File size: 6,719 Bytes
2955ecc | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 | """Evidence-weight model (draft14 maths) evaluated on the v0.3 sweep fixtures."""
import json,math,glob,os,collections,datetime as dt
T=1.0e25; MARGIN=0.10; ETA=0.40; WSC=5.11; L0=-2.0
wF,wC,wN=2.5,2.2,0.85; BW,BM,BH=1.0,2.0,3.0
SEV={'info':.05,'low':.15,'medium':.35,'high':.60,'critical':.85}
psi=lambda z:2*z-1
g=lambda z,t:min(1.0,max(0.0,z/t)) if t>0 else 0.0
sig=lambda x:2**x/(1+2**x)
def num(v,d=0.0):
try:
f=float(v); return d if f!=f else f
except (TypeError,ValueError): return d
def secs(w):
try:
a=dt.datetime.strptime(w['start'],'%Y-%m-%dT%H:%M:%SZ');b=dt.datetime.strptime(w['end'],'%Y-%m-%dT%H:%M:%SZ')
return (b-a).total_seconds()
except Exception: return 0.0
def evaluate(site):
s=site.get('normalized_signals',{}) or {}; cv=site.get('coverage',{}) or {}
raw=site.get('raw_features',{}) or {}
C=lambda k: max(0.0,min(1.0,num(cv.get(k),0.0))) # absent -> 0
dur=secs(site.get('audit_window',{}) or {})
cnt=sum(num(r.get('count')) for r in raw.get('accelerator_count_by_family_sku',[]) or [])
rate=max([num(r.get('peak_rate')) for r in raw.get('advertised_peak_rate_by_precision',[]) or []] or [0.0])
adj=num(s.get('capacity_adjustment_factor'),1.0)
B=cnt*rate*dur*adj
hidden=bool(s.get('hidden_or_unmonitored_capacity_possible',False))
units=bool(s.get('capacity_unit_normalized',True))
sigma_scope=0.0 if hidden else 1.0
cstar=C('capacity')*sigma_scope*C('clock_alignment')
admissible = units and rate>0 and cnt>0 and dur>0 and adj>=0 and sigma_scope>0
mu = B/(cstar*T) if cstar>0 else float('inf')
ruled = admissible and mu < 1/(1+MARGIN)
# achieved
O=num(s.get('achieved_operations'),0.0)
o_norm=bool(s.get('achieved_operations_unit_normalized',True))
A=max(0.0,min(1.0,num(s.get('activity_score'))))
if ruled: pi_sc=0.0
elif O>=T and o_norm and C('achieved_ops')>=0.75: pi_sc=1.0
else:
Ch=max(ETA*A*B, O)
r=-3.0 if Ch<=0 else max(-3.0,min(3.0,math.log10(Ch/T)))
pi_sc=sig(WSC*r)
# ---- witness coherence chi
wit=[]
if O>0 and o_norm and B>0: wit.append(B*(1+MARGIN)/O)
pc=num(s.get('participant_count'))
if pc>0 and cnt>0: wit.append(cnt*(1+MARGIN)/pc)
for rec in raw.get('allocated_accelerator_count_by_sku',[]) or []:
v=num(rec.get('count'))
if v>0 and cnt>0: wit.append(cnt*(1+MARGIN)/v)
chi=min(wit) if wit else None
chi_fail=(chi is not None and chi<1.0)
# ---- stage C explanations -> b
serv=num(s.get('serving_counterevidence_score')); servov=num(s.get('serving_activity_overlap_fraction'))
e_serv=g(serv,.70)*g(servov,.50)*C('serving')
stov=num(s.get('storage_operation_overlap_fraction')); stby=num(s.get('bytes_explained_fraction'))
e_st=g(stov,.80)*g(stby,.70)*C('storage_operations')
breg=num(s.get('benchmark_regularity_score')); bdur=num(s.get('benchmark_duration_seconds'),dur)
e_bm=g(breg,.90)*(1.0 if 0<bdur<=7200 else 0.0)*C('benchmark_hpc')
hpc=num(s.get('hpc_mpi_score')); hov=num(s.get('hpc_overlap_fraction'))
e_hpc=g(hpc,.60)*g(A,.50)*g(hov,.50)*C('benchmark_hpc')
bF=min(1.0,max(e_serv,e_bm,e_hpc)); bC=min(1.0,e_st)
# ---- stage B
F=num(s.get('collective_cadence_score')); rAF=num(s.get('activity_fabric_overlap_fraction'))
