Repository navigation
Conversation
Eleven notes under org/review/2026-11/science/threads/, with the programs and saved outputs for each under checks/. New and checked by separate code: all 2,288 realizations of the largest 11-vertex rigid graph are real, exact counts of 49 and 64 for eight and nine point charges, at least 2,418 sharper entries in the code tables for GF(7), GF(8) and GF(9), and 200 more proved recurrences. The gravity threshold fails against published data in its zero-parameter form and in its falling form. No research repository is changed.
Add the full working report of each of the eleven threads, a table of the outside sources they cached with sizes and hashes, and the small state files needed to resume or re-derive a result. A manifest lists the hash of every file in the threads folder.
…g of the pass Adds one table of 763 systems with the fits across all classes, a check that part of the velocity-scale result is forced by definition, and a baryon-only law. Corrects the early-type galaxy test (a wrong table column), with unchanged conclusions, and removes an inference the new fit does not support. Adds the working log, the audit of the four-speed coincidence and two structure studies of the code-table and recurrence results.
States and proves that the recurrence for grids with one qualifying pair is the square of the recurrence for grids with none, and uses it for a cheaper proof. All 211 recurrences of the first pass are proved again with the same starting index, and the 27 entries skipped for cost are proved.
… result Adds a study of where a light fermion fails to match the cluster data and what could explain each mismatch, with a halo model, a correction to the external-field sizes, and a map of particle mass against population. Adds the 12-vertex rigid graph: 3,904 of 6,180 real, certified, and a proof that this graph cannot raise the growth bound. Corrects an earlier statement in the 11-vertex note.
… tests Writes down, as a conjecture, the picture the threads suggest when taken together: one acceleration scale for galaxies and one light fermion capped by the exclusion principle. Gives its numerical predictions, six observations that would kill it, and the mathematical links between the threads.
…of the charge configurations Traces, for every certified point-charge configuration, which pair of charges each index-2 balance point separates. The loop count of that network equals the index-1 count in all eight cases. Maxwell's bound is the count of a network with exactly one wall per pair, and the configurations that beat it put several walls between the same pair. Adds the links to the planar lensing bound and the tests that follow.
Measures the share of real realizations near the all-equal point and far from it for the largest graphs on 6 to 11 vertices. Realizations are far more often real near it, random directions are almost never all real, and all-real directions exist as thin open cones. The 880 record lies on such a direction. Also corrects one count in the code-table note.
…urations Adds the note four-charges-eleven.md and checks/four with both certificates, the three checkers' outputs, the controls and the programs of both routes.
…e flat configurations
… points, flat-family survey, synthesis and log brought up to date
…ts for the octahedron
… and proved lower bounds for 9 and 21 charges; push from 11 finds nothing higher
…), a faulty earlier check withdrawn
|
phew what an anthology by fable 5.2 here |
orda-aska
left a comment
There was a problem hiding this comment.
Request changes.
The review re-ran the exact checkers across all threads, then attacked the surviving claims with fresh corrupted inputs. Every proved count that was re-run reproduced, with matching box counts and centres. The full record sits in the addendum draft, including every finding word for word in findings.json.
Five items block the merge. One gives a number the shipped data contradicts. Two are verdict and checker claims stated wider than the shipped programs support. One proof cites hand derivations that are not in the record. One set of threads cannot be re-run from this checkout at all. Each blocking item has an inline comment at its line.
Commit a61e289 landed during the review. Its changed programs were re-run. The new stack_check.py proves 41 points at D=6 in exact arithmetic under -O, and levels_check.py now raises on a mismatch. That part of item three is closed. The rest of that item stands.
Major items beyond the inline comments: REPORT.md is cited for two derivative bounds and is absent, though both bounds re-derive correctly. The input certificates for three to seven charges are absent and had to be rebuilt. Two of three exact checkers for 2,288 share one proof template. The fermion window fits and tests on the same 51 clusters, with near-zero per-system correlation in its own saved test.
Confirmed by re-run: 11 balance points for four charges in three configurations, counts 4, 9, 24, 29, 44, 49 and 64, lower bounds 26 and 87, 2,288 real realizations, 1,519 code certificates, 211 recurrences, at least 20, 42 and 67 central-configuration classes, lattice results in dimensions 2 to 9, and a clean 1,437-file manifest.
