Skip to content
Closed
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions research/round6-pr71/.gitignore
Original file line number Diff line number Diff line change
@@ -0,0 +1 @@
baseline/
58 changes: 58 additions & 0 deletions research/round9-balanced-columns/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,58 @@
# Balanced columns with envelope ordering and carried-signal exchanges

The selected conditional witness has κ = **51915547765237/10^18**
(**5.1915547765237e-5**) and bit saving **3244890195589/62500000000000000**.
It combines balanced coarse columns with anchored split frames, envelope
ordering and carried-signal exchanges. Physical width is **133,585,058**,
with auxiliary role counts **27,128 / 35,598**.

Compared with PR #74 at `5830ccfca8466aff613767222558e29b57fa8225`, the
conditional κ improves by approximately **0.163125609609%**. Compared with
PR #76 it improves by **0.3418559067%**. These are comparisons of pinned
conditional claims; no global optimality or unconditional theorem is claimed.

[Proof and attribution](selected-both/PROOF.md) ·
[Exact certificate](selected-both/arithmetic.json) ·
[Source and word manifest](selected-both/MANIFEST.json) ·
[Completed finite validation](package-validation-receipt.json) ·
[Exact comparison](selected-both/comparison.json)

Passed checks include dense scalar/data-output identities, full compiler
input/dirty bases, independent serialized-word replay in both orientations,
exact frame/event reconstruction, regenerated fixed-basis profiles with zero
CRT disagreements, independent rational moments, all 47 strict assembly
constraints and seven margins, and adjacent-grid rejection. Fresh portable
deterministic compilation and broad repository checks remain pending.

The package validation receipt independently reproduces all saved-word, profile
and arithmetic checks against the current manifest. Its mathematical results
match the original receipt exactly. The original validation receipt binds
the original freeze manifest archived
in `provenance/frozen-manifest.json`. Runtime sources, words, profiles and
arithmetic are byte-identical to that freeze; only proof/comparison metadata
has been updated for this publication. The current manifest records those
two documentation changes. No CI configuration or dependencies are changed.

From the repository root, restore the exact inherited source archive:

```sh
git fetch https://github.com/CrocSwap/integer-mult-bounds.git 1bef94fd40a746452548c84a4a8f8834670a3113
mkdir -p research/round6-pr71/baseline
git archive 1bef94fd40a746452548c84a4a8f8834670a3113 | tar -x -C research/round6-pr71/baseline
python3 research/round9-balanced-columns/dense_audit.py
python3 research/round9-balanced-columns/validate_selected.py --bundle research/round9-balanced-columns/selected-both --output /tmp/balanced-columns-validation.json
```

The proof gives sequential compiler regeneration commands. Use assert-enabled
Python and C++17; `CXX` can select a compiler and SDK. All inherited all-size
compiler, fixed-tape, analytic, routing, prime-selection, precision and recovery
hypotheses remain conditional. Finite checks do not establish practical speedup.

