not finite_verification or flags['verified_vectors_saved']))
if attempted:
totals['attempts']+=1;totals['kept']+=name=='kept';totals['fallbacks']+=name!='kept'
check(label+'aux_status_gate',flags['aux_reported_gate']==(int(row['aux_status'])==0))
check(label+'bounded_iterations',0<=int(row['aux_iterations'])<=summary['coupled_maximum_iterations'])
cap=summary['coupled_maximum_correction_ratio']*a*pn
check(label+'correction_cap',close(float(row['correction_cap']),cap,cap))
if rhs is not None:
wanted=[u-v for u,v in zip(prediction,vectors['before'][1])]
check(label+'aux_rhs',near_vector(rhs,wanted,norm(prediction)+norm(vectors['before'][1])))
rhs_norm=norm(rhs)
if math.isfinite(rhs_norm):
threshold=max(summary['coupled_absolute_tolerance'],summary['coupled_relative_tolerance']*rhs_norm)
check(label+'aux_rhs_norm',close(float(row['aux_rhs_norm']),rhs_norm,rhs_norm))
check(label+'aux_threshold',close(float(row['aux_threshold']),threshold,threshold))
if flags['verified_vectors_saved']:
delta=minus(action,rhs);error=norm(delta)
check(label+'verified_defect_vector',near_vector(saved_defect,delta,norm(action)+norm(rhs)))
check(label+'verified_defect_norm',close(float(row['aux_verified_true_norm']),error,norm(rhs)))
check(label+'saved_defect_norm',close(float(row['verified_action_defect_norm']),error,norm(rhs)))
check(label+'independent_aux_gate',flags['aux_verified_gate']==(error<=float(row['aux_threshold'])))
cn=norm(correction)
check(label+'correction_norm',close(float(row['correction_norm']),cn,cn))
check(label+'correction_ratio',close(float(row['correction_ratio']),cn/(a*pn),cn/(a*pn)))
expected_core=1 if not flags['aux_reported_gate'] or not flags['aux_verified_gate'] else (2 if cn>cap else 0)
check(label+'core_decision',int(row['aux_core_status'])==expected_core)
core=int(row['aux_core_status'])
if core in (0,1,2):
check(label+'finite_core_complete_evidence',flags['rhs_saved'] and flags['correction_saved'] and
flags['aux_verification_applied'] and flags['verified_vectors_saved'])
if core!=0:
check(label+'failed_core_falls_back',name=={1:'linear_failure',2:'norm_limit',3:'non_finite'}.get(core))
check(label+'failed_core_no_preparation',int(row['soc_preparations'])==0 and int(row['soc_restorations'])==0 and not flags['corrected_metric_observed'])
else:
check(label+'admissible_core_requires_gates',flags['aux_reported_gate'] and flags['aux_verified_gate'] and
flags['verified_vectors_saved'] and float(row['correction_norm'])<=cap)
if flags['corrected_metric_observed']:
check(label+'corrected_requires_admissible_core',core==0)
check(label+'corrected_displacement',near_vector(vectors['corrected'][0],
[v+w for v,w in zip(vectors['before'][0],correction)],norm(vectors['before'][0])+norm(correction)))
better=float(row['corrected_merit'])<float(row['before_merit'])
check(label+'full_merit_selection',name==('kept' if better else 'not_improved'))
if name=='kept':
check(label+'kept_observed',flags['corrected_metric_observed'])
check(label+'kept_preparations',int(row['soc_preparations'])==1 and int(row['soc_restorations'])==0)
elif core==0:
check(label+'mutated_fallback_restored',int(row['soc_restorations'])==1 and int(row['soc_preparations']) in (1,2))
else:
check(label+'no_attempt_status',name==('original_accepted' if before_accept else 'not_attempted'))
check(label+'no_aux_certificate',int(row['aux_status'])==-1 and int(row['aux_core_status'])==-1 and
not flags['aux_reported_gate'] and not flags['aux_verified_gate'] and not flags['aux_verification_applied'])
check(label+'no_aux_work',int(row['aux_iterations'])==0 and int(row['soc_preparations'])==0 and int(row['soc_restorations'])==0)
check(label+'no_corrected_candidate',not flags['corrected_metric_observed'])
check(label+'selected_prepared',flags['selected_prepared']==flags['before_prepared'])
if flags['selected_prepared']:
expected=vectors['corrected'] if name=='kept' else vectors['before']
check(label+'selected_state',vectors['selected'][0]==expected[0])
check(label+'selected_residual',near_vector(vectors['selected'][1],expected[1],norm(expected[1])))
same=vectors['selected'][0]==vectors['before'][0]
check(label+'selected_original_identity',flags['selected_is_original']==same)
if same:
difference=norm(minus(vectors['selected'][1],vectors['before'][1]))
check(label+'fallback_difference',close(float(row['fallback_residual_difference']),difference,norm(expected[1])))
check(label+'fallback_residual_preserved',difference<=1024*EPS*max(norm(expected[1]),1.e-28))
else:check(label+'no_original_identity',not flags['selected_is_original'])
check(label+'corrected_armijo',flags['corrected_accepted']==(flags['corrected_metric_observed'] and float(row['corrected_merit'])<=armijo))
check(label+'selected_armijo',flags['selected_accepted']==(usable and flags['selected_prepared'] and float(row['selected_merit'])<=armijo))