The corrected verifier in the point-charges give-back also passes all three eleven-point certificates under -O. That closes the second exact program in practice. Nothing inside this pull request changed for it.
Merge order in the body holds. The three earlier review pull requests carry two cited checkers, the verdict words, and the thread list, with zero file overlap here.
|
|
||
| - [F] For 8, 9 and 10 vertices a directed climb found directions u whose whole ray is all real, at every eps tested from 0.001 down to 0.00001. For 10 vertices there are at least two, 126 degrees apart. | ||
| - [F] The certified 880 point lies on such a ray. Along its direction the count is 880 of 880 at every eps from its own 0.033 down to 0.00001. | ||
| - [F] Along the direction of the published lengths (Bartzos and co-authors) the count is 868 at every eps tested. |
There was a problem hiding this comment.
Blocking. certray_868.json in this pull request records 856 at eps 0.01 and 860 at eps 0.003 along this direction. The value 868 appears at eps 0.00188 and below. As written, the sentence disagrees with the shipped data. The direction conclusion survives on the smaller eps values.
| - [F] A brute-force count that shares no code with the machine builder matches the machines and the OEIS data for small n. | ||
| - [F] A third program, written for this review without reading the other two (`checks/oeis/spot_proof.py`), builds its own machine straight from the OEIS wording. It proves 46 of the entries, one from each combination of neighbour kind, alphabet, condition and orientation. All 46 give the same starting index as the certificates. | ||
|
|
||
| Verdict: **VERIFIED** for all 211. A later study (`square-law.md`) proved every one of them again by a different argument with separately written code, and the starting index agrees in all 211. The 46 above have a third proof. |
There was a problem hiding this comment.
Blocking. This line gives VERIFIED for all 211. README.md gives VERIFIED for 46 and REPRODUCED for the rest, and square-law.md adds 27 PROVEN. PROVEN is defined only outside this pull request. The three labels cannot all stand. The README label is the one the re-runs support.
|
|
||
| ### 6. Checkers that pass when they should fail | ||
|
|
||
| [F] The repository verifier for the point charges prints VERIFIED on a corrupted certificate when Python runs with `-O`. The science pass found the same fault in the gravity scripts and in vitruvian3d. That is three repositories. Every checker written in this pass raises an error instead of using `assert`, and was run under `-O` against corrupted input. |
There was a problem hiding this comment.
Blocking in part. Commit a61e289 fixed levels_check.py and added stack_check.py, and both now raise on failure. This sentence still covers print-only programs such as laman_count.py. It also covers verify_arb.py, which the four-charge note states ran without -O. The universal form remains wider than the record.
|
|
||
| ## The structure of the problem | ||
|
|
||
| All of these are proved by hand in the thread's notes, `checks/three/NOTES.md`. |
There was a problem hiding this comment.
Blocking. NOTES.md at the cited path holds one-line lemma statements. The derivations behind this sentence are not in the record. A reviewer rebuilt the valley formulas and the injectivity argument and found both sound. Until they are written here, the proof is not self-contained as archived.
|
|
||
| ## Rerun | ||
|
|
||
| Programs and saved outputs are in `checks/gravity/`. `t1`, `t1b` and `t2` read the SPARC tables from `astroweb.case.edu/SPARC/` and the TeX source of arXiv:2109.03828. `t3` to `t6` read the published tables named at the top of each script. `model.py` re-implements the model without importing the repository. |
There was a problem hiding this comment.
Blocking. The programs named here read a sources tree that STATE-NOTES.md excludes from this checkout. Some paths predate a folder move, and n4_window.out has no program at all. That output is the only support for the 9 eV end of the fermion window. A reviewer working from this checkout cannot re-run these threads.
| zeros = len(res) - (last_bad + 1) | ||
| start_n = (last_bad + 1) + d + off | ||
| if zeros < D: return f"NOT PROVED: only {zeros} consecutive zero residuals, need {D}" | ||
| claimed = cert.get("proof", {}).get("start_n", cert.get("start_n")) |
There was a problem hiding this comment.
