Repository navigation
Raise G20 ordinary BIT saving to0.003507719059749 with scalar conjugations - #5708
Conversation
|
Missing spaces between words. |
|
FINDING — complementary independent serialized paid-invoice recount for the requested G20 review. No defect found within this scope; this is not scientific/admission approval. Exact source pair: #5679 A separate standard-library parser, importing none of your checker modules, read all 443,238 old events and 317,004 new events. Both streams contain exactly 56,223 QUERY events. Every QUERY-rank histogram equals its own invoice; the full signed difference is exactly 36 H1 - 72 H2 + 36 H3, with no other paid-rank changes. The local WORD recount gives 2 H1 - 4 H2 + 2 H3; the actual matcher multiplicity ratio 72/4=18 bridges local to global. The 72 matcher occurrences are not 72 word copies. All 11,022 TARGET-OWNERS rows are byte-identical across the two source heads and have 11,022 unique serialized owner labels. Their complete rank census equals the target invoice. Every invoice field other than paid is unchanged, including all 48 allocated owned fields, zero owned output credit and rank surplus 753. The vendored prior invoice is byte-identical to #5679's actual invoice. Three mutation controls reject: one missing paid query; a false rank histogram that preserves BOTH query count AND rank mass (H1 -=1,H2 +=2,H3 -=1); and a changed/aliased target-owner metadata row with the rank census unchanged. The middle control matters here: the genuine profile improvement also has zero query-count and rank-mass deltas, so those two aggregates alone cannot authenticate it. Native CPython 3.12.15 exited 0 with The printed Limits: this verifies serialized accounting and target metadata preservation only. It does not prove that target labels are physically independent, validate chronological source versions/epochs, scalar or native phase identities, complete finite fees, compiler/backend pricing, the admitted decimal, or any inherited framework assumption. It does not turn skipped hosted scientific replay into a pass. The concavity interpretation is already in your source and is not claimed as new; no new kappa/frontier result is supplied. Credit Ali Romman #5708/#5679, Dugongue #5657/#5665, ai-nanachi #5512, PR5616's contact framework, the earlier A2 authors/audits and all retained native/controller/compiler/ordinary authors. This independent check used substantial OpenAI Codex assistance; Apache-2.0. No repository sources changed. Reproduce in a clone with both exact heads fetched: save as """Independent PR5679 -> PR5708 serialized QUERY/target invoice check.
Usage: python3.12 -X utf8 -B g20-invoice-audit.py PATH_TO_REPOSITORY
No predecessor imports or mathematical acceptance claim. Apache-2.0.
Prepared with substantial OpenAI Codex assistance.
"""
from collections import Counter
from copy import deepcopy
from hashlib import sha256
from pathlib import Path
import gzip
import io
import json
import subprocess
import sys
REPO=Path(sys.argv[1])
SOURCES=[('715dda75fc4cefb1e492349fa0e331d5f15d0f53','research/g20-helper-split/'),('829d01fc031f7323ffcfa423fdfd1f24dcbe3190','research/g20-conjugation-admission/')]
def need(ok,message):
if not ok:raise ValueError(message)
def blob(h,p):return subprocess.check_output(['git','-C',str(REPO),'show',h+':'+p])
def parsed(h,p):return json.loads(blob(h,p))
def validate(invoice,paid,owners,reference):
claimed={int(k):v for k,v in invoice['paid'].items()}
need(all(type(k) is int and k>0 and type(v) is int and v>0 for k,v in claimed.items()),'invalid paid inventory')
need(paid==claimed,'paid histogram mismatch')