for key in ('raw_seconds','gravity_seconds','soc_seconds','soc_refresh_seconds','soc_solve_seconds','soc_verification_seconds',
'soc_preparation_seconds','capture_seconds','total_sample_seconds'):
check(label+key,math.isfinite(float(row[key])) and float(row[key])>=0)
parts=sum(float(row['soc_'+k]) for k in ('refresh_seconds','solve_seconds','verification_seconds','preparation_seconds'))
check(label+'soc_time_scope',float(row['soc_seconds'])+1.e-6>=parts)
check(label+'sample_time_scope',float(row['total_sample_seconds'])+1.e-6>=sum(float(row[k]) for k in ('raw_seconds','gravity_seconds','soc_seconds','capture_seconds')))
for key in ('preparations','restorations','solve_seconds','verification_seconds','preparation_seconds','refresh_seconds'):
totals[key]+=float(row['soc_'+key])
totals['iterations']+=int(row['aux_iterations'])
outcomes.append({'sample':i,'alpha':a,'transaction_status':name,'before_usable':usable,'before_armijo':before_accept,
'correction_attempted':attempted,'aux_verified_gate':flags['aux_verified_gate'],'corrected_metric_observed':flags['corrected_metric_observed'],
'selected_armijo':flags['selected_accepted'],'before_residual_norm':float(row['before_norm']),
'selected_residual_norm':float(row['selected_norm']),'full_merit_improved':name=='kept',
'correction_norm':float(row['correction_norm']),'correction_cap':float(row['correction_cap']),
'auxiliary_iterations':int(row['aux_iterations']),'auxiliary_verified_defect_norm':float(row['aux_verified_true_norm']),
'corrected_affine_target_defect_norm':norm(minus(vectors['corrected'][1],prediction)) if flags['corrected_metric_observed'] else None,
'auxiliary_seconds':{k:float(row['soc_'+k]) for k in ('refresh_seconds','solve_seconds','verification_seconds','preparation_seconds')},
'residual_norms_by_block':{stage:{block:norm(value[1][start:start+n]) for block,(start,n) in rl.items()}
for stage,value in vectors.items()}})
check('complete_block_coverage',seen_blocks==set(block_map))
if outer_gate:
for key,value in totals.items():check('soc_total:'+key,close(summary['soc_'+key],value,value))
preparations=rows(folder/(P+'preparations.csv'))
check('preparation_sequence',[int(r['sequence']) for r in preparations]==list(range(1,len(preparations)+1)))
check('explicit_preparation_count',summary['explicit_probe_preparations']==len(preparations))
for r in preparations:
check('preparation_time:'+r['sequence'],math.isfinite(float(r['seconds'])) and float(r['seconds'])>=0)
if r['kind']!='raw':check('base_restoration_prepares:'+r['sequence'],r['prepared']=='1' and close(float(r['residual_norm']),f0,f0))
raw_preparations=[r for r in preparations if r['kind']=='raw']
check('raw_preparation_coverage',[int(r['sample']) for r in raw_preparations]==[int(r['sample']) for r in trial_rows])
for prep,trial in zip(raw_preparations,trial_rows):
label=prep['sample']+':raw_preparation:'
for key in ('prepared','rejection_reason','rejection_stage'):
check(label+key,prep[key]==trial['raw_'+key])
check(label+'alpha',float(prep['alpha'])==float(trial['alpha']))
measured=float(prep['residual_norm']);reported=float(trial['raw_norm'])
check(label+'residual',close(measured,reported) if prep['prepared']=='1' else math.isnan(measured) and math.isnan(reported))
check(label+'time_scope',float(prep['seconds'])<=float(trial['raw_seconds'])+1.e-6)
check('preparation_time_total',close(summary['explicit_probe_preparation_seconds'],sum(float(r['seconds']) for r in preparations)))
check('final_restoration_present',bool(preparations) and preparations[-1]['kind']=='final_base_restore')
check('total_time',math.isfinite(summary['total_seconds']) and summary['total_seconds']>=0)
physical_observations=inspect_physical(folder,trial_rows,meta,check)
check('no_unexplained_vector_artifacts',{p.name for p in folder.glob(P+'*.txt')}==
{name for name in required if name.startswith(P) and name.endswith('.txt')})
passed=all(r['passed'] for r in checks)
return sanitize({'passed':passed,'checks':checks,'observations_valid':passed,'metadata':meta,'summary':summary,
'required_artifacts':sorted(required),'outer_linear_gate':outer_gate,'trial_outcomes':outcomes if passed else [],
'physical_observations':physical_observations if passed else [],
'scope':'Independent saved-vector arithmetic and transaction/provenance audit. Frozen alpha samples are not a nonlinear trajectory. Physical replay metrics are checked for internal consistency and reported separately; no independent FEM quadrature is performed.',
'gravity_scope':'Saved vectors independently check the affine gravity target defect. This schema does not independently reconstruct the gravity inner solve or its contribution-based arithmetic allowance.',
'operator_scope':'Saved SOC J actions verify the reported algebraic defect; this checker does not perform another FEM action.'})
def inspect_physical(folder,trials,metadata,check):
"""Check replay binding and scalar identities without certifying the star."""
