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
1 change: 1 addition & 0 deletions research/p104-loan-snapshot-engine-on7621/EXPECTED.json
Original file line number Diff line number Diff line change
@@ -0,0 +1 @@
{"kappa_prefix": "0.0036624027915539"}
23 changes: 23 additions & 0 deletions research/p104-loan-snapshot-engine-on7621/NOTICE
Original file line number Diff line number Diff line change
@@ -0,0 +1,23 @@
Apache-2.0. DreamingOfClouds with Claude Opus 5.5 (Anthropic), via Claude Code.

The device is SYLforge's: the stationary-lifetime construction of #7616 and #7620 (a register drops a middle stop
while a partner inserts one, the dropped work rides a loan through a carrier, and stationary owned snapshots carry
the rest). #7620 is SYLforge's composition of the 167-CLEAR stationary primal/dual bypasses on #7616 (candidate work
with Claude Opus 5.5 and Antigravity, validation with OpenAI assistance) and builds on Ali Romman's #7605, our #7611
stay drops, Clear1nsight's #7612 contact rechart, Hao Yan/FanPUYen's #7604, SYLforge's #7603, ToikY's #7546/#7544
and Rohan Arun's signed #7113 caller. The base word is our #7621 (courier + helper re-stop on #7620), which builds
on our #7615 and #7617 (courier lineage; #7617 supplies the control word) and on the crowd's courier pattern
(Clear1nsight's #7244/#7274/#7314/#7334/#7481, rohanarun's #7480 and #7277, CubeJerry's #7577, petalas's #7326).

Vendored data (unmodified, data only): data/base-7621-word.json.gz and data/invoice-7621.json are our #7621's
data/word.json.gz and data/invoice-new.json; data/affected-pairs-7620.json is the 837-pair census pinned in #7620's
native runner; engine/data/word-7620.json.gz is #7620's word and engine/data/word-7617.json.gz is #7617's word
(controls only); engine/data/7113_CALLER.json is Rohan Arun's signed #7113 caller (used to price kappa).

engine/ is our own code except three helpers written earlier in this project and vendored with only their path lines
changed: pricekit.py (the exact pricer), fastx.py (randomised-exact replay) and jl.py (kappa under the #7113
caller). No other author's code was run; their words are read as data. check.py adapts the replay, census,
crossing and column-test sections of our #7621 package; its column tests are written from the description of
the affected-pair tests in #7620's README and native runner (512 forward and 512 adjoint Gaussian columns, signed
Pauli action, absolute phase, wrong-phase rejection), following the repository's frozen interpreter and compiler
documented in main.
127 changes: 127 additions & 0 deletions research/p104-loan-snapshot-engine-on7621/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,127 @@
# CONDITIONAL κ = 0.0036624027915 — 14 loan + stationary-snapshot instances on #7621, with the engine that finds them (premises: exactly #7620's, over 180 CLEARs)

**κ = 0.00366240279155398…, ×1.00025679 (+0.025679%) over our
[#7621](https://github.com/CrocSwap/integer-mult-bounds/pull/7621)** (κ 0.0036614625547). The official
`evaluate_recursive_power` bracket (256 bits) is [0.0036624027915, 0.0036624027916]; the claim 0.0036624027915 is
below the exact crossing 0.00366240279155398.

This PR ships the word **and the engine that produced it** (`engine/`), so anyone can run these incremental
passes on any P104 native word. The device is [SYLforge](https://github.com/CrocSwap/integer-mult-bounds/pull/7620)'s
(#7616, #7620); the engine finds, solves and exactly checks instances of it automatically.

> **Premises (inherited; one reading widens).** Conditional on exactly #7620's premises: its five framework
> assumptions (`native_all_columns`, `owned_erasure`, `compiler`, `precision_tape`, `analytic_recovery`) and its
> fixed allocation of 17·68 = 1,156 owned banks. Its CLEAR reading now covers **180 CLEARs** (74 original + 106
> stationary), not 167: the 13 new CLEARs belong to 13 new stationary owned lifetimes of the commonest shape among #7620's own
> (INIT, one write, one read, dirty CLEAR, no MOVE; 82 of #7620's lifetimes have it), each with an exact CLEAR certificate
> under the rule of #7620's `verify_certificates` (data/clear-certificates.json). The word's rows, end charts, 47 owned MOVEs,
> peak 50, 68 labels, all 1,018 old owned observations and the old 167 CLEAR contents are #7621's. If #7620's premises
> are refused, this word is void with #7620 and #7621.