Major. The value computed on this line is never used, and the key it reads does not appear in the certificates. The matching starting indices in the note were confirmed by a reviewer's own count, with zero mismatches. This program does not assert them. It can also exit 0 after choosing zero entries.
|
|
||
| - [F] Every class is proved to exist, to be the only solution in its box and to be non-degenerate. The classes are proved different from one another. | ||
| - [F] The Morse index and the exact symmetry group of every class are proved as well. | ||
| - [F] Two programs that share no code accept every certificate, one in ball arithmetic and one in exact rational arithmetic. Both reject 15 kinds of corrupted input, also under `python3 -O`. |
There was a problem hiding this comment.
Major. The corruption suite holds seven kinds for the rational verifier and one kind for the ball-arithmetic program, across 15 run lines. The second program never sees six of the seven kinds. reports/cc.md states the count in that form. This sentence gives both programs all 15.
|
|
||
| ## A halo with no free normalisation | ||
|
|
||
| [F] The halo study computed the largest halo a fermion can make: an isothermal Fermi gas at the gas temperature, in each cluster's measured field, with its central occupation at the maximum a thermal relic allows. The particle mass needed for that halo to hold the extra mass is 4.7 to 5.9 eV in the median system and 8 to 17 eV for 95% of systems, depending on the corrections. The hardest systems are cool groups. |
There was a problem hiding this comment.
Major. The saved halo output gives a maximum of 100.00 eV in three of four correction sets. That value is the search ceiling, reached by one system. The 95% framing in this sentence passes over it. The window as stated holds only after that system is set aside or heavily corrected.
| | Strand | Where | Status | | ||
| |---|---|---| | ||
| | Galaxies follow one relation with one acceleration scale, up to redshift 2.5 | `gravity.md`, `acceleration-evolution.md` | [F] holds in every sample tested | | ||
| | That scale does not grow with the expansion rate | `acceleration-evolution.md` | [F] the one fit possible leans against growth, not decided | |
There was a problem hiding this comment.
Major. This row marks the first claim with the thread's own words, leans against, not decided. The strength table above it in this file and in convergences.md tags the same claim [F]. The conjecture frame at the top of this file is clear. The per-strand tags read stronger than that frame.
|
|
||
| **Step 4. Charges.** Start from a point that is flat to order d - 1. Small changes of the charges set the terms of degree 1 to d - 1 to any small values, here the stacked levels of step 3. The balance points near the flat point are then the critical points counted in step 3, once the scale is small enough. | ||
|
|
||
| [I] With about d²/3 charges and about d³/6 balance points, the count grows like 0.87 n^(3/2). |
There was a problem hiding this comment.
Major. The exponent here comes from the four-step sketch, and the note labels it as one. The constant 0.87 and the crossing near 160 charges come from the solved flat family at about d squared over 3 charges. That family is numerical. No existence proof for general d sits in the record, so the constant carries more precision than its support.
…the checkers support Addresses review items on pull request 58: the 868 ray sentence now matches certray_868.json, ray statements name the eps values tested, the 2,288 claim is stated as a proved lower bound with the completeness step cited, the checkers are described by method, and codes.md separates what the checkout re-derives from what rests on the saved tables.
…ted to match them Adds checks/three/DERIVATIONS.md with the hand proofs the review found missing (planarity, index count, inverse map, minimum criterion, the local lemma, and the mathematics of the certificate). The note now says what each kind of box certificate proves and what the 55.1 million counts.
…ences against data Counts the 119 findings of the team review by kind, layer and direction (classify.py). No proved result failed; 46 of 68 problems are in prose and 41 of 42 directed errors overclaim. Adds check_claims.py, which fails when a listed sentence leaves its note or its data file stops supporting it.
…ifier and input certificates in the folder) Adds checks/charges/LEMMAS.md with the proofs of the two derivative bounds and of each step of the exact count, the corrected repository verifier, the five input certificates for three to seven charges and a rerun script. The notes say which corrupted inputs are rejected and name the program's known limits.
…-run, the third program checks what its note says All 238 square-law proofs now pass the verifier (A269766 added, two more moved to the dense engine). spot_proof.py compares starting indexes, fails on empty input and no longer drops an entry. The notes label 46 entries VERIFIED and 192 REPRODUCED, with the reason. Ledger entries added for these counts.