records=rows(folder/(P+'physical.csv'))
measured={(int(r['sample']),r['source'],int(r['quadrature']),r['metric']):float(r['value']) for r in records}
check('physical:unique_measurements',len(measured)==len(records))
used=set()
def get(sample,source,q,key):
index=(sample,source,q,key);used.add(index);return measured[index]
for key in ('accepted_buffers_unchanged_before_restoration','accepted_buffers_unchanged_after_restoration',
'normalization_factors_unchanged','direction_buffers_unchanged','accepted_operator_and_inverse_current'):
check('physical:'+key,get(-1,'restoration',0,key)==1)
selected=[r for r in trials if r['selected_prepared']=='1']
check('physical:snapshot_count',get(-1,'replay',0,'selected_snapshots_measured')==len(selected))
check('physical:sample_coverage',{k[0] for k in measured if k[0]>=0}=={int(r['sample']) for r in selected})
observations=[]
for trial in selected:
sample=int(trial['sample']);source=P+'selected_physical_a'+str(sample)+'.txt'
saved=vector(folder/source);label='physical:'+str(sample)+':'
check(label+'fresh_preparation',get(sample,source,0,'fresh_preparation_passed')==1)
for q in (14,18):
prefix=label+str(q)+':'
m={key[3]:value for key,value in measured.items() if key[:3]==(sample,source,q)}
used.update(key for key in measured if key[:3]==(sample,source,q))
check(prefix+'quadrature',m['quadrature_order']==q)
check(prefix+'physics',m['polytropic_index']==float(metadata['n']) and
m['polytropic_constant']==float(metadata['K']) and m['target_mass']==float(metadata['M']))
for key in ('stellar_samples','valid_stellar_samples','invalid_stellar_samples','negative_density_samples',
'negative_enthalpy_samples','invalid_density_eos_samples','invalid_enthalpy_eos_samples'):
check(prefix+key,math.isfinite(m[key]) and m[key]>=0 and m[key]==int(m[key]))
check(prefix+'sample_accounting',m['valid_stellar_samples']+m['invalid_stellar_samples']==m['stellar_samples'])
complete=m['invalid_stellar_samples']==0 and m['valid_stellar_samples']>0 and math.isfinite(m['volume']) and m['volume']>0
check(prefix+'complete_measurement',m['complete_volume_measurement']==int(complete))
hp=complete and m['invalid_enthalpy_eos_samples']==0
rp=complete and m['negative_density_samples']==0 and m['invalid_density_eos_samples']==0
check(prefix+'enthalpy_pressure_definition',m['pressure_from_enthalpy_defined_everywhere']==int(hp))
check(prefix+'density_pressure_definition',m['pressure_from_density_defined_everywhere']==int(rp))
check(prefix+'snapshot_roundtrip',0<=m['saved_prepared_l2_difference']<=1024*EPS*(1+norm(saved)) and
0<=m['saved_prepared_maximum_difference']<=m['saved_prepared_l2_difference']+1024*EPS and
0<=m['saved_prepared_maximum_scaled_difference']<=1024*EPS)
selected_norm=float(trial['selected_norm'])
# Repreparing a denormalized snapshot introduces roundoff. This
# tolerance checks replay identity, not physical residual accuracy.
check(prefix+'selected_residual_binding',math.isfinite(m['normalized_bordered_residual']) and
abs(m['normalized_bordered_residual']-selected_norm)<=1024*EPS*(1+selected_norm))
check(prefix+'central_border_finite',math.isfinite(m['central_border']) and math.isfinite(m['normalized_border_action']))
if metadata['central_phase']=='0':
check(prefix+'unbordered_model',m['central_border']==0 and m['normalized_border_action']==0 and
close(m['normalized_unbordered_residual'],m['normalized_bordered_residual'],selected_norm))
if complete:
check(prefix+'mass_error',close(m['mass_absolute_error'],abs(m['mass']-m['target_mass']),abs(m['mass'])+m['target_mass']))
check(prefix+'sampled_signs',
(m['minimum_density']<0)==(m['negative_density_samples']>0) and
(m['minimum_enthalpy']<0)==(m['negative_enthalpy_samples']>0))
w=m['binding_energy'];wf=m['force_binding_energy'];t=m['kinetic_energy']
virials={}
for prefix_pressure,defined,pressure_key in (('',rp,'pressure_integral'),('enthalpy_',hp,'enthalpy_pressure_integral')):
pressure=m[pressure_key]
check(prefix+pressure_key+':definition',math.isfinite(pressure) if defined else math.isnan(pressure))
for force,binding in (('',w),('force_',wf)):
key=prefix_pressure+force+'virial_';signed=m[key+'signed'];error=m[key+'error']
valid=defined and math.isfinite(w) and w<0
if valid:
expected=(2*t+binding+3*pressure)/abs(w)
scale=(2*abs(t)+abs(binding)+3*abs(pressure))/abs(w)
check(prefix+key+'identity',close(signed,expected,scale) and close(error,abs(signed),scale))
else:check(prefix+key+'undefined',math.isnan(signed) and math.isnan(error))
virials[key+'within_1e5']=(error<=1.e-5) if valid and math.isfinite(error) else None
observations.append({'sample':sample,'quadrature':q,'selected_armijo':trial['selected_accepted']=='1',
'complete_volume_measurement':complete,'pressure_from_density_defined':rp,'pressure_from_enthalpy_defined':hp,
'negative_density_samples':m['negative_density_samples'],'minimum_density':m['minimum_density'],
'negative_enthalpy_samples':m['negative_enthalpy_samples'],'minimum_enthalpy':m['minimum_enthalpy'],
'global_enthalpy_certificate':'not_measured','virial_status':virials,'reported_metrics':m})
check('physical:source_and_quadrature_coverage',used==set(measured))
return observations
def suppress_observations(results):
for result in results.values():
result.update(passed=False,observations_valid=False,trial_outcomes=[],physical_observations=[])
def inspect_campaign(campaign):
campaign=Path(campaign).resolve();checks=[];results={}
def check(name,value):checks.append({'check':name,'passed':bool(value)})
try:
provenance=json.loads((campaign/'provenance.json').read_text());cases=provenance['cases']
check('executable_hash',digest(campaign/'plateau_diagnostic')==provenance['binary_sha256'])
check('mesh_hash',digest(Path(provenance['mesh']))==provenance['mesh_sha256'])
check('unique_cases',bool(cases) and len(cases)==len({c['id'] for c in cases}))
check('manifest_snapshot',json.loads((campaign/'cases.json').read_text())==cases)