Credit eumemic's PR #69 balanced coarse columns; huxint's PR #77 anchored
adapter; Thomas DiFiore's PR #74 envelope ordering; Chafik Boukhalfa's PR #71
anchored splits and next-use scoring; Alejandro Zarzuelo Urdiales's PR #70
exchanges; Rohan Arun's PR #65/#67 schedules, costs and arithmetic; Dominik
Scholz's PR #68/#76 controls and composition; Avi Eisenberg's PR #62,
eumemic's PR #57 and the complete inherited contributor chain. This composition
was prepared by Dominik Scholz with substantial OpenAI GPT-6 Astra assistance.
All original notices and Apache-2.0 terms remain preserved.
14 changes: 14 additions & 0 deletions research/round9-balanced-columns/SOURCE.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,14 @@
{
"pr77_commit": "83298467291d3bd4b53f76faaef60dbf585be171",
"pr76_base_manifest_sha256": "52c1a59a4372376f64df937a2f19268e2ebd553574386fcdda08b01db9a5c345",
"base_pair_graph_sha256": "3d47e89d2649c90800a276bdd0480131def50e0bdf92c122a19494b55bf17420",
"base_pair_graph_matches_pr77_bytes": true,
"inherited_files": {
"inherited/SOURCE.json": "813564d76569fd54a5aa1645da15f1a33b12de8a0bcc53056e69087394b0cfcd",
"inherited/graph.py": "349c74a3684cc26c14f4b7001efb086825a4398461709ba13ade2cd10e336d23",
"inherited/NOTICE": "3bfe7a693e10fbe710e7fadedb7caeb21e272ce5899f48d043f636e2b06562b0",
"inherited/README.md": "e9d7ae955d27fb1832744583f1d251263e9fc499bae2d664b31f28fc83dc2592",
"inherited/vendor/pr69-compiler.py": "bd8169bf9f5df185a8aa4aec22234b4fd483000a1cecc78b71bf32e479af1499"
},
"scope": "Only PR69/77 balanced coarse-column graph mechanism composed with PR76 node ordering + carry exchanges; no PR77 cost-basis compiler changes."
}
25 changes: 25 additions & 0 deletions research/round9-balanced-columns/cache-stale-failure.log
Original file line number Diff line number Diff line change
@@ -0,0 +1,25 @@
Traceback (most recent call last):
File "/Users/dominik/integer-mult-bounds-incremental/research/round9-balanced-columns/run.py", line 102, in <module>
if __name__=='__main__':main()
~~~~^^
File "/Users/dominik/integer-mult-bounds-incremental/research/round9-balanced-columns/run.py", line 93, in main
try:result,word=compiler.compile_(a.h,matching=True,reclaim=True,dirty=True)
~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/dominik/integer-mult-bounds-incremental/research/round6-pr71/baseline/scripts/experiments/split_pair_engine.py", line 120, in compile_
t0=time.time();c,blocks,uses,value_uses,owner,signal,order,contains=build(h);v=len(c.inputs)
~~~~~^^^
File "/Users/dominik/integer-mult-bounds-incremental/research/round6-pr71/baseline/references/frame-compiler/pr65/research/reordered-rank-pair/verify.py", line 34, in build
parts=list(original(h))
~~~~~~~~^^^
File "/Users/dominik/integer-mult-bounds-incremental/research/round6-pr71/baseline/scripts/experiments/split_pair_engine.py", line 39, in build
c=graph(h);groups={};owner={};signal={};blocks=[]
File "/Users/dominik/integer-mult-bounds-incremental/research/round9-balanced-columns/run.py", line 33, in <lambda>
compiler.graph=lambda h: schedule_variant.reorder(coarse_graph.graph(h,coarse='columns'));compiler.build=producer.ordered_build(compiler.build,a.h)
~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/dominik/integer-mult-bounds-incremental/research/round9-balanced-columns/schedule_variant.py", line 34, in reorder
c.verify()
~~~~~~~~^^
File "/Users/dominik/integer-mult-bounds-incremental/research/round6-pr71/baseline/references/frame-compiler/pr48/scripts/partial_swap/shared.py", line 111, in verify
assert not A&B
^^^^^^^
AssertionError
68 changes: 68 additions & 0 deletions research/round9-balanced-columns/coarse_graph.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,68 @@
"""Controlled combinations of the PR62, PR69 and PR71 scalar graphs.

PR62 graph: Avi Eisenberg with Claude assistance. PR59 split operation:
Rohan Garg. PR69 balanced coarse columns: eumemic with Codex assistance.
PR71 anchored split vector: Chafik Boukhalfa with Codex assistance.
This configurable composition/search adapter: huxint with Codex assistance.
Apache-2.0 and all inherited notices apply. No global search claim.
"""
import sys
if sys.flags.optimize:
raise ValueError('Assertions must remain enabled')
from hashlib import sha256
from pathlib import Path

