Skip to content
Draft
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
34 changes: 34 additions & 0 deletions research/p104-prepared-split-reuse/PRICING-EDITS.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,34 @@
[
[
"Q('.00309839'), Q('.0030983905')",
"Q('.00309871'), Q('.0030987101')"
],
[
"PREVIOUS = Q('.00308189')",
"PREVIOUS = Q('.00309839')"
],
[
"13084",
"13066"
],
[
"66852",
"66816"
],
[
"'kappa_decimal':'.00309839'",
"'kappa_decimal':'.00309871'"
],
[
"previous_open_pr5580",
"previous_open_pr5599"
],
[
"'peak_owned_fields':1",
"'peak_owned_fields':local['peak_owned_fields']"
],
[
"'local_paid_rank_mass':local['paid_rank_mass']",
"'local_paid_rank_mass':sum(r*n for r,n in local['paid'].items())"
]
]
149 changes: 149 additions & 0 deletions research/p104-prepared-split-reuse/PROOF.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,149 @@
# Two paid queries replaced by a prepared split helper

Use PR5599's exact complete word at
`0644592d421b375c221045954fdf13932575ffa8`.
Its fifteen helper splits, scalar bases, phase conventions and caller remain
credited to that source. This construction changes only the preparation
and two queries described below.

## Exact source-version identities

Write `h_r` for original input r and `x_r` for its current normalized value.
The source has 750 independent productive inputs and eight owned roles.
Fresh exact replay, before any preparation, proves:

| Source event | Current-value identity | Paid source operation |
| --- | --- | --- |
| 13670 | `x573-x572+2*x732 = -2*h573` | `MOVE 573: 3 -> 595`, rank 1 |
| 13691 | `x573-x572-2*x710 = -2*h573` | `MOVE 573: 595 -> 603`, rank 1 |

The identities include every source input coefficient and current data version.
They are derived by `verify.py:source_relations`, rather than read from a
saved relation table. The source scalar basis changes are included in replay.

Source helper 573 starts at frame 3, the zero subspace, and finishes at
frame 603, a two-dimensional subspace. It is the rank-two prefix target
created by PR5599's eleventh split. The rank-seven successor has its own
independent input and remains in the word.

Declare role 573 owned and initialize it to zero before the source's entry
basis program. Its arbitrary dirty entry is erased. Thus `h573=0` is an
internal preparation, not a promise on any retained productive input.
Remove its rank-two target from the productive invoice.

All other source output rows are independent of `h573`. Exact replay of the
new word checks all 749 retained independent inputs, 208 main outputs,
541 helper outputs and every owned return. A dense integer transpose checks
the same complete linear map through a separate arithmetic path.

## Explicit replacements

At source event 13670, roles 572 and 732 are already at frame 595.
Replace the rank-one MOVE with:

```text
CLEAR 573 at 3
INIT 573 at 595
COPY 573 <- 572
ADD 573 += -2*732
```

The first identity proves that this gives the required value in frame 595.
Keep the following source ADD and every later source instruction.

At source event 13691, roles 572 and 710 are already at frame 603.
Replace the rank-one MOVE with:

```text
CLEAR 573 at 595
INIT 573 at 603
COPY 573 <- 572
ADD 573 += 2*710
```

The second identity proves that this gives the required value in frame 603.
In particular, the following source SCALE and split basis operations remain
unchanged. No operation is moved across their reads or writes.

After the source program, clear role 573 at frame 603. Its exact output was
already zero because its prepared input was zero. The eight original owned
roles retain their source operations and cleanup.

The word has 758 total roles, nine owned roles, eleven owned lifetimes and
peak owned occupancy two. All nine original dirty owned columns vanish from
all outputs. The new instructions use only the source INIT, CLEAR, COPY and
ADD operations. Neither a zero retag nor a new Gaussian operator is needed.

## Native meaning and every positive width

The normalized row in a frame denotes a linear combination of arbitrary
original fields under that frame's native operator. Both reconstructed
values use sources already in the required physical frame.
The COPY and scalar ADD operations therefore act on the actual physical
fields, with no omitted transport or phase correction.

For each legal positive tensor width f, substitute the tensor power of every
source native chart. Keep rational scalar coefficients outside the tensor
power. Equal charts commute with these whole-field scalar combinations.
The two source identities therefore prove both replacements at every width,
under the same native premise as the source.

No retained independent input is copied into another productive input slot.
Role 573 is owned and receives no target credit. Its old external entrance
normalizer can be omitted because its input is replaced by zero. Its old
exit normalizer can be omitted because its output is owned zero.
All other helper normalizers and both main interfaces remain unchanged.

The verifier checks that every retained paid-map pair occurs in the source.
The source native evaluator rechecks the signed Pauli images and absolute
phase anchors of all those maps, all original helper endpoints, and the
208 main adapters. Checking the discarded helper endpoint is harmless extra
evidence; its target and two unused normalizers are removed from accounting.

## Complete invoice and conditional bound

The local paid profile is:

```text
{1:2209, 2:459, 3:494, 4:94, 5:164, 6:18, 7:14}.
```

The retained productive helper profile is:

```text
{1:4, 2:4, 4:90, 5:10, 7:5, 8:18, 9:410}.
```

Relative to PR5599, the local paid delta is `-2 H1`, and the target delta
is `-H2`. For `0<t<1`, `2^t<2`. The target-minus-paid moment therefore
increases by `2-2^t>0`. At t=1 the delta is zero.