for case in cases:
name=case['id'];folder=campaign/name;args=case['arguments'];record=json.loads((campaign/(name+'_run.json')).read_text())
check(name+':completed',bool(record.get('finished_utc')) and record.get('exit_code')==0 and not record.get('timed_out',False))
check(name+':command',record.get('id')==name and record.get('command')==[str(campaign/'plateau_diagnostic'),*args,'--mesh',provenance['mesh'],'--output',str(folder)])
expected={}
for option,saved in (('--probe-state','base_physical'),('--probe-direction','direction')):
if option not in args:raise ValueError('A frozen coupled trial must bind '+option)
path=Path(args[args.index(option)+1]);path=path if path.is_absolute() else Path(provenance['source'])/path
expected[str(path.resolve())]=saved
check(name+':required_input_hashes',expected.keys()<=record.get('input_state_hashes',{}).keys())
for path,value in record.get('input_state_hashes',{}).items():check(name+':input:'+path,digest(Path(path))==value)
for filename,value in record.get('output_hashes',{}).items():check(name+':output:'+filename,digest(folder/filename)==value)
for path,saved in expected.items():
source=vector(Path(path));target=vector(folder/(P+saved+'.txt'))
check(name+':input_binding:'+saved,source==target if saved=='direction' else near_vector(source,target,norm(source)))
result=inspect_case(folder);results[name]=result;meta=result.get('metadata',{})
check(name+':required_output_hashes',set(result['required_artifacts'])<=record.get('output_hashes',{}).keys())
check(name+':expert_probe','--coupled-trial-probe' in args and meta.get('coupled_trial_probe')=='1')
bindings={'--index':'n','--angular-momentum':'J','--linear-tol':'linear_tol','--coupled-tol':'coupled_tolerance',
'--coupled-iterations':'coupled_iterations','--coupled-restart':'coupled_restart','--coupled-max-ratio':'coupled_maximum_ratio',
'--coupled-attempts':'coupled_attempts_per_iteration','--gravity-tol':'gravity_tolerance',
'--gravity-iterations':'gravity_iterations','--gravity-restart':'gravity_restart'}
for option,key in bindings.items():
if option in args:check(name+':requested_'+key,key in meta and float(meta[key])==float(args[args.index(option)+1]))
if '--strategy' in args:check(name+':strategy',meta.get('strategy')==args[args.index('--strategy')+1])
if '--no-central' in args:check(name+':no_central',meta.get('central_phase')=='0')
check(name+':independent_alpha_metadata',meta.get('coupled_trial_alphas_are_independent')=='1')
check(name+':sampled_enthalpy_scope',meta.get('coupled_trial_admissibility')=='sampled_enthalpy_without_repair')
check(name+':seed_control',meta.get('coupled_seed_control')==str(int('--coupled-seed-control' in args)))
passed=all(c['passed'] for c in checks) and all(r['passed'] for r in results.values())
if not all(c['passed'] for c in checks):
suppress_observations(results)
return {'passed':passed,'source_commit':provenance['commit'],'checker_sha256':digest(Path(__file__)),
'checker_utility_sha256':digest(Path(__file__).with_name('check_nonlinear_step.py')),
'provenance_sha256':digest(campaign/'provenance.json'),'provenance_checks':checks,'cases':results,
'scope':'Artifact validity and numerical/physical success are separate conclusions.'}
except (OSError,ValueError,KeyError,TypeError,IndexError) as error:
suppress_observations(results)
return {'passed':False,'provenance_checks':checks+[{'check':'complete_readable_campaign','passed':False,'error':str(error)}],'cases':results}
def main():
parser=argparse.ArgumentParser(description=__doc__)
parser.add_argument('campaign',type=Path);parser.add_argument('--output',type=Path,required=True)
args=parser.parse_args()
if args.output.resolve()==args.campaign.resolve() or args.campaign.resolve() in args.output.resolve().parents:
parser.error('Write the review outside the immutable campaign.')
result=inspect_campaign(args.campaign);args.output.write_text(json.dumps(result,indent=2,allow_nan=False)+'\n')
print('PASS' if result['passed'] else 'FAIL',args.output)
for name,case in result['cases'].items():print(name,'artifacts=',case['passed'],'outer_linear=',case.get('outer_linear_gate'))
return 0 if result['passed'] else 1
if __name__=='__main__':raise SystemExit(main())

335
experiments/test_coupled_trial_checker.py Normal file
View File

@@ -0,0 +1,335 @@
#!/usr/bin/env python3
"""Synthetic counterexamples guard against false coupled-correction successes."""
import csv
import json
import math
from pathlib import Path
import tempfile
import unittest
import check_coupled_trial as c
def csv_file(path,data,fields=None):
with path.open('w',newline='') as stream:
writer=csv.DictWriter(stream,fieldnames=fields or list(data[0]));writer.writeheader();writer.writerows(data)
def vec_file(path,values):
path.write_text(str(len(values))+'\n'+'\n'.join(repr(v) for v in values)+'\n')
class CoupledTrialCheckerTests(unittest.TestCase):
def setUp(self):
self.temp=tempfile.TemporaryDirectory();self.addCleanup(self.temp.cleanup)
self.campaign=Path(self.temp.name).resolve();self.folder=self.campaign/'synthetic';self.folder.mkdir()
self.make_case()
def make_case(self,mode='kept'):
for path in self.folder.iterdir():path.unlink()
folder=self.folder;nan=math.nan
x=[0.,0.,0.];f=[1.,1.,1.];p=[-1.,-1.,-1.];j=p[:];raw_state=p[:]
factors=[2.]*3;fn=math.sqrt(3.);pn=fn;alpha=1.
raw_prepared=mode!='raw_failure';attempted=mode not in ('original_accepted','raw_failure')
before=[.5,0.,0.] if mode=='original_accepted' else [10.,0.,0.]
rhs=[-v for v in before]
correction=[-5.,0.,0.] if mode=='failed_gate' else [-10.,0.,0.]
action=correction[:];defect=[v-w for v,w in zip(action,rhs)]
cn=c.norm(correction);dn=c.norm(defect)
cap_ratio=1. if mode=='cap' else 10.;cap=cap_ratio*pn
core=1 if mode=='failed_gate' else (2 if mode=='cap' else 0)
corrected_observed=attempted and core==0
corrected=[12100.,0.,0.] if mode=='worse' else ([2.,0.,0.] if mode=='improved_not_accepted' else [0.,0.,0.])