ROOT = Path(__file__).resolve().parent.parent/'round6-pr71/baseline/references/frame-compiler/pr65'
BASE = ROOT/'research/pair-assembly/pair_graph.py'
BEFORE = ''' coarse = {(i, j): self.total([e(a, b) for a in groups[i] for b in groups[j]])
for i, j in combinations(range(ng), 2)}'''
AFTER_COLUMNS = ''' def coarse_sum(i, j):
columns = [self.total([e(a,b) for a in groups[i]]) for b in groups[j]]
return self.total(columns)
coarse = {(i,j):coarse_sum(i,j) for i,j in combinations(range(ng),2)}'''
AFTER_ROWS = ''' def coarse_sum(i, j):
rows = [self.total([e(a,b) for b in groups[j]]) for a in groups[i]]
return self.total(rows)
coarse = {(i,j):coarse_sum(i,j) for i,j in combinations(range(ng),2)}'''


def graph(h, groups=(1, 1, 2), anchor='first', coarse='none'):
if h not in (23, 25) or any(type(x) is not int or not 0 <= x <= 3 for x in groups):
raise ValueError('Unsupported finite graph configuration')
source = BASE.read_text()
if coarse != 'none':
assert source.count(BEFORE) == 1
source = source.replace(BEFORE, {'columns': AFTER_COLUMNS, 'rows': AFTER_ROWS}[coarse])
namespace = {'__file__': str(BASE), '__name__': 'next_use_graph_composition'}
exec(compile(source, str(BASE)+':next_use_composition', 'exec'), namespace)
original_class = namespace['circuit_class']

def circuit_class(dimension):
class SplitPair(original_class(dimension)):
def grouping(self, points):
partition = super().grouping(points)
choice = groups[self._level] if self._level < len(groups) else 0
if len(points) > 3 and choice:
index = {1: 0, 2: (len(partition)-1)//2, 3: len(partition)-1}[choice]
if len(partition[index]) == 2:
partition[index:index+1] = [[x] for x in partition[index]]
assert all(len(g) in (1, 2) for g in partition)
assert len(partition) < len(points)
assert sorted(x for g in partition for x in g) == sorted(points)
return partition
return SplitPair

namespace['circuit_class'] = circuit_class
original_points = namespace['alternating_points']
if anchor != 'none':
def points(dimension, common):
old = original_points(dimension, common)
pairs = [(a, a+1) for a in range(0, dimension-1, 2) if common not in (a, a+1)]
chosen = {'first': pairs[0], 'second': pairs[min(1, len(pairs)-1)], 'last': pairs[-1]}[anchor]
result = list(chosen)+[x for x in old if x not in chosen]
assert sorted(result) == [x for x in range(dimension) if x != common]
return result
namespace['alternating_points'] = points
result = namespace['graph'](h)
result.verify()
return result
28 changes: 28 additions & 0 deletions research/round9-balanced-columns/dense-audit.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,28 @@
{
"scope": "Finite dense scalar/data-output identity audit",
"axes": {
"23": {
"h": 23,
"inputs": 1771,
"additions": 67460,
"outputs": 5336,
"all_dense_additions_disjoint": true,
"all_dense_partial_outputs_exact": true,
"original_common_point_envelopes_exact": true
},
"25": {
"h": 25,
"inputs": 2300,
"additions": 94993,
"outputs": 6925,
"all_dense_additions_disjoint": true,
"all_dense_partial_outputs_exact": true,
"original_common_point_envelopes_exact": true
}
},
"sources": {
"dense_audit.py": "b282e743d9d8c48cbc97eb2c2597a6c52894d33ea1967dd1ff4c269b0e0a72e3",
"coarse_graph.py": "b5bbabb8e5bf69769af4b6cd67163954d990dc489e3c9488ed84bd09f0a767c8",
"schedule_variant.py": "3ce3b4c0886868e3fcd2b75c54b61180d103447cfdb9d3ece1de47dd6d46eecc"
}
}
53 changes: 53 additions & 0 deletions research/round9-balanced-columns/dense_audit.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,53 @@
"""Independent dense supports for both balanced-column scalar graphs.

