Skip to content
Open
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
11 changes: 11 additions & 0 deletions research/longer-coupled-port-caller/CERTIFICATE-SUMMARY.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,11 @@
{
"orders": 75600,
"degree_counts": {
"2": 72020,
"3": 3580
},
"largest_witness_terms": 77,
"uncompressed_bytes": 6472926,
"compressed_bytes": 108057,
"sha256": "2e9279611423b0fc0488780e10cc8ac4ce3397f9c3f1809dd0ab294f4987ed1a"
}
Binary file not shown.
93 changes: 93 additions & 0 deletions research/longer-coupled-port-caller/DISCOVERY.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,93 @@
# Longer coupled-port search

Ali Romman with substantial OpenAI Codex assistance. Apache-2.0.

This file records discovery steps. The final result is in PROOF.md.
No new kappa is established.
Base: 6cf7a683f447b9f1e2333791907fd92be5009a63.
Claim: https://github.com/CrocSwap/integer-mult-bounds/discussions/5593#discussioncomment-19186397

## A useful necessary condition

Work in characteristic two. Let R_i be square-zero matrices and let their
sum be a diagonal projector E. Let P_i be diagonal coordinate projectors
with P_i R_i = R_i P_i = R_i. A coordinate on which E is one must occur in
at least three P_i.

Proof: zero occurrences fail the diagonal equation. With one occurrence,
the row and column of that R_i are those of E, so its square has diagonal
entry one. With two occurrences, call their row vectors a and e_j+a,
and their column vectors b and e_j+b. Square zero gives a*b=0 and
(e_j+a)*(e_j+b)=0. But a_j=b_j, so the latter expression is 1+a*b=1.
Both cases are contradictions.

For full rank-q queries in rank-2q coordinate supports, 3*rank(E) is at
most 2*q*L. The proof does not assume separate ports or any chronology.
If every full call has a private stage axis, the target basis makes its
projector diagonal. Thus this condition applies at every coupled rank.

A second consequence applies when each stage occurs twice and has no
rank loss. Its two rank-2q projectors sum to a rank-4q idempotent. The
projectors are orthogonal. The sum of the two compatible queries is
square zero and is supported on the diagonal target stage projector.
The same three-occurrence argument applies to these grouped queries.
It rules out full stage capacity whenever 3*N > 2*q*L, at all lengths
and all coupled ranks. No chronological assumption is needed for this
necessary condition. Rank losses and noncommuting local projectors need
further work.

## Query frame complement

For sum R_i = I_N, equal rank q, write R_i=U_i V_i with V_i U_i=0.
Set U=[U_i], V=[V_i]^T, K=VU. Then UV=I and C=I-K is an idempotent
of rank h=q*L-N, with q by q identity diagonal blocks.
Factor C=A B, BA=I_h. Each B_i A_i is a rank-q idempotent and their sum
is I_h. When L is even, the complementary idempotents I_h-B_i A_i
have rank h-q and also sum to I_h. Therefore L*(h-q)>=h. In particular,
N<=q*L-q-ceil(q/(L-1)). The trace condition also requires N even over F2.
These are necessary frame conditions, not a caller construction.

## Numerical discovery screens

The initial screen used the pinned PR5599 local invoice. It assumed all
2*q*L stage incidences could be used and distributed them as evenly as
possible. For q=104, eight calls on 720 fields appeared to have a raw
crossing near .00348661. That frame is impossible: h=112, while the
complement inequality requires 8*(112-104)>=112. Other paired-stage
screens also lose capacity before an operator is built. These floating
numbers are rejected discovery hypotheses, not kappa claims.

## Next experiments

The remaining construction must use noncommuting local stage projectors
and pay for lost stage capacity, or change another stated hypothesis.
The current search keeps the local P104 supplier and independent helpers.
Use PR5631's mixed chronology condition before a full word replay.

## Completed continuation

The initial finite search checked 24^3 coupled splits in each of 144
two-sweep orders: 1,990,656 assignments, with no chronology survivor.
The rank-two coupled search used Z3 5.3.2.0 on 144 orders and found all
unsatisfiable; this result alone was not used as a proof. Exact free-algebra
witnesses then covered all 75,600 order types, with no matrix-size bound.
The final theorem and its reduction are in PROOF.md.

DISCOVERY-RESULTS.json.gz preserves all five JSON discovery reports. Its
SHA256 is ec6fbe62a74cb05750db45c24c2a14ca18cb6b4f6bfb88a8151b76df40120ef8. The retained scripts
can regenerate these reports. The registered checker uses only the exact
order witnesses and standard-library matrix checks.

