Skip to content
Merged
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
36 changes: 36 additions & 0 deletions research/p112-mixed-support38/CHECKS.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,36 @@
{
"status": "local registered replay, committed export and designated peer check passed",
"utc": "2026-10-17T20:01:52.057442+00:00",
"baseline": "463113d675af7410ac91d53a4fed51ec6494eb9a",
"environment": "CPython3.12.15 macOS arm64; standard library",
"registered": "PASS all three commands with -W error; controls under -O",
"primary": "3780orthogonalplanes,896live-axis fibers times10795sixspaces; A2 and chamber receipts match RESULT.json",
"independent": "788035ambient sixspaces,16support20sets,120fourframes,1920intersections and3780orthogonal-pair orbit",
"controls": "7groups passed; includes malformed inventory, corrupted source, symlink, oversized file and absent-puncture control",
"diagnostics": "Ruff check and format check passed",
"limits": [
"Global g(5)=12/g(6)=20 historical censuses inherited, not rerun",
"Weyl-chamber theorem is a written dependency",
"Independent checker shares authenticated geometry; no independent A2/global-capacity proof",
"No Lean, Docker, full selected cold replay, supplier or kappa claim"
],
"peer_review": {
"reviewer_worker": 1,
"reviewed_commit": "fb75f16629b3d664488e9f9a186c9933b0901235",
"result": "No blocking issue at stated characteristic-zero commoncut38 ranks5,5,6 scope after singleton-domain clarification.",
"resolved": "Singleton sites/J_i explicitly restricted to U outsideS in PROOF.md; proportional recheck accepted SHA25674b3e610a69a2f5e372595efbc17f3215c3f1af53a0f536dbbc12e5abb821327.",
"receipt": "REVIEW.json"
},
"committed_export": {
"commit": "fb75f16629b3d664488e9f9a186c9933b0901235",
"file_count": 18,
"sha256": "1b6780a64e025756c347930849aa2eb25bd5f1ca59cc30e520d1099df2629f30",
"result": "All3directcommands passed optimized Python3.13.16 in exact committed read-only Git-free export. Final accepted proof-only correction and review-metadata update do not change executables/sources/RESULT."
},
"submission_refresh": {
"canonical_main": "463113d675af7410ac91d53a4fed51ec6494eb9a",
"priority_revision": 6,
"cutoff": "2026-10-17T20:03:02Z",
"coverage": "Complete311openPRs/2101discussionrecords,no warnings. Worker2 near-synonym duplicate released. No novel-coverage inference from no-match."
}
}
60 changes: 60 additions & 0 deletions research/p112-mixed-support38/ENGINEERING.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,60 @@
# Verification design

The executable claim is finite geometry supporting the written mixed-rank
support 38 theorem. No supplier implementation is changed.

Separate Before Serializing Shared State shaped the workspace boundary. This
package, branch, status file and session cache have one writer. Shared Git
metadata operations use the batch's OS-released lock. Research arithmetic runs
without that lock. Peer checks read frozen source and write only reviewer-owned
receipts.

Prove It Works shaped the publication gate. Run the actual registered commands,
then replay exact committed export bytes from a read-only Git-free tree under
a second Python version before freezing the peer-review tree.