… table replaces the universal checker claim README, convergences, growth, central configurations, the wall table and the flat-centres note are corrected against the review findings. scan_checkers.py writes checkers.md from the source of each checking program. The verdict words are restated in the README. Ledger entries added for the corrected sentences.
… checkout; fermion note states its circularity fetch_sources.py downloads the published tables with recorded checksums and check_all.py re-runs each program and compares outputs byte for byte. The missing program for n4_window.out is added and old paths are fixed. neutrino-like.md says the construction is circular by design, quotes the per-system test that fails, and names the system at the search ceiling. fabric.md and convergences.md marks now match the thread verdicts.
…erns recur critical-circles.md answers an open avenue of the review's exploration note: an axisymmetric harmonic polynomial of degree k+1 has exactly k critical saddles in the meridian plane, so coaxial rings give about 3 balance points per charge. framework.md names the four tools that recur across the threads, what was put in by the choice of threads, and two rules that could be tested.
…rt sentences, a bit-order comment)
…and to check the folder The two programs of the review's exploration note reproduce, and its code-table counts are recomputed from this folder's files. verify_folder.py checks the manifest, the claims ledger and the checker scan in one run. check_all.out is the author's own run of the gravity and neutrino checks: 41 passed, 1 skipped.
…o 193 sentences An audit of README, convergences, LOG and SYMBIOSIS found 12 more places that said more than the files, close to the 11 the overlap of review passes had predicted. All are corrected. The three 12-vertex checkers were run again and their outputs saved, since the folder held none. review-meta.md records the test of its own estimate.
…the literature findings The author of #58 answered the review's blocking and major items in twelve commits while this branch did the same. Their version of every file in threads/ is taken as it stands, and this branch's own fixes and its second three-charge write-up are dropped. Both write-ups found the same two over-claims in three-charges.md. What remains in threads/ from this branch is what their repeated literature search did not find, read from the full texts: - the 2007 remark of 11 effective cells at the regular tetrahedron for a large exponent (Gabrielov, Novikov, Shapiro), now cited; - arXiv:2501.05315 described as it reads (11 co-located, with a minimum, not realised), not as "three saddles" only; - "for the Coulomb potential" in the novelty sentences; - Theorem 1.1 of arXiv:2609.03965: at most 570 for four charges; - the source of the rounding sentence, the later checker, and the T(4,4,4) identification. checks/verify_folder.py passes: 1,481 files, 193 of 193 ledger claims. pr58-trust/CHECKER-AUDIT.md is removed: the author's generated checker table replaces it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012tWK9X51ZybvPt7yCm7Zbt
|
Independent review of PR 58 at I rebuilt the four charge certificate from the potential definition with my own solver. I also re-ran the shipped exact lower bound prover. Both give 11 distinct nondegenerate balance points, 7 of index 2 and 4 of index 1. The index difference is 3, equal to four charges minus one. Four shipped certificate hashes match the note. A third, standard library only exact checker is offered in PR 61. I ran it locally and it proves exactly 11 with no floating point. Three independent routes now agree. The local multiplicity at the tetrahedron centre is 11. My exact local algebra gives graded dimensions 7, 10, 11, 11, 11, 11, 11 for the quotient at the origin. This holds with and without the degree 6 term. The shipped PR 61 identifies the point as T(4,4,4) with Milnor number 11. It explains the earlier figure of 27 as a global quotient dimension with 16 simple points away from the origin, leaving 11 at the centre. The 11 versus 27 discrepancy is resolved. Status of the five earlier blockers at this head:
I re-ran the shipped claim ledger, the folder verifier and the corrupted sentence control. The ledger passes 193 of 193. The folder verifier passes all three checks. Restoring the old false sentence makes the ledger fail 1 of 193, as intended. Two limits and one open item remain. The growth claim stays a lead, not a theorem. For 21 charges the shipped output proves at least 87 points near the flat centre. Its index difference is 7 while 21 charges require 20. The note says at least 13 more points sit further out. That step needs every balance point to be nondegenerate, or a local degree version of the index sum. PR 60 reviews Two sentences in I confirmed both against the primary sources. The Coulomb result itself is not contradicted. These are suggestions for the author, not blockers I would raise on the code. I have no remaining merge blocker at
|
Follow-up to work package F7 from issue 50. Draft. A person merges.