## Run it yourself

```sh
python3 engine/run_on.py WORD.json.gz [OUT_PREFIX] # census -> greedy -> re-census, about 1 minute, one process, under 1 GB
python3 engine/run_on.py data/base-7621-word.json.gz /tmp/out # reproduces this PR: /tmp/out_14.json.gz == data/word.json.gz
python3 -I -B check.py engine # the same, as a test (JSON-equal word and equal instance log)
python3 engine/control.py # controls on #7617 with #7620's instances as data
```
Python 3.12 and the standard library only; `pricekit.py`, `fastx.py`, `jl.py` and the #7113 caller are vendored in
`engine/`, and every path is relative to the script. Full notes are in [engine/ENGINE.md](engine/ENGINE.md).

**The move.** A register r drops an interior stop X from its path A→X→B. A partner p takes one extra stop: on side A
(all 14 here) p is the reader of r at X and inserts A before X. The work r would have done at X is carried by (i) a
loan through a carrier c, a public register whose path runs A→X→B consecutively, (ii) stationary owned snapshots
(INIT at X, one write, one read, dirty CLEAR) when r's X-writes must reach p, and (iii) a signed role swap between r and p, with
a settle at B. Natively each instance swaps one 3|1 hop merge for one 1+2 split.

**How the engine works.** `lb.py` indexes a word (register stays, ops, owned live intervals, projected-state checkpoints)
and solves a candidate exactly: the template coefficients that are unknown enter linearly, so a windowed replay (mod 2^61−1, 4-dim random
projection) is affine in them, is solved, rationalised and re-verified, and must reproduce the window-end state, every pre-existing
owned observation and every CLEAR content. `tpl.py` holds the templates and the legality-keeping slides. `run7621.py` takes an exact census of
both shapes ordered by the convexity gain of x ln x, tries instances greedily, accepts one only if `check.full` passes, the owned
peak stays ≤ 50 and the PRICEKIT κ rises, then re-censuses until a pass lands nothing. `control.py` rebuilds #7620's own instances
on #7617 (read as data): 39 of 44 shape instances rebuilt, 258@829 and 464@526 identical to SYLforge's (same carriers, same coefficients).

**Known blockers on this word (85-place re-census after the 14 landed).**

| reason | places | meaning |
|---|---|---|
| no solution | 20 | contamination reaches registers that never meet r at B |
| no free label (peak 50 binds) | 17 | owned peak 50 is full over the snapshot interval |
| no carrier | 17 | no register runs A→X→B |
| A window | 15 | the carrier leaves A before r or p arrive |
| slides | 10 | a slide partner is touched in between |
| write after read / read before p at X | 2 + 2 | side-A write-after-read; r read before p arrives at X |

**Next levers.** A second carrier, so a contaminated register can be settled where it meets a register holding ρ rather than r;
shared snapshots or a reorder of busy owned lifetimes to free peak-50 events; carriers that are not consecutive with in-transit
settles; census shapes with several sources or readers.

## The edits

73 deletions and 154 insertions on #7621's word (77,196 → 77,277 events; ADD +55, INIT +13, CLEAR +13;
[data/edits.json](data/edits.json), instances in [data/instances.json](data/instances.json)). Fourteen instances, all side A:

| r | X | p | carrier c |
|---|---|---|---|
| 382 | 438 | 553 | 1144 |
| 422 | 492 | 577 | 1182 |
| 392 | 455 | 562 | 1154 |
| 391 | 451 | 558 | 1153 |
| 390 | 453 | 560 | 1152 |
| 386 | 452 | 559 | 1148 |
| 374 | 433 | 548 | 1136 |
| 366 | 420 | 541 | 1128 |
| 364 | 419 | 540 | 1126 |
| 360 | 417 | 538 | 1122 |
| 359 | 413 | 534 | 1121 |
| 358 | 415 | 536 | 1120 |
| 356 | 416 | 537 | 1118 |
| 348 | 399 | 528 | 1110 |

In each, r's route loses X and p's route gains the chart before X; exactly these 28 registers change chains. Thirteen instances add a
snapshot (348@399 reroutes r's X-writes straight into p). 24 new MOVE pairs appear, all strictly nested; 14 leave the MOVE set.

## Price

Native paid delta per call {1:+1, 2:+13, 3:−28, 4:+13, 5:+1}; over 17 calls {1:+17, 2:+221, 3:−476, 4:+221, 5:+17}.
Paid positions 110,145 and rank surplus S = 1,424 are unchanged. Official tools (`price.py`): `compare_recursive_invoices` against #7621 gives
`uniformly_better`, strictly negative for r^(1−C); `evaluate_recursive_power` (256 bits) gives `new_cost_lower` at C = 0.0036624027915, 0.0036614625547 (#7621)
and 0.0036611026181 (#7620), `new_cost_higher` at C = 0.0036624027916.