## Power of Ten

1. The three scripts use structured loops and explicit ValueError exits. There
is no recursion. File contexts and temporary-directory contexts own cleanup.
2. Root inventories have 120 full/112 live entries. Binary inputs have at most 512
labels of at most 18 bits. D6 has 3780 planes,448 minimum planes and 896 live-axis
fibers. A fiber has 256 functionals and 10795 dual two-spaces. The alternate
ambient census has 788035 three-dimensional annihilators. A2 has 1120 planes
and its orbit queue can contain at most 1120 items. The alternate D6 orbit
queue has at most 3780 items. Source inventories have 5 rows and 4 retained rows.
Registered command timeouts are 120 seconds, sequentially.
3. CPython requires dynamic allocation. This managed-runtime deviation covers
only the bounded inventories above. Source files are at most 1 MB each; the
largest retained primary census serialization has 896 rows with bounded
histograms. Bit masks have at most 33 bits for D6 fibers. The alternate full
inventory retains only 16 maximal supports. No asynchronous work or subprocess
remains after a command. Direct allocation instrumentation is unrun.
4. The largest new function is a 2_receipt at 60 physical lines after Ruff format.
Other functions are below 60 lines. It owns one plane classification and
orbit/fiber receipt; splitting it further would add a single-use forwarding
boundary without a separate ownership requirement.
5. Runtime require checks validate source pins, exact cardinalities, label
ranges, census totals, orbit coverage and arithmetic implications. They
remain active with Python optimization. Controls run with -O.
6. Module globals identify immutable paths and finite expected inventories.
Geometry, bit masks, queues and receipts stay local. No peer artifact is
writable through the verifier.
7. Read and parse failures propagate. Output uses exclusive creation. Source
inventory lengths and exact allowed paths are checked before reading or
importing sources. Hash and Git-blob checks bind retained bytes without Git.
8. The fixed geometry module is loaded after authentication. No input supplies
code, module names or call targets. No generated code or eval is used.
9. The import loader is the sole dynamic call boundary and loads one fixed
authenticated local module. All mathematical helpers otherwise have visible
direct call targets. This Python import boundary is a bounded runtime
deviation from a ban on indirect calls.
10. Registered replays enable -W error. Ruff check and formatting checks cover
every new Python file. There is no compiled component or compiler warning
configuration. The standard-library runtime and inherited geometry internals
retain their existing contracts.

The managed-runtime allocation and fixed import-loader deviations stay confined
to these offline checks. Revisit them if inputs become user-sized, the module
path becomes configurable, or the verifier becomes a service. Keep the bounds
and authentication tests when changing the implementation.
173 changes: 173 additions & 0 deletions research/p112-mixed-support38/PROOF.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,173 @@
# Mixed-rank shortening through common support 38

Let K be any characteristic-zero field. On the fixed 112 nonframe E8 root
lines, let C be the root-evaluation code, using doubled roots r_t:

```
C = {(r_t dot h)_t : h in K^8}.
```

Binary support rank is the rank over F2 of the pinned nine-bit labels in a
vector's support. Let pi_S delete ONE common coordinate set S.

**Theorem.** If |S| <= 38, and three residual pieces q_1,q_2,q_3 have binary
support ranks at most 5,5,6, then

```
dim_K(pi_S(C) intersect span_K(q_1,q_2,q_3)) <= 1.
```

The pieces can be dependent and have arbitrary coefficients. Rank constraints
apply only outside S. The theorem is sufficient; sharpness is not claimed.
PR3260 proves support 37 and leaves the 30-root D6 classification branch
open. This proof closes that branch at 38. It supplies no new historical
profile exclusion, full C5, decoder-floor theorem, supplier or kappa.

## Dependencies and exact finite coverage

The source-pinned root-simplex proof and receipt supply global P112 capacities
12 and 20 for binary ranks five and six. Their complete historical flat
censuses are inherited, not rerun here. The standard real Weyl-chamber theorem
and the rational-nullspace transfer below are written dependencies.

The new replay checks the 28 six-node parabolics. Only D6 and E6 have at
least 30 root lines, with counts 30 and 36. All 56 five-node parabolics have
at most 20 root lines. The normal planes of the two qualifying six-node
parabolics contain respectively two orthogonal root lines and an A2 triple.
Thus their Weyl images are covered by orthogonal root pairs and A2 planes.
The source module also checks all 960 simple reflections.

The replay checks all 1,120 A2 planes, their live union spectrum 78/80,
and transitivity under eight simple reflections. Each reflection induces
an invertible linear map of the nine-bit labels. The three representative
27-line projective fibers each span a binary seven-space. Their 2,667
five-spaces have fiber capacity ten. Their 127 hyperplanes have the following
complete full-E8 support / fiber-support counts:

```
(0,0):1, (12,11):27, (12,12):36, (16,15):36, (16,16):27.
```

These joint counts follow PR3262's support-plus-singleton argument; they are
rechecked here. The independent D6 check does not independently re-prove
these A2 premises or the global 12/20 capacities.

## Classify a two-dimensional intersection

Suppose the displayed intersection has dimension at least two. Choose two
independent restricted codewords and lifts c_h,c_k. Their full support union U
lies in S together with the three piece supports. Extend each residual piece
by zero on S. The global capacities give

```
|U| <= |S| + 12 + 12 + 20 <= 82.
```

The lifts are independent, since their restrictions are independent. Therefore
the rational span Z of all full E8 roots perpendicular to h and k has rank
at most six. At least 112-82=30 live roots belong to Z.