Opened through the zero-aska account. An approval from this login cannot count for this work.
What this adds
org/review/2026-11/science/threads/: what came out when each thread listed innew-ground.md(#56) was pulled as far as it would go, and what the first results led to. Twenty-three notes, a log, a sources list, a record of what went back to each source repository, and the programs and saved outputs for each underchecks/. Start withREADME.md.Nothing outside that folder changes. No research repository is edited by this pull request.
The main result
Four positive point charges with exactly 11 balance points. Maxwell's bound for four charges is 9, and the counterexample published in July 2026 starts at five charges.
four-charges-eleven.mdwith its eleven points.indep_count.pyand the point-charges repository's own verifier. A third program, written for this result, proves at least 11.A second result, at the level of an argument
Balance points can grow like n^(3/2) with the number of charges n. Every construction known before grows like 10n.
growth.md.Other results
l2288.mdrigid-realizations.mdpoint-charges.mdcodes.mdrecurrences.md,square-law.mdincentre-lattices.mdgravity.mdacceleration-evolution.mdcentral-configurations.mdthree-charges.mdflat-centres.mdrelational-geometry.mdunit-point.mdneutrino-like.md,fabric.mdconvergences.mdindexes the patterns found across threads, including three that did not survive.SYMBIOSIS.mdlists what was returned to each source repository as draft pull requests and issues.Response to the team review
The review asked for changes on five blocking items and a list of major ones. Every proved result it re-ran reproduced. All five blocking items and the major items below are fixed on this branch. The review's own record and its exploration note are in pull request 60, whose base is this branch. The two should land together.
unit-point.mdnow gives the counts in the data file, value by valueREADME.mdchecks/claims/checkers.md, a table written by a scan of the 22 checking programschecks/three/DERIVATIONS.mdhas the full derivations, and the note is corrected to matchchecks/check_all.py: 41 jobs reproduce byte for byte, 1 long fit is skipped by defaultchecks/charges/LEMMAS.md, the five input certificates and the corrected verifier are in the folderneutrino-like.mdsays so, quotes the per-system test that fails, and names the outlierNot fixed, and said so in the notes: the saved code-table pages and one outside library are still not in the folder, and three checks cited from the science pass are in another open pull request.
review-meta.mdcounts the review's 119 findings. No proved result failed. 46 of the 68 problems are in prose, and 41 of the 42 with a direction claim more than the evidence.In answer,
checks/claims/check_claims.pynow tests 193 listed sentences against the data files beside them. An audit of the summary pages against the data then found 12 more sentences that said too much, and those are corrected too.checks/verify_folder.pyruns the manifest, the ledger and the checker scan in one command.The exploration note's first avenue is closed by the verifier run above. Its third is answered in
critical-circles.md: rings on one axis cannot beat a flat point.framework.mdtakes up why the same patterns recur.How it was checked
Where a result is marked VERIFIED, a program written apart from the one that found it accepts it:
checks/four/verify/lower_bound_check.py, plus the two exact programs, both now in the folder. Controls fail as they should.checks/growth/verify/stack_check.pyproves the stacked counts for three to seven levels,levels_check.pysearches the one-level counts, andlower_bound_hp.pyproves the two lower bounds.checks/oeis/spot_proof.py. The square law has a written proof, and all 238 proofs pass the thread's verifier.checks/charges/indep_count.pyreads only positions and charges.checks/claims/checkers.mdlists the 22 checking programs. None has a bareassert, all can stop with an error, and 15 have a corruption test saved in the folder.For the owners
Nothing here changes a claim label or a published record. These items need an owner's call:
SYMBIOSIS.mdeach carry their own list of decisions.Size
1,482 files, about 28 MB. Most of it is certificates and saved search states. No downloaded tables or pages are included.
Merge order
After #55, which holds the verdict words, and #51, which holds two check programs that
codes.mdandrigid-realizations.mdcite. The review asks that those and the thread-list pull request land first. This one and the review addendum then land together.Replaces: no pull request.