Commands, from the repository root:

```sh
python3 -B research/longer-coupled-port-caller/search_star.py
python3 -B research/longer-coupled-port-caller/search_orders.py --degree 2
python3 -B research/longer-coupled-port-caller/nc_star.py --degree 3 --limit 144
python3 -B research/longer-coupled-port-caller/build_certificates.py
```

The optional search_coupled.py command requires z3-solver==5.3.2.0. It is
not a verification dependency. Its finite negative output is superseded
by the exact algebra certificates. The numeric invoice screens do not
supply an operator, compiler or kappa result.
Binary file not shown.
197 changes: 197 additions & 0 deletions research/longer-coupled-port-caller/PROOF.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,197 @@
# All-order exclusion of the coupled star query frame

Ali Romman with substantial OpenAI Codex assistance. Apache-2.0.

## 1. Result and scope

The constant-loss frame in PR5631 passes the earlier stage-rank and
intersection cuts at arbitrarily large sizes. This package permits arbitrary
matrix coupling inside every good stage pair. It excludes every call order,
every coupled channel dimension, and every odd leaf count g>=3 for this
frame. It also supplies a small exact proof format for other matrix searches.

This is a theorem in the binary normalized model. It is not a reduction of
all complex or dyadic native callers to that model. It gives no new kappa.
Other query frames and connector contracts remain open.

Let F be any field of characteristic two. Let W have positive dimension d.
The main space is W_0 plus W_1 plus ... plus W_g, with W_0 the core. Let
E_j select W_0 and W_j. On these two blocks put

T_j = [[I,I],[I,I]].

All other blocks of T_j are zero. Split each good stage into two calls.
Their projectors P_j and P'_j have sum E_j, are orthogonal, and their
queries R_j and R'_j have sum T_j. Require P_j R_j = R_j P_j = R_j and
P'_j R'_j = R'_j P'_j = R'_j. These conditions cover every compatible
split at full stage rank. No separate-port restriction is imposed.

The four remaining queries are F_C, G_C, F_(I+C), G_(I+C), where C is any
idempotent on W. F_H has block H from each leaf input to the core output.
G_H has block H from the core input to each leaf output. All other blocks
are zero. These are the connector queries of the source star frame, after
an arbitrary common channel basis change. Their stage projectors are not
used in the exclusion. They can be thin calls. There can be arbitrary
constant basis changes already absorbed into the displayed matrices.

For odd g, the sum of all queries is I: the sum of the T_j has identity
diagonal and identity core/leaf cross blocks; the four connectors cancel
the cross blocks. Thus the frame passes the query-sum condition.

For the native equal-rank instance, d=2q, rank A_j=rank C=q. Each good
projector has rank 2q, each query has rank q, N=2q(g+1), and L=2g+4.
The theorem is stronger than that instance: its algebra does not use the
ranks of A_j or C.

**Theorem.** No order of these calls has an endpoint with the central
involution Q on every main field and the declared independent good-stage
projectors E_j, if g>=3 is odd. This holds for every d>0 and every F of
characteristic two. Failure already occurs in a mixed first derivative.

## 2. All coupled splits have a matrix parameterization

Since the two projectors are orthogonal and sum to E_j, both act inside
W_0 plus W_j. Also P_j T_j = T_j P_j = R_j. The centralizer of T_j has
block form

P_j = [[A_j+B_j, B_j], [B_j, A_j+B_j]],
R_j = [[A_j, A_j], [A_j, A_j]],
P'_j = E_j + P_j, R'_j = T_j + R_j.

To see completeness, write a general four-block matrix and equate its
products with T_j on the two sides. Its two diagonal blocks are equal,
and its two off-diagonal blocks are equal. Idempotence is equivalent to

A_j^2 = A_j,
A_j B_j + B_j A_j = B_j.

For example, in a basis with T_j=[[0,I],[0,0]], the projector is
[[A_j,B_j],[0,A_j]]. This also gives rank P_j=2 rank A_j. These are full
matrix parameters. Their entries can mix every channel. They are not
scalar choices or copies of a one-port calculation.

## 3. Necessary chronology equations

Put Q=1+e and z_j=1+d_j, with e^2=d_j^2=0. After normalization a good
call is (I+d_j P)(I+e R). Calls act in listed order, with later factors
on the left. The query sum is I. The coefficient of e*d_j is

E_j + sum_(i in stage j) [P_i, sum_(k later than i) R_k].

The desired endpoint has coefficient E_j. Therefore its defect must be
zero. This is PR5631's mixed chronology condition. The bracket is AB+BA
in characteristic two. An exact local matrix word supplies a separate
check of this formula in the verifier.