Dominik Scholz with substantial OpenAI GPT-6 Astra assistance. Apache-2.0.
Does not use the graph's memoized/provenance-based support routine.
"""
from hashlib import sha256
from pathlib import Path
import json
import run


def audit(h):
c = run.schedule_variant.reorder(run.coarse_graph.graph(h, coarse='columns'))
support = [0] * len(c.args)
point_rows = [0] * h
for i, triple in enumerate(c.inputs):
support[i + 1] = 1 << i
for point in triple:
point_rows[point] |= 1 << i
additions = 0
for node in sorted(c.active):
if not c.args[node]:
assert node <= len(c.inputs)
continue
a, b = c.args[node]
assert a < node and b < node
assert support[a] and support[b]
assert not support[a] & support[b]
support[node] = support[a] | support[b]
assert c.core[node] == c.core[a] & c.core[b]
assert c.union[node] == c.union[a] | c.union[b]
additions += 1
for (common, target), node in c.outputs.items():
expected = point_rows[common]
for point in target:
if point != common:
expected &= ~point_rows[point]
assert support[node] == expected
return dict(h=h, inputs=len(c.inputs), additions=additions,
outputs=len(c.outputs), all_dense_additions_disjoint=True,
all_dense_partial_outputs_exact=True,
original_common_point_envelopes_exact=True)


if __name__ == '__main__':
root = Path(__file__).resolve().parent
result = {'scope': 'Finite dense scalar/data-output identity audit',
'axes': {str(h): audit(h) for h in (23, 25)},
'sources': {name: sha256((root/name).read_bytes()).hexdigest()
for name in ('dense_audit.py', 'coarse_graph.py',
'schedule_variant.py')}}
(root/'dense-audit.json').write_text(json.dumps(result, indent=2) + '\n')
print(json.dumps(result, indent=2))
28 changes: 28 additions & 0 deletions research/round9-balanced-columns/inherited/NOTICE
Original file line number Diff line number Diff line change
@@ -0,0 +1,28 @@
Balanced coarse sums on anchored split frames
Prepared for huxint with substantial OpenAI Codex assistance, October 9, 2026.
Distributed under the repository's Apache License 2.0.

The selected finite network composes previously attributed constructions.
Preserve all repository notices and these source-specific attributions:

eumemic: PR57 joint reversible frame compiler and independent physical replay;
PR69 balanced coarse column sums, with OpenAI Codex assistance.
Avi Eisenberg / ikeboy: PR62 interval strips and core-aware pair assembly,
with Anthropic Claude assistance.
Rohan Garg / rohangar1: PR59 split-pair operation, with OpenAI Codex assistance.
Rohan Arun: PR65 schedule/refined arithmetic and PR67 profile-cost compiler,
with OpenAI Codex assistance.
Dominik Scholz: PR63 ranked composition and PR68 compatible pending controls,
with OpenAI GPT-6 Astra / Codex assistance.
Chafik Boukhalfa: PR60 ranked reclamation and PR71 anchored split graph and
next-use scoring, with OpenAI Codex assistance.

The original code comments and preserved vendor snapshots retain their own
attributions. The inherited framework also credits icekylinx, RaD / hipotures,
James Chang, Zhihao Chen / jacklightChen, Aurel Prosz / Paureel, Swapnil Jain,
Douglas Colkitt, OpenAI, David Harvey, Joris van der Hoeven, and other
contributors identified in the root NOTICE and inherited source files.

The new package's finite checks are conditional evidence. They are not a
complete theorem formalization, an independent human referee report, or a
practical multiplication benchmark.
90 changes: 90 additions & 0 deletions research/round9-balanced-columns/inherited/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,90 @@
# Balanced coarse sums on anchored split frames

The selected finite construction gives the conditional integer-multiplication
bound

\[
T(n)=O\bigl(n(\log n)^{1-\kappa}\bigr),\qquad
\kappa=\frac{12886963972497}{250000000000000000}
=5.1547855889988\times10^{-5}.
\]

This is **0.2590891819% above PR71** at commit
`1bef94fd40a746452548c84a4a8f8834670a3113`. It is a comparison of conditional
asymptotic exponent bounds, with no global-optimality or measured-runtime claim.