corrected_state=[v+w for v,w in zip(raw_state,correction)]
kept=corrected_observed and mode!='worse'
status='kept' if kept else ('not_improved' if mode=='worse' else ('linear_failure' if mode=='failed_gate' else
('norm_limit' if mode=='cap' else ('original_accepted' if mode=='original_accepted' else 'not_attempted'))))
selected_state=corrected_state if kept else raw_state[:]
selected=corrected if kept else before[:]
aux_preps=1 if kept else (2 if mode=='worse' else 0)
aux_restores=int(mode=='worse')
def save(name,values):vec_file(folder/(c.P+name+'.txt'),values)
for name,values in {'base_normalized':x,'base_physical':x,'base_residual':f,'state_factors':factors,
'residual_factors':[3.]*3,'direction':p,'base_action':j,'base_defect':[0.]*3,'restored_residual':f,
'restored_action':j,'prediction_a0':[0.]*3,'raw_state_a0':raw_state}.items():save(name,values)
if raw_prepared:
for name,values in {'raw_residual_a0':before,'before_state_a0':raw_state,'before_residual_a0':before,
'selected_state_a0':selected_state,'selected_residual_a0':selected,
'selected_physical_a0':[v/2 for v in selected_state]}.items():save(name,values)
if attempted:
for name,values in {'soc_rhs_a0':rhs,'soc_correction_a0':correction,'soc_action_a0':action,'soc_defect_a0':defect}.items():save(name,values)
if corrected_observed:
for name,values in {'corrected_state_a0':corrected_state,'corrected_residual_a0':corrected,
'corrected_physical_a0':[v/2 for v in corrected_state]}.items():save(name,values)
layout=[{'kind':kind,'block':name,'offset':i,'size':1} for kind,names in (
('value',['density','surface_deformation','specific_enthalpy']),
('residual',['closure','shape','hydrostatic'])) for i,name in enumerate(names)]
csv_file(folder/(c.P+'layout.csv'),layout)
def merit(v):return .5*c.norm(v)**2
cutoff=(1-1.e-4)*1.5
row={'sample':0,'alpha':alpha,'gravity_path':0,'raw_prepared':int(raw_prepared),
'raw_rejection_reason':-1 if raw_prepared else 2,'raw_rejection_stage':-1 if raw_prepared else 6,
'raw_norm':c.norm(before) if raw_prepared else nan,'raw_merit':merit(before) if raw_prepared else nan,
'gravity_attempted':0,'gravity_reported_gate':1,'gravity_prepared_gate':1,'gravity_target_defect':nan,
'gravity_affine_disagreement':nan,'gravity_target_threshold':nan,'before_prepared':int(raw_prepared),
'before_usable':int(raw_prepared),'before_norm':c.norm(before) if raw_prepared else nan,
'before_merit':merit(before) if raw_prepared else nan,'before_accepted':int(raw_prepared and merit(before)<=cutoff),
'soc_called':int(raw_prepared),'soc_status':c.STATUS.index(status),'soc_attempted':int(attempted),
'aux_status':0 if attempted else -1,'aux_reported_gate':int(attempted),'aux_iterations':1 if attempted else 0,
'aux_rhs_norm':c.norm(rhs) if attempted else nan,'aux_reported_true_norm':0. if attempted else nan,
'aux_verified_true_norm':dn if attempted else nan,'aux_verified_gate':int(attempted and dn<=.03*c.norm(rhs)),
'aux_threshold':.03*c.norm(rhs) if attempted else nan,'aux_verification_applied':int(attempted),
'correction_norm':cn if attempted else nan,'correction_cap':cap if attempted else nan,
'correction_ratio':cn/pn if attempted else nan,'corrected_metric_observed':int(corrected_observed),
'corrected_norm':c.norm(corrected) if corrected_observed else nan,
'corrected_merit':merit(corrected) if corrected_observed else nan,
'corrected_accepted':int(corrected_observed and merit(corrected)<=cutoff),'selected_prepared':int(raw_prepared),
'selected_norm':c.norm(selected) if raw_prepared else nan,'selected_merit':merit(selected) if raw_prepared else nan,
'selected_accepted':int(raw_prepared and merit(selected)<=cutoff),'selected_is_original':int(raw_prepared and not kept),
'fallback_residual_difference':0. if raw_prepared and not kept else nan,
'raw_seconds':.1,'gravity_seconds':0.,'soc_seconds':.3 if attempted else 0.,
'soc_refresh_seconds':.02 if attempted else 0.,'soc_solve_seconds':.1 if attempted else 0.,
'soc_verification_seconds':.01 if attempted else 0.,'soc_preparation_seconds':.05*aux_preps,
'soc_preparations':aux_preps,'soc_restorations':aux_restores,'aux_operator_applications':2 if attempted else 0,
'aux_inverse_applications':1 if attempted else 0,'verified_action_defect_norm':dn if attempted else nan,
'capture_seconds':.1,'total_sample_seconds':2.,'aux_core_status':core if attempted else -1,
'rhs_saved':int(attempted),'correction_saved':int(attempted),'verified_vectors_saved':int(attempted)}
csv_file(folder/(c.P+'trials.csv'),[row])
blocks=[]
for stage,values in [('raw',before),('before',before),('corrected',corrected),('selected',selected)]:
if not raw_prepared or (stage=='corrected' and not corrected_observed):continue
for name,start,size in [('all',0,3),('closure',0,1),('shape',1,1),('hydrostatic',2,1)]:
blocks.append({'sample':0,'alpha':1.,'stage':stage,'block':name,'offset':start,'size':size,