Every root in the rational span of Z is also perpendicular to h,k. A generic
real vector perpendicular to Z therefore has exactly those full E8 roots as
its perpendicular root set. Indeed, finitely many additional hyperplanes can
be avoided. Move this real vector into a closed Weyl chamber. Its perpendicular
root system is a parabolic subsystem of rank at most six. The checked counts
force rank six and type D6 or E6. Its two-dimensional normal plane equals the
K-span of h,k, since rational matrix rank is unchanged by characteristic-zero
extension. In particular it has a rational basis.

This argument uses the full perpendicular root system, even though the count
30 comes from live roots. It does not assume any individual piece is a
restricted codeword or any quotient-basis normalization.

## Exclude E6 using joint incidence capacity

In this branch the normal plane is an A2 root plane. Replace h,k by a rational
basis of that plane. Their live support union has at least 78 coordinates.
On the complement of S there is a fixed two-by-three matrix A with

```
(c_h(t), c_k(t))^T = A (q_1(t), q_2(t), q_3(t))^T.
```

At a coordinate in U outside S where exactly one piece q_i is nonzero,
the nonzero evaluation pair is proportional to column i of A. Such singleton
sites for q_i lie in ONE projective fiber. A zero column has no such sites. Besides three large
27-line fibers, the A2 plane has three one-line fibers.

Let M_i be total residual support of piece i and let J_i count its singleton
sites in U outside S.
For i=1,2, the global capacity 12 and fiber capacity ten give M_i+J_i<=22.
For the rank-six piece, M_3+J_3<=32. To see the latter, a one-line fiber gives
the stronger bound 21. For a large fiber in a seven-space W, let V be the
actual binary support span of q_3. If dim(V intersect W)<=5, then J_3<=10
and M_3<=20, giving at most 30. Otherwise V is a six-space contained in W.
The complete hyperplane inventory above gives M_3+J_3<=32. Supports can be
proper subsets of their binary closures; this only lowers these quantities.

Let N=|U minus S|. Each union coordinate has at least one piece incidence;
a nonsingleton has at least two. Thus

```
2N <= sum_i(M_i+J_i) <= 22+22+32 = 76.
2N >= 2*(78-38) = 80.
```

This contradiction permits cancellations, arbitrary columns of A, and piece
support outside U. Extra support only increases M_i. No common shift is
assumed modal.

## Exclude the D6 equality case with the actual puncture

Every D6 normal plane is spanned by two orthogonal E8 roots a,b. The full E8
nonzero evaluation union has 90 lines. Its four projective fibers have sizes
33,33,12,12. The two large fibers are the two coordinate axes. The exact live
union histogram over all 3,780 orthogonal root pairs is

```
82:448, 84:2688, 85:448, 86:168, 88:28.
```

Since |U|<=82, only the 448 minimum planes can occur. Then |U|=82 and |S|=38
are forced. Their residual union has at least 44 coordinates, while the
three pieces have at most 12+12+20=44 incidences in total. Equality follows
throughout. The pieces have disjoint supports of sizes 12,12,20, every support
coordinate lies in U outside S, and every coordinate is a singleton.

Consequently the rank-six piece's entire support of size 20 lies in ONE
projective fiber. A small fiber has only twelve full E8 coordinates. Each
large fiber in a minimum plane has 29 live labels of binary rank eight.
For each of the 448 planes and both large fibers, the primary replay checks
all 10,795 six-spaces in that eight-space via unique two-dimensional dual
spaces. Their maximum live intersection is always 18. A binary space of rank
at most six meets the fiber in a subset of one such six-space. This contradicts
support size 20 and proves the theorem.

The puncture is essential to this finite premise. A full 33-line axis has
six-space capacity 20, with sixteen maximizing supports. A separate checker
enumerates all 788,035 ambient six-spaces, rather than using the primary
fiber coordinates, and reconstructs those sixteen sets. It checks all 120
orthogonal four-root subsets inside the axis. Each support 20 meets each such
four-frame in two or four roots, with multiplicities 1,440 and 480. All 896
actual minimum-plane axis punctures are such four-frames. A separate orbit
check covers all 3,780 orthogonal pairs under the simple reflections.