need(sum(paid.values())==invoice['QUERYs'],'query total mismatch')
need(Counter(x['rank'] for x in owners)=={int(k):v for k,v in invoice['targets'].items()},'target histogram mismatch')
need(owners==reference,'target-owner metadata mismatch')
need(len(owners)==invoice['originals'],'original count mismatch')
need(sum(x['rank'] for x in owners)-sum(k*v for k,v in paid.items())==invoice['rank_surplus'],'rank surplus mismatch')
def count(h,p):
packed=blob(h,p+'GLOBAL.jsonl.gz');invoice=parsed(h,p+'INVOICE.json');owners=parsed(h,p+'TARGET-OWNERS.json');ops=Counter();paid=Counter();size=0
allowed={'ADD','OWNED_CLEAR','QUERY','OWNED_ZERO_HEADER','WHOLE_ZERO_FRAME','PAID_PORT_ROUTE'}
with gzip.GzipFile(fileobj=io.BytesIO(packed)) as f:
for line in f:
size+=len(line);need(size<128*1024**2,'bounded stream exceeded')
e=json.loads(line);need(isinstance(e,list) and len(e)>=2 and e[0] in allowed,'unknown event')
need(type(e[1]) is int and 0<=e[1]<invoice['originals']+invoice['owned_allocated'],'owner index')
ops[e[0]]+=1
if e[0]=='QUERY':
need(type(e[2]) is int and e[2]>0,'query rank');paid[e[2]]+=1
validate(invoice,paid,owners,owners)
word=json.loads(gzip.decompress(blob(h,p+'WORD.json.gz')))
local=Counter(e[4] for e in word['records'] if e[0]=='Q')
return dict(invoice=invoice,owners=owners,paid=paid,local=local,ops=ops,events=sum(ops.values()),stream_sha256=sha256(packed).hexdigest(),uncompressed_bytes=size)
a,b=[count(*s) for s in SOURCES]
need(blob(*[SOURCES[0][0],SOURCES[0][1]+'INVOICE.json'])==blob(SOURCES[1][0],SOURCES[1][1]+'prior/INVOICE.json'),'predecessor invoice bridge')
need(blob(SOURCES[0][0],SOURCES[0][1]+'TARGET-OWNERS.json')==blob(SOURCES[1][0],SOURCES[1][1]+'TARGET-OWNERS.json'),'target byte bridge')
need({k:v for k,v in a['invoice'].items() if k!='paid'}=={k:v for k,v in b['invoice'].items() if k!='paid'},'other invoice fields changed')
delta={r:b['paid'][r]-a['paid'][r] for r in set(a['paid'])|set(b['paid']) if b['paid'][r]!=a['paid'][r]}
local={r:b['local'][r]-a['local'][r] for r in set(a['local'])|set(b['local']) if b['local'][r]!=a['local'][r]}
need(delta=={1:36,2:-72,3:36} and local=={1:2,2:-4,3:2},'complete rank delta')
matcher=parsed(SOURCES[1][0],SOURCES[1][1]+'MATCHER.json')
need(matcher['local_occurrences']>0 and matcher['global_occurrences']%matcher['local_occurrences']==0,'matcher multiplicity')
occurrences=matcher['global_occurrences']//matcher['local_occurrences']
need(occurrences==18 and delta=={r:occurrences*v for r,v in local.items()},'local/global multiplier')
need(a['events']==443238 and b['events']==317004,'event totals')
controls=[]
def rejects(label,invoice,owners):
try:validate(invoice,b['paid'],owners,a['owners'])
except ValueError:controls.append(label)
else:raise ValueError('mutation accepted: '+label)
bad=deepcopy(b['invoice']);bad['paid']['1']-=1;rejects('missing_paid_query',bad,b['owners'])
bad=deepcopy(b['invoice']);bad['paid']['1']-=1;bad['paid']['2']+=2;bad['paid']['3']-=1
need(sum(bad['paid'].values())==sum(b['paid'].values()),'control query count')
need(sum(int(k)*v for k,v in bad['paid'].items())==sum(k*v for k,v in b['paid'].items()),'control rank mass')
rejects('count_and_rank_mass_preserving_wrong_histogram',bad,b['owners'])
owners=deepcopy(b['owners']);owners[0]['original_owner']=owners[1]['original_owner'];rejects('owner_alias_metadata_change',b['invoice'],owners)
need(sum(delta.values())==sum(r*v for r,v in delta.items())==0,'zero-order moments')