'residual_norm':c.norm(values[start:start+size])})
csv_file(folder/(c.P+'blocks.csv'),blocks,['sample','alpha','stage','block','offset','size','residual_norm'])
preparations=[]
for kind,prepared,rn in [('initial_base_restore',True,fn),('capture_base',True,fn),('raw',raw_prepared,c.norm(before)),('final_base_restore',True,fn)]:
preparations.append({'sequence':len(preparations)+1,'kind':kind,'sample':0 if kind=='raw' else -1,
'alpha':1. if kind=='raw' else 0.,'prepared':int(prepared),'rejection_reason':-1 if prepared else 2,
'rejection_stage':-1 if prepared else 6,'seconds':.1,'residual_norm':rn if prepared else nan})
csv_file(folder/(c.P+'preparations.csv'),preparations)
summary={'schema_version':1,'independent_frozen_alphas':1,'accepted_commits':0,'outer_linear_tolerance':.03,
'sufficient_decrease':1.e-4,'gravity_path':0,'seed_control':1,'coupled_maximum_correction_ratio':cap_ratio,
'coupled_relative_tolerance':.03,'coupled_absolute_tolerance':1.e-14,'coupled_maximum_iterations':10,'coupled_restart':4,
'gravity_relative_tolerance':1.e-10,'gravity_absolute_tolerance':1.e-14,'gravity_maximum_iterations':100,'gravity_restart':20,
'polynomial_direction_repairs':0,'sampled_enthalpy_policy':1,'coupled_attempt_budget_per_independent_alpha':1,
'alpha_0':1.,'base_residual_norm':fn,'base_merit':1.5,'direction_norm':pn,'base_merit_slope':-3.,
'outer_verified_defect_norm':0.,'outer_verified_relative_defect':0.,'outer_verified_threshold':.03*fn,
'outer_verified_gate':1,'outer_gate_failure_skips_trials':0,'soc_attempts':int(attempted),'soc_kept':int(kept),
'soc_fallbacks':int(attempted and not kept),'soc_iterations':int(attempted),
'soc_solve_seconds':row['soc_solve_seconds'],'soc_verification_seconds':row['soc_verification_seconds'],
'soc_preparation_seconds':row['soc_preparation_seconds'],'soc_refresh_seconds':row['soc_refresh_seconds'],
'soc_preparations':aux_preps,'soc_restorations':aux_restores,'probe_exception':0,
'accepted_buffers_bitwise_identical':1,'normalization_factors_bitwise_identical':1,'direction_buffers_bitwise_identical':1,
'direction_buffers_unchanged_before_restoration':1,'maximum_base_restoration_error':0.,'restored_base_action_error':0.,
'fresh_operator_and_inverse_current':1,'explicit_probe_preparations':4,'explicit_probe_preparation_seconds':.4,
'explicit_probe_inverse_refreshes':3,'explicit_probe_inverse_refresh_seconds':.06,
'explicit_probe_operator_applications':2,'explicit_probe_operator_seconds':.02,'total_seconds':10.}
summary.update({'soc_status_'+name:i for i,name in enumerate(c.STATUS)})
csv_file(folder/(c.P+'summary.csv'),[{'metric':k,'value':v} for k,v in summary.items()])
meta={'n':1.5,'K':1.,'M':1.,'J':0,'central_phase':0,'coupled_trial_probe':1,'linear_tol':.03,'coupled_tolerance':.03,
'coupled_iterations':10,'coupled_restart':4,'coupled_maximum_ratio':cap_ratio,'coupled_attempts_per_iteration':1,
'coupled_seed_control':1,'gravity_tolerance':1.e-10,'gravity_iterations':100,'gravity_restart':20,
'strategy':'coupled','coupled_trial_alphas_are_independent':1,'coupled_trial_admissibility':'sampled_enthalpy_without_repair'}
(folder/'metadata.txt').write_text('\n'.join(f'{k}={v}' for k,v in meta.items())+'\n')
physical=[]
def measurement(sample,source,q,key,value):
physical.append({'sample':sample,'source':source,'quadrature':q,'metric':key,'value':value})
for key in ('accepted_buffers_unchanged_before_restoration','accepted_buffers_unchanged_after_restoration',
'normalization_factors_unchanged','direction_buffers_unchanged','accepted_operator_and_inverse_current'):
measurement(-1,'restoration',0,key,1)
measurement(-1,'replay',0,'selected_snapshots_measured',int(raw_prepared))
if raw_prepared:
source=c.P+'selected_physical_a0.txt';measurement(0,source,0,'fresh_preparation_passed',1)
for q in (14,18):
metrics={'quadrature_order':q,'polytropic_index':1.5,'polytropic_constant':1.,'target_mass':1.,
'stellar_samples':10,'valid_stellar_samples':10,'invalid_stellar_samples':0,'negative_density_samples':0,
'negative_enthalpy_samples':0,'invalid_density_eos_samples':0,'invalid_enthalpy_eos_samples':0,
'volume':1.,'complete_volume_measurement':1,'pressure_from_enthalpy_defined_everywhere':1,
'pressure_from_density_defined_everywhere':1,'saved_prepared_l2_difference':0.,
'saved_prepared_maximum_difference':0.,'saved_prepared_maximum_scaled_difference':0.,
'normalized_bordered_residual':c.norm(selected),'normalized_unbordered_residual':c.norm(selected),
'normalized_border_action':0.,'central_border':0.,'mass':1.,'mass_absolute_error':0.,
'minimum_density':0.,'minimum_enthalpy':0.,'binding_energy':-3.,'force_binding_energy':-3.,
'kinetic_energy':0.,'pressure_integral':1.,'enthalpy_pressure_integral':1.}
for prefix in ('','enthalpy_'):
for force in ('','force_'):
metrics[prefix+force+'virial_signed']=0.;metrics[prefix+force+'virial_error']=0.