The unchanged physical controller has 18 local calls, 15 thin query stages
and 15 whole idle stages. Each call uses its own 541 fresh helper inputs.
It invokes the complete new supplier, which returns all owned storage zero.
The global deltas are `-36 H1` paid and `-18 H2` productive.
There are 66,816 paid jobs and 13,066 independent productive fields.
The complete rank surplus stays 760.

The new word has 25,510 instructions and scalar allowance 220,432.
All initialization, erasure, copying, scalar operations and routes are charged.
The route allowance includes eleven owned lifetimes and peak occupancy two.
The local cap 10^7 and global cap 10^9 remain sufficient.
The added owned role replaces a productive role, so the total field count
stays 758. The native paid-rank sum is recomputed from the changed word.

All scalar coefficients are dyadic and satisfy the source growth bound.
There are fewer than 26,000 operations, each with growth below 2^16.
The same precision and tape premise covers this explicit finite template
with its charged constants. The rounded multiplier is not materialized.

The source pricing code is retained with the exact edits listed in
`PRICING-EDITS.json`. The target debit is included in both completion tails.
The source's tail decay 39/40 and compiler logarithm power 1024 are retained.
Two rational moment engines test saving 0.0030987101 and fund the claim
**kappa=0.00309871**. They also reproduce PR5599's predecessor crossing.
The same run checks the compiler threshold, BIT stock, 54 strict outer
predicates and four functional predicates.

The five inherited framework assumptions remain. This is an author-checked
conditional construction, without independent human or proof-kernel acceptance.
84 changes: 84 additions & 0 deletions research/p104-prepared-split-reuse/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,84 @@
# Prepared split helper: conditional kappa 0.00309871

This construction raises the complete conditional bound from
**0.00309839 to 0.00309871**. It composes an owned-helper preparation with
Ali Romman's fifteen independent helper splits in
[PR5599](https://github.com/CrocSwap/integer-mult-bounds/pull/5599), pinned at
`0644592d421b375c221045954fdf13932575ffa8`.
The selected repository record remains unchanged.

The exact crossing lies in `(0.00309871051439, 0.00309871051440)`.
It exceeds PR5599's crossing by between `0.00000031699227` and
`0.00000031699229`.

The rank-two prefix helper 573 becomes owned storage, initialized to zero.
Two exact source relations then replace its two rank-one queries with
ordinary copies and additions. Its rank-two output is debited. The independent
rank-seven successor from PR5599 remains productive.
For every `0<t<1`, the complete local moment improves by `2-2^t>0`.

The complete caller applies the change in all 18 local calls. It has
**66,816 paid jobs and 13,066 independent productive fields**, with rank
surplus 760. The local interface has 749 independent productive inputs and
nine owned roles. All owned roles return zero; peak occupancy is two.
There are no new paid native maps or new instruction types.

The word keeps the source register numbers. Its legacy `public` slot limit
is 750; `independent_public` lists the 749 actual productive inputs.
`owned_roles` includes role 573 and the eight original owned roles.

These are conditional, author-checked results. The five inherited framework
assumptions remain: native implementation at every width, owned erasure,
uniform compilation, precision and tape costs, and analytic recovery.
The complete uniform word and rounded multiplier are not materialized.
Independent human review, proof-kernel acceptance and frontier promotion
are not claimed.

Read [PROOF.md](PROOF.md) for the source identities and physical implementation.
[construct.py](construct.py) emits the complete word.
[RESULT.json](RESULT.json) records the exact checks and complete bill.
[VALIDATION.md](VALIDATION.md) records the executed validation.

Reproduce with Python 3.12 and the standard library:

```sh
python3 -B scripts/research_package.py --package research/p104-prepared-split-reuse
```

The verifier checks all 758 endpoint rows and independent transpose columns.
It derives both relations before specializing the source input.
It checks every retained native map with absolute phase, the full physical
caller, two rational moment engines and all 58 outer predicates.
Seven semantic controls test initialization, reconstruction, chart matching,
owned cleanup, paid accounting and the productive debit.

[SOURCES.json](SOURCES.json) pins the complete PR5599 word, its local and
pricing code, its split witnesses, proof and notices. The unchanged PR5545
export supplies 33 authenticated runtime files. Replay needs no network,
solver or pending Git objects. The source scalar basis programs remain intact.

This changes the independent-input contract and actual paid word. It falls
outside the fixed-input passive-history obstruction in
[PR5598](https://github.com/CrocSwap/integer-mult-bounds/pull/5598).
No global optimality or worldwide novelty claim is made.

Ali Romman supplied PR5599's fifteen splits, PR5580, and the P104 caller and
pricing machinery. GamingPuzzled supplied PR5569; DreamingOfClouds supplied
PR5561; thu-nmrc supplied PR5551. David Arias, Cinna Davis and Gareth Davidson
supplied the underlying P104 word and private-copy construction.
The preparation and two query replacements are by Scratch-net, with
substantial OpenAI Codex assistance. No subagents were used.

The reuse question comes from Dugongue's PR5586 atlas. Cached-relation
precedents include Ali Romman's PR5383 and Dillon Amburgey's PR5402.
Retain Chris R/Clear1nsight's review and native interpreter, Cao Yifan's
native lineage, Jacob Alan Sussman's productive recursion, zxv's compiler
and outer machinery, and all authors in the source closure and
[retained PR5599 notice](inputs/pr5599-NOTICE). New work is Apache-2.0.

## Open next questions

Independent review should check the changed owned interface and the two
reconstruction identities, then replay the full caller and funding checks.
All source dependencies are present. The five framework assumptions and
independent scientific acceptance remain open.
Loading