For each good stage j, retain only defect blocks whose row and column
are not both in {0,j}. A nonzero retained block involves stage j and
one other leaf k. A query at a third leaf contributes zero to this block:
its only nonzero rows and columns are at the core and that third leaf.
A connector contributes H or zero, independent of the number of leaves.
Thus these retained equations depend only on the relative order of the
two good stages and the four connectors.

Choose any three good stages. Delete every other good call from the order.
The retained equations on those three stages do not change. Rename the
three stages in order of first occurrence. If a stage's second labelled
call occurs first, replace A_j by I+A_j and leave B_j unchanged. This
preserves all local equations and exchanges the two calls. Consequently
there are exactly

10! / (2^3 * 3!) = 75600

order types to check. This finite reduction is independent of g and d.
It is the step that turns a finite order certificate into an unbounded
length and matrix-size theorem.

## 4. Exact certificates in a free algebra

Use the free associative algebra F2<A0,B0,A1,B1,A2,B2,C>. A word records
the order of its factors. In particular A0*B1 and B1*A0 are different
words. The equations are the six local relations in Section 2, C^2+C,
and the retained chronology blocks from Section 3.

For each of the 75600 order types, ORDER-CERTIFICATES.json.gz gives
explicit finite words u_l,v_l and equation indices i_l such that

sum_l u_l * equation_(i_l) * v_l = 1.

The separate checker expands this sum and cancels equal words modulo two.
There are 72020 witnesses of degree at most two and 3580 of degree at most
three. The largest has 77 terms. No SAT or numerical solver is needed to
check them. The search code uses finite linear algebra on words to find
these identities; the checker does not import that code.

If a tuple of matrices solved the equations, evaluation of each witness
would give 0=I_d. This is impossible for d>0. Evaluation preserves word
order, so no commutativity of the matrices is assumed. Since F contains
F2, the same identities work over every characteristic-two field.
This proves the theorem.

For a short example, put all first good calls before all connectors, and
put second good calls in reverse order. Four retained equations are

I + A0 + A1 + B1 = 0,
B1 = 0,
I + A0 + A2 + B2 = 0,
I + A1 + A2 + B2 = 0.

Their sum is I=0. The other order types need the certificates above;
this example is not used as a substitute for their complete coverage.

## 5. What the search removes

The source relaxed profile has 2q jobs of rank 1+8g and 2gq jobs of rank
9, against 2g+4 full calls and their complete independent helper bill.
For q=104 and g=1023, the PR5599 bill gives a positive relaxed moment at
saving .003125. The exact integer-power enclosure is in RESULT.json.
This exceeds the selected .00310045 but is not an attainable caller.
The theorem excludes every coupled split and every order of this frame,
including large cases that were not individually instantiated.

The preliminary eight-call, 720-field price screen also failed a frame
condition. If rank-q square-zero queries sum to I_N, their Gram complement
has rank h=qL-N and identity q by q diagonal blocks. Factoring it gives
rank-q idempotents E_i on an h-dimensional space with sum I_h. For even L,
the complements I_h-E_i also sum to I_h. Hence L(h-q)>=h. The proposed
q=104,L=8,N=720 has h=112 and 8(h-q)=64, so it is impossible.

A separate useful test is the three-occurrence lemma. If square-zero
matrices R_i sum to a diagonal projector E and are supported on diagonal
projectors P_i, each coordinate where E=1 must occur in at least three
P_i. With two occurrences, their rows are a and e_j+a and their columns
are b and e_j+b. Square zero gives a*b=0, but
(e_j+a)*(e_j+b)=1+a*b because a_j=b_j. This is a contradiction. One
occurrence also contradicts square zero. Thus 3 rank E <= sum rank P_i.
This covers arbitrary coupled queries whose local projectors commute.
For paired stages at full rank, group the two orthogonally supported
queries. Their sum is square zero and supported on the target stage
projector. The same test applies. Rank loss requires another argument;
the star theorem above supplies one for the declared frame.

## 6. Verification boundary and next routes

The order reduction and free-algebra evaluation argument are written
proofs. The executable checker verifies the complete finite order cover
and every exact identity. A different matrix implementation checks the
local identities, query frame and mixed coefficient formula on coupled
noncommuting examples. The one-leaf case is a genuine full polynomial
word and is retained as a positive control. Corrupted certificates and
changed query data must fail. No proof assistant or independent human
acceptance is claimed.

This theorem does not exclude a different query Gram frame, a different
connector map, more connector queries interleaved with good calls, nonfull
stage ranks, different main outputs, noninvolutive stages, helper feedback,
or arbitrary complex/dyadic callers. It does not price or certify a new
native multiplication algorithm. A future search can change one of these
premises, or use the same free-algebra certificate format on another exact
frame before it allocates large matrices and a complete native bill.