for key,value in metrics.items():measurement(0,source,q,key,value)
csv_file(folder/(c.P+'physical.csv'),physical)
state=self.campaign/'input_state.txt';direction=self.campaign/'input_direction.txt';vec_file(state,x);vec_file(direction,p)
binary=self.campaign/'plateau_diagnostic';binary.write_bytes(b'synthetic executable identity')
mesh=self.campaign/'mesh.smesh';mesh.write_text('synthetic mesh identity')
args=['--coupled-trial-probe','--probe-state',str(state),'--probe-direction',str(direction),'--strategy','coupled','--coupled-seed-control']
cases=[{'id':'synthetic','arguments':args}]
provenance={'commit':'synthetic','source':str(self.campaign),'mesh':str(mesh),'mesh_sha256':c.digest(mesh),
'binary_sha256':c.digest(binary),'cases':cases}
(self.campaign/'provenance.json').write_text(json.dumps(provenance));(self.campaign/'cases.json').write_text(json.dumps(cases))
record={'id':'synthetic','finished_utc':'2026-09-18T00:00:00Z','exit_code':0,
'command':[str(binary),*args,'--mesh',str(mesh),'--output',str(folder)],
'input_state_hashes':{str(state):c.digest(state),str(direction):c.digest(direction)},
'output_hashes':{p.name:c.digest(p) for p in folder.iterdir()}}
(self.campaign/'synthetic_run.json').write_text(json.dumps(record))
def alter(self,name,callback):
path=self.folder/(c.P+name+'.csv');data=c.rows(path);callback(data);csv_file(path,data)
def valid(self,result):
failed=[x for x in result.get('checks',result.get('provenance_checks',[])) if not x['passed']]
failed += [x for case in result.get('cases',{}).values() for x in case['checks'] if not x['passed']]
self.assertTrue(result['passed'],failed)
def test_exact_mixed_correction_and_provenance(self):
self.valid(c.inspect_campaign(self.campaign))
def test_failed_verified_gate_is_valid_evidence_but_not_a_correction_success(self):
self.make_case('failed_gate');r=c.inspect_case(self.folder);self.valid(r)
self.assertFalse(r['trial_outcomes'][0]['aux_verified_gate']);self.assertFalse(r['trial_outcomes'][0]['full_merit_improved'])
def test_cap_rejection_preserves_original_trial(self):
self.make_case('cap');r=c.inspect_case(self.folder);self.valid(r)
self.assertEqual(r['trial_outcomes'][0]['transaction_status'],'norm_limit')
def test_worse_corrected_merit_requires_fallback(self):
self.make_case('worse');r=c.inspect_case(self.folder);self.valid(r)
self.assertEqual(r['trial_outcomes'][0]['transaction_status'],'not_improved')
self.assertFalse(r['trial_outcomes'][0]['selected_armijo'])
def test_kept_correction_does_not_imply_armijo_acceptance(self):
self.make_case('improved_not_accepted');r=c.inspect_case(self.folder);self.valid(r)
self.assertTrue(r['trial_outcomes'][0]['full_merit_improved']);self.assertFalse(r['trial_outcomes'][0]['selected_armijo'])
def test_already_acceptable_trial_needs_no_auxiliary_solve(self):
self.make_case('original_accepted');r=c.inspect_case(self.folder);self.valid(r)
self.assertFalse(r['trial_outcomes'][0]['correction_attempted']);self.assertTrue(r['trial_outcomes'][0]['selected_armijo'])
def test_failed_raw_preparation_has_no_false_correction_certificate(self):
self.make_case('raw_failure');r=c.inspect_case(self.folder);self.valid(r)
self.assertFalse(r['trial_outcomes'][0]['before_usable']);self.assertFalse(r['trial_outcomes'][0]['selected_armijo'])
def test_failed_outer_gate_is_valid_evidence_without_trials(self):
for path in self.folder.glob(c.P+'*_a*.txt'):path.unlink()
for name,values in {'base_action':[-.5]*3,'restored_action':[-.5]*3,'base_defect':[.5]*3}.items():
vec_file(self.folder/(c.P+name+'.txt'),values)
for name in ('trials','blocks'):
path=self.folder/(c.P+name+'.csv');csv_file(path,[],list(c.rows(path)[0]))
updates={'base_merit_slope':-1.5,'outer_verified_defect_norm':math.sqrt(.75),
'outer_verified_relative_defect':.5,'outer_verified_gate':0,'outer_gate_failure_skips_trials':1,
'explicit_probe_preparations':2,'explicit_probe_preparation_seconds':.2}
self.alter('summary',lambda rows:[r.update(value=updates[r['metric']]) for r in rows if r['metric'] in updates])
def preparations(rows):
rows[:]=[r for r in rows if r['kind'] in ('initial_base_restore','final_base_restore')]
for i,row in enumerate(rows):row['sequence']=i+1
self.alter('preparations',preparations)
def physical(rows):
rows[:]=[r for r in rows if r['sample']=='-1']
next(r for r in rows if r['metric']=='selected_snapshots_measured')['value']=0
self.alter('physical',physical)
result=c.inspect_case(self.folder);self.valid(result)
self.assertFalse(result['outer_linear_gate']);self.assertEqual(result['trial_outcomes'],[])
def test_gravity_path_then_full_correction_has_distinct_saved_targets(self):
def save(name,values):vec_file(self.folder/(c.P+name+'_a0.txt'),values)
save('gravity_correction',[2.,0.,0.]);save('before_state',[1.,-1.,-1.]);save('before_residual',[0.,10.,0.])
for name in ('soc_rhs','soc_correction','soc_action'):save(name,[0.,-10.,0.])
for name in ('corrected_state','selected_state'):save(name,[1.,-11.,-1.])
for name in ('corrected_physical','selected_physical'):save(name,[.5,-5.5,-.5])
self.alter('trials',lambda rows:rows[0].update(gravity_path=1,gravity_attempted=1,
gravity_target_defect=0.,gravity_affine_disagreement=0.,gravity_target_threshold=1.e-12))
self.alter('summary',lambda rows:next(r for r in rows if r['metric']=='gravity_path').update(value=1))
meta=self.folder/'metadata.txt';meta.write_text(meta.read_text().replace('strategy=coupled','strategy=curvature-coupled'))
self.alter('layout',lambda rows:[r.update(block={'density':'gravity_potential','closure':'poisson_balance'}.get(r['block'],r['block'])) for r in rows])
def blocks(rows):
for row in rows:
if row['stage']=='before' and row['block']=='closure':row['residual_norm']=0.
if row['stage']=='before' and row['block']=='shape':row['residual_norm']=10.