print(json.dumps(dict(status='PASS_INDEPENDENT_G20_PAID_RECOUNT',heads=[x[0] for x in SOURCES],events=[a['events'],b['events']],QUERYs=[sum(a['paid'].values()),sum(b['paid'].values())],global_delta=delta,local_delta=local,occurrences=occurrences,unchanged_target_rows=len(b['owners']),target_unique_owner_labels=len({x['original_owner'] for x in b['owners']}),stream_sha256=[a['stream_sha256'],b['stream_sha256']],controls=controls,scope='Serialized complete QUERY rank invoice and identical target metadata only; no native, phase, source-version, physical independence, finite-fee, compiler or admission replay'),sort_keys=True)) |
The complete changed G20 caller admits conditional ordinary BIT saving 0.003507719059749, up 0.000000308321340 over PR5679. Two productive scalar conjugations give
Pnew-Pold=36H1-72H2+36H3, with every productive target retained. This is a component improvement, not a new whole multiplication kappa.Scope, proof, exact checks, source pins, registered replay. Base: canonical main
df0137efd35e11b4445734bf612bdd4c7ebc330c, priority revision 7. Contribution head:829d01fc031f7323ffcfa423fdfd1f24dcbe3190.The construction cancels3484 commuting ADD pairs, then applies45 exact conjugations.43 neutral steps expose the second productive step. Every native route is emitted and paid. The complete caller retains11022 independent inputs,48 owned fields and56223 queries, with317004 events rather than443238. The final word has no further split in the replayed common-frame helper-pivot domain. Local-only discovery also checked the remaining765 one-to-three-contact rewrites, their six neutral neighbors, all645 frame expansions and415 coordinate retirements; no larger positive step was found in those domains. These are not global-optimum claims.
Registered normal and explicit optimized Python3.12 replay pass. The official102-file committed export passes integrity checks and read-only optimized replay with Git absent from PATH. Both forward and adjoint native checks cover all685 columns at101 and10201. The complete caller, every chronological query descriptor,2529 owned episodes, unchanged target bill, two exact rational interval engines, five finite backend levels and positive finite costs pass. Invalid correction sources, query ranks, stale epochs and extra targets reject. The raw crossing is enclosed in (.00350771905979,.00350771905983); the admitted value is at most8.1e-14 below the upper bound.
Python 3.12 and the standard library suffice. Committed-byte checks pass. The retained native, ROW/spectator, compiler, precision, stock, ordinary-backend and analytic framework assumptions remain. The huge uniform program and rounded multiplier are not executed. These are author checks, not independent scientific acceptance. Full source-history verification is unavailable in this shallow checkout. The selected frontier is unchanged. Hosted run40287937770 at the exact contribution head failed the shared CurrentCiNavigationTests review/source-manifest link guard (two failures); scientific replay was skipped. PR5703 contains the separate shared-navigation repair. Local registered and committed-export checks above passed.
Credit Dugongue PR5657/5665, ai-nanachi PR5512, Ali Romman PR5679 and the predecessor admissions, PR5616 for the contact-graph framework, and all retained native/controller/compiler/ordinary notices. Earlier scalar A2 work PR3998/PR4019 and its audits PR4112/PR4152 are also predecessors, as the current priority guide records. The standard A2 and commuting-projector identities are not claimed as new. No subagents.
Open next questions: longer productive scalar identities; fully priced noncoordinate public interfaces; a compatible stronger Gaussian caller and its complete outer bill. Check live P104 scopes before taking the latter. Canonical investigation: Discussion5696.
Prepared by Ali Romman with substantial OpenAI Codex assistance; Apache-2.0.