Credit Cinna Davis / 1337isnot1337 for the controller, simplex and private
projector lineage; PR5617 for stage/intersection cuts; PR5631 for the
constant-loss star and mixed chronology equation; and the P104, helper,
native, compiler and outer-framework authors retained by those sources.
Free associative algebras and linear combinations of polynomial equations
are standard tools. The all-order coupled-frame reduction and its exact
witnesses are the contribution of this package.
74 changes: 74 additions & 0 deletions research/longer-coupled-port-caller/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,74 @@
# An all-order obstruction for coupled star callers

The constant-loss star frame cannot be repaired by port coupling or by a
longer order. In the binary normalized model, this package excludes every
odd leaf count g>=3, every coupled channel dimension and every order of
its calls. It keeps the star query frame and four complementary connector
queries fixed. **No new kappa is claimed.**

[PROOF.md](PROOF.md) gives the complete contract and the three-stage
reduction. There are only 75,600 relative order types after that reduction.
For each type, an exact identity in a free noncommutative algebra proves
0=I. The identities apply to matrices of every size. This is stronger than
a bounded search over matrix entries or port counts.

[RESULT.json](RESULT.json) records the independent replay. The compressed
[certificates](ORDER-CERTIFICATES.json.gz) contain every witness. The
[checker](check_certificates.py) expands them without importing the search
code or using a solver. There are 72,020 witnesses of degree at most two
and 3,580 of degree at most three. None has more than 77 terms.

The package also checks actual coupled matrices, the mixed coefficient
formula, a complete one-leaf positive word, and corrupted data. The pinned
PR5599 bill gives a positive relaxed moment at saving 1/320 for g=1023;
the theorem shows why that price screen cannot become a caller in this
frame. The selected record advanced upstream to .00310045 during this
work. It was not changed by this contribution.

## Reproduce

Use Python 3.12 and the standard library:

```sh
python3 scripts/research_package.py --package research/longer-coupled-port-caller --check-only
python3.12 scripts/research_package.py --package research/longer-coupled-port-caller
python3.12 -I -O -B research/longer-coupled-port-caller/verify.py
```

The registered check uses the saved exact witnesses. To rebuild witnesses:

```sh
python3.12 -B research/longer-coupled-port-caller/build_certificates.py
```

The optional discovery searches and their complete receipts are described
in [DISCOVERY.md](DISCOVERY.md). Their Z3 results are not proof inputs.
Exact source commits, Git blobs and SHA256 hashes are in
[SOURCES.json](SOURCES.json). The package began at canonical main
6cf7a683f447b9f1e2333791907fd92be5009a63; the refresh at
df0137efd35e11b4445734bf612bdd4c7ebc330c changed the selected record but
not the priority or working instructions.

## Open next questions

- Change the query Gram frame. Test its chronological matrix equations
before any full native construction or favorable price claim.
- Change the connector maps or add interleaved paid queries. Recompute
every equation and keep the complete bill.
- Establish a valid characteristic-zero statement for a concrete caller.
General dyadic coefficients cannot be reduced modulo two.

Other frames, connector contracts, nonfull stage ranks, noninvolutive
stages, changed targets and productive helper feedback remain outside the
theorem. No native replay, proof-assistant proof, independent human review
or global obstruction to improving kappa is claimed.

## Credit

Ali Romman with substantial OpenAI Codex assistance; no subagents.
Apache-2.0. Credit Cinna Davis / 1337isnot1337 for the controller and simplex
lineage, PR5617 for the stage/intersection cuts, PR5631 for the star frame
and mixed chronology condition, and all P104/helper/native/compiler/outer
source authors retained by those packages. Free-algebra identities and
finite linear algebra are standard tools. The new work is the coupled
parameterization, the all-order reduction and its exact certificates.
17 changes: 17 additions & 0 deletions research/longer-coupled-port-caller/REFRESH.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,17 @@
{
"checked_at_utc": "2026-10-26T01:59:00Z",
"main": "df0137efd35e11b4445734bf612bdd4c7ebc330c",
"selected": "0.00310045",
"selected_record_sha256": "2a0220e866417774606e0f991e41b9c1d26b1f43b97857238cf478dbc7d330b5",
"priorities_or_working_instructions_changed": false,
"active_claims_checked": [
5600,
5636,
5638,
5642,
5125,
1166
],
"own_scope": "Global caller with fixed supplier; no overlap with local contact/history/G20 construction",
"coverage": "Live latest Discussion updates plus the initial full open-PR and Discussion overlap cache; sparse local notes remain a limitation"
}
Loading
Loading