adur=num(s.get('activity_duration_seconds'),dur)
Ck=num(s.get('checkpoint_periodicity_score')); rAC=num(s.get('checkpoint_activity_adjacency_fraction'))
nb=num(s.get('checkpoint_burst_count')); N=num(s.get('non_serving_score'))
aF=C('fabric')*g(A,.55)*g(rAF,.50)*g(adur,1800)
aC=C('storage')*g(A,.50)*g(rAC,.50)*g(nb,2)
aN=C('serving')*g(C('serving'),.80)*g(A,.55)
LF=aF*(1-bF)*wF*psi(F); LC=aC*(1-bC)*wC*psi(Ck); LN=aN*wN*psi(N)
Lam=LF+LC+LN; Lcov=C('fabric')*wF+C('storage')*wC+C('serving')*wN
# ---- stage C discrepancies -> D
D=1.0
ratio=(O/B) if B>0 else 0.0
disc=[]
if o_norm and ratio>1.10 and C('capacity')>=0.75 and num(s.get('unit_mismatch_or_hidden_capacity_explanation_score'))<0.70:
disc.append((min(1.0,(ratio-1.10)/1.10),C('capacity'),'critical'))
if A>=0.70 and adur>=600 and num(s.get('attribution_overlap_fraction'),1.0)<=0.05 and C('attribution')>=0.80 \
and num(s.get('benign_attribution_explanation_overlap_fraction'))<0.80:
disc.append((1.0,C('attribution'),'high'))
for k in ('physical_timeline_conflict','health_throttle_conflict','topology_route_conflict','power_activity_conflict'):
if bool(s.get(k)): disc.append((1.0,1.0,'high'))
for t_r,c_r,sev in disc: D*= (1-t_r*c_r*SEV[sev])
D=1-D
pi_ref=(0.0 if ruled else sig(WSC*max(-3,min(3,math.log10(max(ETA*B,1e-30)/T)))))*sig(L0)
pi=(1-D)*pi_sc*sig(L0+Lam)+D*pi_ref
# ---- route
if not admissible or Lcov<1.0 and not ruled and B>0 and Lcov<1.0 and mu>=1/(1+MARGIN) and Lcov<1.0:
pass
if bool(s.get('decision_blocking_missingness')) or not admissible: route='inconclusive_due_to_missingness'
elif chi_fail or D>=0.5: route='integrity_review_required'
elif ruled: route='capacity_ruled_out_for_scope'
elif Lcov<1.0: route='inconclusive_due_to_missingness'
elif Lam>=BH: route='high_training_like_warning'
elif Lam>=BM: route='medium_training_like_warning'
elif Lam>=BW: route='weak_training_like_candidate'
elif max(bF,bC)>0.5: route='candidate_explained_or_demoted'
elif pi_sc>=0.5: route='weak_training_like_candidate'
else: route='no_training_like_candidate_detected_in_covered_live_segment'
return dict(route=route,chi=chi,Lam=Lam,Lcov=Lcov,mu=mu,pi=pi,pi_sc=pi_sc,D=D)
WARN={'high_training_like_warning','medium_training_like_warning'}
CAUGHT=WARN|{'integrity_review_required'}
if __name__=='__main__':
base='xx_claude/results/sweep_v0_3_20260613_01'
out={}
for d in sorted(os.listdir(base)):
p=os.path.join(base,d,'sites_all.jsonl')
if not os.path.exists(p): continue
rc=collections.Counter(); lam=[]; n=0
for line in open(p):
rec=json.loads(line); site=rec.get('site',rec)
r=evaluate(site); rc[r['route']]+=1; lam.append(r['Lam']); n+=1
out[d]=dict(n=n,routes=dict(rc),
warn=sum(rc[x] for x in WARN)/n, caught=sum(rc[x] for x in CAUGHT)/n,
mean_lambda=sum(lam)/n)
print(f"{d:45s} n={n:5d} warn={out[d]['warn']*100:6.2f}% caught={out[d]['caught']*100:6.2f}% meanL={out[d]['mean_lambda']:+.3f}")
json.dump(out,open('/private/tmp/claude-501/-Users-idacy-Develop-datacenter-verification/23d7cbd6-c237-44fb-bca4-24d8ac2a4692/scratchpad/newmodel_results.json','w'),indent=1)
print("\ntotal:",sum(v['n'] for v in out.values()))
|