This alternate check excludes a surviving support 20. It does not by itself
establish the stronger punctured capacity 18 for subsets of smaller full
intersections; the primary complete census supplies that stronger value.

## Credit and limits

Nick Petalas's PR3260 supplies the mixed 5,5,6 question and support 37 proof.
PR3262 supplies the joint A2 support/singleton argument. The source-pinned
root-simplex work, PR3249 and PR3255 supply the capacity and geometry lineage.
DreamingOfClouds/PR3154 supplies the TWOSIDE motivation; the retained E8/P112
lineage includes joaqcoo. Worker1's simultaneous uniform 5,5,5 support 46 audit
identified the same D6 prerequisite and has a distinct equality obligation.
The present theorem uses no unpublished sibling result.

The finite certificate and written characteristic-zero proof do not constitute
proof-kernel or independent human verification. Support39, global sharpness,
a fresh caller, the original C5 reductions, physical/recursive realization,
restoration, the complete consumer bill and multiplication transfer remain
separate obligations. No selected result is changed.
70 changes: 70 additions & 0 deletions research/p112-mixed-support38/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,70 @@
# P112 mixed-rank shortening at common support 38

Three residual P112 pieces of binary support ranks at most 5,5,6, after ONE
common deletion set of size at most 38, meet the restricted root-evaluation
code in dimension at most one over every characteristic-zero field. This
improves PR3260's sufficient threshold 37. It does not prove sharpness or a
new historical profile exclusion. [PROOF.md](PROOF.md) gives the complete
argument and inherited dependencies.

The new step closes the D6 boundary. A putative equality requires a rank-six
piece supported on 20 singleton coordinates in one projective fiber. All 896
relevant live axis fibers have rank-six capacity 18. An independent ambient
annihilator/puncture check also excludes support 20. The E6 branch uses
PR3262's rechecked joint incidence argument, with allowance76 versus80.
Exact coverage and full-census digest are in [RESULT.json](RESULT.json).

## Reproduce

Python 3.12 or later, standard library only, from the repository root:

```sh
python3.12 -W error -B scripts/research_package.py \
--package research/p112-mixed-support38
```

The three registered commands run sequentially with 120-second limits. Direct
Git-free export replay uses these commands, from the exported root:

```sh
python3 -W error -B research/p112-mixed-support38/verify.py
python3 -W error -B research/p112-mixed-support38/independent_check.py
python3 -W error -O -B research/p112-mixed-support38/test_verify.py
```

The export needs the package and its three declared baseline inputs at their
original relative paths. Pending PR proof/receipt bytes are retained locally
with exact commit, Git blob and SHA256 pins in [SOURCES.json](SOURCES.json).
No pending Git objects or Git subprocesses are needed for direct replay.
Authentication bounds inventory length, paths and source sizes before import.
`verify.py --output NEW_EXTERNAL_PATH` creates an external receipt and refuses
to overwrite an existing file. Set TMPDIR to an external session directory to
keep control-fixture files outside the checkout.

## Open next questions

Does support 39 admit a counterexample or a larger sufficient theorem? The D6
incidence equality used here no longer follows at 39. First bind the actual
puncture and mixed support-overlap patterns before another census. This is
unrun and not evidence that 38 is sharp. A new TWOSIDE application also needs
a named caller, current public-exclusion reconciliation and its exact retained
code dimension. [TASK.json](TASK.json) records these questions and blockers.

## Sources and verification limits

Credit Nick Petalas/PR3260/3262, DreamingOfClouds/PR3154 and the retained
root-simplex/E8/P112 lineage including joaqcoo. Worker1's concurrent uniform
rank-five support 46 theorem shares the D6 prerequisite but proves a different
predicate. This package uses no unpublished sibling theorem. All copied
predecessor bytes stay unchanged. New work and retained sources are Apache-2.0.

The global five-/six-flat capacities 12/20 and standard Weyl-chamber theorem
are inherited; their historical global censuses are not rerun. The independent
checker shares authenticated root/label reconstruction and does not independently
establish A2 capacities or characteristic-zero classification. Internal review
is not independent human or proof-kernel review. [CHECKS.json](CHECKS.json)
records executed commands, export evidence and review limits.

No full C5, decoder floor, native or physical supplier, restoration, complete
consumer, all-size multiplication theorem, new kappa or selected-frontier
update is claimed. Lean, Docker and the selected full cold replay are unrun.
Loading