The construction combines eumemic's PR69 balanced coarse columns with Chafik
Boukhalfa's PR71 anchored `[1,1,2]` split graph. It uses the inherited joint
reversible frame compiler and compatible pending controls, and orders both
eligible live controls and retired dependency candidates by their actual
profile cost. The latter changes the temporary dependency basis, not just the
choice among the first basis's dependent rows. Pending controls are priced
against their known next use. Every selected clearing expression is emitted
as literal XORs, and all frame moves and final cleanup remain charged.

| Quantity | Pinned PR71 | This construction |
|---|---:|---:|
| Roles at h=23 | 27,455 | 27,338 |
| Roles at h=25 | 36,015 | 35,939 |
| Physical width W | 135,075,665 | 134,671,969 |
| Complete rank mass | 77,666,660,475 | 77,434,535,275 |
| Rank deficit | 1,846,900 | 1,846,900 |

The strict bit-saving witness is `51550513208569/10^18`. The full paid child
list of PR71 has a rigorous moment lower bound above one at this value. The
new list has a rigorous upper bound below one. Its next `10^-18` bit grid
point is rejected, as is the next kappa grid point in the unchanged balanced
assembly. All 47 strict inequalities and seven margins are checked exactly.

## Reproduction

From the repository root, with Python 3.11+ and a C++17 compiler:

```sh
make -C research/balanced-split-reclamation verify
make -j1 verify
```

The first command regenerates both selected words byte-for-byte, independently
replays every input/target/dirty basis vector in both orientations, derives
all physical transitions from literal XOR incidences, recomputes every
fixed-basis profile with the inherited bounded-minor/CRT checks, rebuilds
the complete global profile and exact arithmetic, and runs twelve failure-mode
tests. A second logarithm/exponential evaluation uses a different range
reduction and Taylor degree. This is a second arithmetic implementation,
not independent expert review of the theorem.

`selected/` contains the words, actual profiles, receipts and exact certificate.
`SOURCE.json` freezes consumed sources and deliverables; verification never
rewrites this manifest. `pin.py` is an explicit maintenance operation.
`generate.py --output <directory>` produces fresh artifacts without modifying
the selected witness. The proof is in [paper.tex](paper.tex) and [paper.pdf](paper.pdf).

The selected graph uses columns rather than a three-addition coarse chain,
split vector `[1,1,2]`, and the lowest intact pair first in each common-point
ordering. The h23 region order is `cover-core`; h25 uses `reverse-node`.
The oracle's integer entropy score is only a deterministic search heuristic.
The final complete physical profile, not the score, determines acceptance.

## Proof boundary and attribution

The original-envelope interpretation, all-size residual/frame compiler,
fixed finite-alphabet tape layout, semantic and analytic multiplication
reduction, routing, prime setup and exact recovery remain inherited
assumptions. The complex branch is unchanged. Exact finite checks do not
constitute a complete formal proof of the multiplication theorem.

This composition, cost-ordered dependency basis experiment, independent
reproduction and package were prepared for **huxint with OpenAI Codex
assistance**. Credit eumemic for PR57's reversible region compiler and PR69's
coarse-column identity; Avi Eisenberg for PR62's interval/core-aware graph;
Rohan Garg for PR59's split operation; Rohan Arun for PR65's schedules and
PR67's cost oracle and numerical refinement; Dominik Scholz for PR63's
composition and PR68's compatible pending controls; and Chafik Boukhalfa for
PR60's ranked reclamation and PR71's anchored split/next-use composition.
All prior contributors retain their original notices. See [NOTICE](NOTICE).

The preserved contributor files in `vendor/` are unmodified snapshots, used
for provenance and the complete PR71 comparison. The inherited repository
files are retained unchanged. No endorsement or completed expert review by
those contributors is implied.
Loading