|
| 1 | +# PR #555's punctured R439 bit word with two A-section transports |
| 2 | + |
| 3 | +Two helper additions of PR #555's word (`data/bitE8mut-p111.json.gz`, v = 111, R = 439) are moved out of the D4 block-frame |
| 4 | +mini-node gates they sit in and rerun at other frames. Nothing else changes: same registers, same scatter, same retained |
| 5 | +totals (nine COPY ranks of 7, decoder constant 63), same frozen hosting selections, same pipelines. |
| 6 | + |
| 7 | +| | PR #555 | this word | change | |
| 8 | +|---|---|---|---| |
| 9 | +| bit c, nine stages (PR #510's chain, 32 replicas, parameterised by `code/adapt9p.py`) | 0.001733052823330162099157 | **0.001733427039589844234749** (`1733427039589844234749/10^24`) | +0.0216% | |
| 10 | +| bit c, eleven stages (PR #493's chain as parameterised by PR #545's `adapt.py`) | 0.001710608806098012673516 | **0.001710951900777583975622** (`855475950388791987811/(5·10^23)`) | +0.0201% | |
| 11 | +| κ with #548's complex b (0.001593519419282307) | 0.001590984155135328 | 0.001590984155135328 | none | |
| 12 | + |
| 13 | +κ is complex-bound for every cited complex supplier (#548, #544, #535, #531), exactly as with #555's word, so this changes |
| 14 | +no κ. The gain is bit headroom only: c is above every cited complex b (the largest, #548's, is 0.0015935), so the bit side |
| 15 | +does not bind. The local paid profile goes from `{1:1321, 2:1342, 3:80, 4:47, 5:102, 6:10, 7:18}` to |
| 16 | +`{1:1323, 2:1342, 3:79, 4:44, 5:102, 6:11, 7:19}` (rank mass 5,129 both; 27,234 records, 24,305 common-frame ADDs, payload |
| 17 | +93 bits, 636 frames, all unchanged); the nine-stage global profile is `{1:381024, 2:400416, 3:22752, 4:26880, 5:29376, |
| 18 | +6:17376, 7:26784, 42:288}`. |
| 19 | + |
| 20 | +## What changed (`moves.json`, applied by `apply.py`) |
| 21 | + |
| 22 | +Both moves take one summand out of a mini-node gate of phase A (the nine phase-A gates at the four D4 block frames 506-509, rank 4, |
| 23 | +that carry a pseudo-target's B-part; gate and register numbers are those of the base word, helper h is register 222 + h): |
| 24 | + |
| 25 | +- **hoist.** A[1993] is `helper 222 += {x17, h49, h71, h130, h191, h246, h248, h249}` at frame 509 (rank 4). The summand |
| 26 | + `helper 222 += helper 248` (register 444 += 470) becomes its own gate at rank-1 frame 79, placed immediately before that gate. |
| 27 | +- **sink.** A[1994] is `helper 191 += {x17, h49, h71, h130, h149, h247, h249}` at frame 509. The summand `helper 191 += helper |
| 28 | + 247` (register 413 += 469) becomes its own gate at rank-7 frame 517, the frame of the gate `helper 257 += helper 191 + helper |
| 29 | + 74` (helper 257 is a retained total), placed immediately after base gate 2028. |
| 30 | + |
| 31 | +Each move factors a multi-source addition, moves one summand across additions that commute with it and gives it another |
| 32 | +frame from the existing catalogue; the dirty-register endpoint is unchanged on every column. C is rebuilt as the literal |
| 33 | +reverse of the z-destination gates of A and B at the full frame (A has 2,075 gates, C 2,574). The profile changes by |
| 34 | +`+2 e1 − e3 − 3 e4 + e6 + e7`, a zero rank-mass change. The independent checker (`xcheck_p.py`, with the interface lemma |
| 35 | +H1-H5 and its word-level mutation controls) and the official cohort checkers accept the word, and the hosting programmes |
| 36 | +rediscover the frozen selections. |
| 37 | + |
| 38 | +## Where the moves come from, and why #541/#542/#548's move lists are not reused |
| 39 | + |
| 40 | +The mechanisms are those of three open PRs on PR #510's R480 word: Dugongue's #541 (factor a multi-source helper addition, move |
| 41 | +one summand across its commuting prefix and give it the source's rank-1 frame: the *hoist* is an instance of the same mechanism, applied to a helper summand already at a rank-1 frame rather than a source), Dugongue's #548 |
| 42 | +(observable pruning, frame-sensitive transports, Steinberg three-to-two cancellations: the *sink* is a transport of that kind) and |