Ordinary bill (17 calls, #7000 formula) 785,022,672 → 785,921,632 < 10^9; precision bit bounds 303 / 299 / 299 unchanged; common grid
2^25·1125 with 40 numerator bits unchanged; non-dyadic literal events 5,903 → 5,903. An independent read-only audit found the word sound relative to
#7620/#7621 (exact replay; official #7000 native engine with bad_inputs 0; strictification gate passed, γ 0.941467, δ 0.908157;
22/22 rejecting controls).

## Checks

```sh
python3.12 -B scripts/research_package.py --package research/p104-loan-snapshot-engine-on7621
```
`VERIFY.json` registers six commands: `core` (words, edit list, exact replay of both, owned data, observations, certificates, structure, invoice, κ),
`controls` (22 rejecting controls, 2 accepting), `columns1` and `columns2` (the affected-pair column tests), `price.py` and `engine` (the engine rerun on #7621's word).

**Affected pairs.** #7621's census is 846 pairs (the 837 pinned by #7620's runner, minus (719,127), plus 10). This word changes 28 more routes; the census
becomes **964 pairs** (+118: the old and new route pairs of the registers not already counted; none dropped; every new MOVE pair is inside).
`check.py` runs 512 forward and 512 adjoint Gaussian columns, signed Pauli action, absolute phase, exact ascent rank and a wrong-phase rejection
on all 964 (964 × 1,024 = 987,136 columns, in two halves of 482 pairs). Census files: `data/affected-pairs-7621.json`, `data/affected-pairs-loan.json`.

## Conditional premises and what is not rerun

The claim carries exactly #7620's premises (above), now over 180 CLEARs, and the signed #7113 caller and BIT premise of the base. This package
does not rerun #7620's full native/signed 17-stage application, catalogue, outer predicates or precision/tape counts (header, rows, grid and owned data
are unchanged); the audit ran the official #7000 native engines on the word. No upstream maintainer has reviewed the package.

## Credit

The device is SYLforge's ([#7616](https://github.com/CrocSwap/integer-mult-bounds/pull/7616), [#7620](https://github.com/CrocSwap/integer-mult-bounds/pull/7620));
the base is our [#7621](https://github.com/CrocSwap/integer-mult-bounds/pull/7621), on [#7617](https://github.com/CrocSwap/integer-mult-bounds/pull/7617)
and [#7615](https://github.com/CrocSwap/integer-mult-bounds/pull/7615) (courier lineage). Signed caller: Rohan Arun's #7113. See [NOTICE](NOTICE).
DreamingOfClouds with Claude Opus 5.5 (Anthropic), via Claude Code. Apache-2.0.

## Open next questions

**The premises.** A ruling on #7620's inherited assumptions and its CLEAR reading (`owned_erasure`, clear-form-gap Q1), now with 180 CLEARs, decides
#7620, #7621 and this word together. See [TASK.json](TASK.json).
30 changes: 30 additions & 0 deletions research/p104-loan-snapshot-engine-on7621/SOURCES.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,30 @@
{
"schema_version": 1,
"base_commit": "e54523517d0f03e58352e647c403a5af3e525a57",
"files": [
{
"path": "certificates/selected-result.json",
"usage": "provenance",
"git_blob": "75d77b328ea02ded535e7f23aee5b1e55d0bb1a8",
"sha256": "154c3d30455d778fd3343b6b4851f0fa4665f0f7625b0adb3675e2bafbbf6bf1"
},
{
"path": "docs/research/agent-task-catalog.json",
"usage": "provenance",
"git_blob": "3404fbdd5e7529a0674eef26114b5071e4586cec",
"sha256": "ead6cdc9d0af6d58e8a49744d8ec8c7c15e13d2ba93c276c434bcf9b51b93a81"
},
{
"path": "scripts/compare_recursive_invoices.py",
"usage": "input",
"git_blob": "fda2a852e9937e42fe40df0707694465ca7c1b48",
"sha256": "3930c22c11da7bc6042cfeb6e459ecb06e4478430037bee3469d034353855fc6"
},
{
"path": "scripts/evaluate_recursive_power.py",
"usage": "input",
"git_blob": "ca50f8021fc3238a58fd420383f4619a8d008ace",
"sha256": "52b20f41fd0bb55fa88dccd6015af2ce358917817b9506720f3e60ce681cc3ad"
}
]
}
Loading
Loading