if row['block']=='closure':row['block']='poisson_balance'
self.alter('blocks',blocks)
self.valid(c.inspect_case(self.folder))
self.alter('trials',lambda rows:rows[0].update(gravity_target_defect=1.e-6))
self.assertFalse(c.inspect_case(self.folder)['passed'])
def test_false_saved_true_gate_cannot_pass(self):
self.make_case('failed_gate');self.alter('trials',lambda rows:rows[0].update(aux_verified_gate=1))
self.assertFalse(c.inspect_case(self.folder)['passed'])
def test_finite_failed_verification_cannot_be_hidden_by_save_flags(self):
self.make_case('failed_gate');self.alter('trials',lambda rows:rows[0].update(verified_vectors_saved=0))
for name in ('soc_action_a0','soc_defect_a0'):(self.folder/(c.P+name+'.txt')).unlink()
self.assertFalse(c.inspect_case(self.folder)['passed'])
def test_looser_reported_controls_cannot_override_requested_controls(self):
self.make_case('failed_gate')
self.alter('summary',lambda rows:next(r for r in rows if r['metric']=='coupled_relative_tolerance').update(value=1.))
self.assertFalse(c.inspect_case(self.folder)['passed'])
def test_conflicting_raw_preparation_is_not_usable_evidence(self):
self.alter('preparations',lambda rows:next(r for r in rows if r['kind']=='raw').update(prepared=0,rejection_reason=2,rejection_stage=6))
self.assertFalse(c.inspect_case(self.folder)['passed'])
def test_physical_replay_is_bound_to_selected_residual(self):
self.alter('physical',lambda rows:next(r for r in rows if r['metric']=='normalized_bordered_residual').update(value=.1))
self.assertFalse(c.inspect_case(self.folder)['passed'])
def test_undefined_density_pressure_is_valid_evidence_not_virial_success(self):
def negative_density(rows):
for row in rows:
if row['sample']=='0' and row['quadrature'] in ('14','18'):
key=row['metric']
if key=='negative_density_samples':row['value']=1
elif key=='minimum_density':row['value']=-.001
elif key=='pressure_from_density_defined_everywhere':row['value']=0
elif key in ('pressure_integral','virial_signed','virial_error','force_virial_signed','force_virial_error'):row['value']=math.nan
self.alter('physical',negative_density);result=c.inspect_case(self.folder);self.valid(result)
self.assertIsNone(result['physical_observations'][0]['virial_status']['virial_within_1e5'])
self.assertTrue(result['physical_observations'][0]['virial_status']['enthalpy_virial_within_1e5'])
self.assertEqual(result['physical_observations'][0]['global_enthalpy_certificate'],'not_measured')
def test_false_physical_virial_cannot_be_declared(self):
self.alter('physical',lambda rows:next(r for r in rows if r['metric']=='enthalpy_pressure_integral').update(value=2.))
self.assertFalse(c.inspect_case(self.folder)['passed'])
def test_missing_physical_restoration_guard_fails(self):
self.alter('physical',lambda rows:rows.__setitem__(slice(None),[r for r in rows if r['metric']!='direction_buffers_unchanged']))
self.assertFalse(c.inspect_case(self.folder)['passed'])
def test_corrupted_corrected_displacement_is_detected(self):
path=self.folder/(c.P+'corrected_state_a0.txt');v=c.vector(path);v[1]+=.1;vec_file(path,v)
self.assertFalse(c.inspect_case(self.folder)['passed'])
def test_rejected_correction_cannot_replace_original_selected_state(self):
self.make_case('worse');path=self.folder/(c.P+'selected_state_a0.txt');vec_file(path,c.vector(self.folder/(c.P+'corrected_state_a0.txt')))
self.assertFalse(c.inspect_case(self.folder)['passed'])
def test_missing_restoration_invariant_is_not_success(self):
self.alter('summary',lambda data:data.__setitem__(slice(None),[r for r in data if r['metric']!='direction_buffers_unchanged_before_restoration']))
self.assertFalse(c.inspect_case(self.folder)['passed'])
def test_false_armijo_acceptance_is_detected(self):
self.make_case('improved_not_accepted');self.alter('trials',lambda rows:rows[0].update(selected_accepted=1,corrected_accepted=1))
self.assertFalse(c.inspect_case(self.folder)['passed'])
def test_unhashed_auxiliary_action_cannot_support_campaign(self):
path=self.campaign/'synthetic_run.json';r=json.loads(path.read_text());del r['output_hashes'][c.P+'soc_action_a0.txt'];path.write_text(json.dumps(r))
self.assertFalse(c.inspect_campaign(self.campaign)['passed'])
def test_incomplete_producer_cannot_support_campaign(self):
path=self.campaign/'synthetic_run.json';r=json.loads(path.read_text());del r['finished_utc'];path.write_text(json.dumps(r))
self.assertFalse(c.inspect_campaign(self.campaign)['passed'])
def test_later_missing_producer_suppresses_earlier_interpretation(self):
path=self.campaign/'provenance.json';provenance=json.loads(path.read_text())
provenance['cases'].append({'id':'unfinished','arguments':provenance['cases'][0]['arguments']})
path.write_text(json.dumps(provenance));(self.campaign/'cases.json').write_text(json.dumps(provenance['cases']))
result=c.inspect_campaign(self.campaign);self.assertFalse(result['passed'])
self.assertFalse(result['cases']['synthetic']['observations_valid'])
self.assertEqual(result['cases']['synthetic']['trial_outcomes'],[])
self.assertEqual(result['cases']['synthetic']['physical_observations'],[])
def test_rehashed_changed_input_direction_is_not_the_measured_direction(self):
path=self.campaign/'input_direction.txt';vec_file(path,[1.,2.,3.]);rp=self.campaign/'synthetic_run.json'
r=json.loads(rp.read_text());r['input_state_hashes'][str(path)]=c.digest(path);rp.write_text(json.dumps(r))
self.assertFalse(c.inspect_campaign(self.campaign)['passed'])
if __name__=='__main__':unittest.main()