| 43 | +rohanarun's #542 (observable slice and least-frame assignment, on the same word). Their published moves are lists of R480 gate |
| 44 | +numbers: the 60 two-x-source helper additions, the 227 zero-response gates of R480's dead totals, 20 transports and 7 |
| 45 | +cancellations by id. PR #555's word comes from the R420 L3 skeleton: it has no two-x-source additions and no zero-response |
| 46 | +gates, and its gate numbering is different, so the lists cannot be applied. `tune.py` therefore re-runs the mechanisms as |
| 47 | +searches on the word's elementary-operation form (one source per gate): single-operation transport with reframing under |
| 48 | +exact commutation and frame nesting, Steinberg 3→2 with frame re-choice, frame reassignment, and the pruning and least-frame |
| 49 | +steps. Run on the R480 word, the same code reproduces #541's local profile `{1:1260, 2:1320, 3:60, 4:180, 6:120, 8:9}` and |
| 50 | +#548's `{1:1257, 2:1330, 3:43, 4:163, 6:137, 8:9}` exactly. Run on PR #555's word to a fixed point, it applies exactly the two |
| 51 | +moves above; the pruning, Steinberg and least-frame steps have nothing to apply that lowers the moment. `apply.py --discover` |
| 52 | +repeats that search inside `verify.sh` and asserts that it returns this word. The greedy search is not claimed optimal. |
| 53 | + |
| 54 | +## Files |
| 55 | + |
| 56 | +- `moves.json`, `apply.py`: the two moves, pinned to the base word's content hash, and their application with an exact replay. |
| 57 | +- `bitE8mut-p111-tuned.json.gz`: the tuned word (rebuilt byte for byte by `verify.sh`). |
| 58 | +- `tune.py`, `wlib.py`: the search engines (numpy) and the exact replay and paid-profile code. |
| 59 | +- `fastc.py`: a quick exact pricer, `c = certify(288 × local paid profile + idle remainder, 64, 28416)`; the remainder is read off |
| 60 | + the base word's hosted audit. `verify.sh` checks that it equals the pipeline's c; it is not a substitute for the pipeline. |
| 61 | +- `verify.sh`, `summary.py`, `EXPECTED.json`: the checks and every number. |
| 62 | + |
| 63 | +## Checks |
| 64 | + |
| 65 | +```sh |
| 66 | +python -m pip install -r research/e8-bit-p111/requirements.txt |
| 67 | +bash research/e8-bit-p111/tune/verify.sh /absolute/new/output |
| 68 | +``` |
| 69 | + |
| 70 | +Python 3.11+ with numpy, sympy and scipy, a C++17 compiler (`CXX`) and Boost headers (`BOOST_INCLUDE`). The script rebuilds the |
| 71 | +base word from #555's frozen plan, applies the moves and compares both byte for byte, runs the independent replay, the interface lemma, |
| 72 | +the word-level and record-level mutation controls, the official legality and formal-column checkers at five, nine and eleven stages, |
| 73 | +the Gram witnesses, and both price pipelines on the tuned word, and asserts the nine-stage c exactly |
| 74 | +(`1733427039589844234749/10^24`) against `EXPECTED.json`. PR #555's own `verify.sh` is unchanged and still passes: it checks the |
| 75 | +untuned word and the controls (`MANIFEST.sha256` lists the files of this directory). |
| 76 | + |
| 77 | +## Credit and caveats |
| 78 | + |
| 79 | +PR #555 (eumemic) is the base: the punctured R439 word, its generator, checkers, pricing pipelines and package structure are |
| 80 | +used unchanged, building on #543 (R420 skeleton), #545 (punctured unit, DreamingOfClouds) and #510 (nine-stage chain). The |
| 81 | +mechanisms are from #541 and #548 (Dugongue) and #542 (rohanarun). The searches, `apply.py`, `verify.sh` and this note were |
| 82 | +prepared with Claude Opus 5.5 (Anthropic), via Claude Code; Apache-2.0. |
| 83 | + |
| 84 | +All retained assumptions of #555 apply unchanged (mutable-source realization, whole-call hosting, the all-size compiler, |
| 85 | +compressed windows, precision and tape contracts, the cited complex suppliers). The moves rewrite the phase-A writes of two |
| 86 | +helpers inside mini-node gates; `xcheck_p.py` and the official checkers verify the result exactly. This is a conditional |
| 87 | +finite construction, not a Lean proof, and it changes no κ. |
0 commit comments