{ "package": "Hexagon source submission for paper 011", "main_source": "011.tex", "main_source_sha256": "d57803d39fa50e3a2606147310c9d71bf1976d263e4d22783b9b032053ac6359", "classification": { "primary_MSC2020": "65K10", "secondary_MSC2020": "15B34" }, "contents": "Unchanged TeX, all 28 Python programs and compact text data. Every uploaded file is below 4 MiB.", "reproducibility": "Use the complete code011 distribution to rerun the full 27-check verification route. This reduced submission omits NumPy binary datasets and nested campaign archives. The saved reproduction.json documents the successful complete-distribution run, not a run on this reduced bundle.", "omitted_from_submission": [ "code011/README.md", "code011/data/catalogue5.npz", "code011/data/haar.npz", "code011/data/model_checks.txt", "code011/data/near_flat.npz", "code011/data/near_real.npz", "code011/data/phase.npz", "code011/data/polar.npz", "code011/data/polar_polish.npz", "code011/data/representatives.txt", "code011/data/stalled5.npz", "code011/daytime/replay-local.zip", "code011/daytime/replay-mac.zip", "code011/daytime/search-local.zip", "code011/daytime/search-mac.zip", "code011/overnight/ascent-compact.zip", "code011/overnight/stationary-compact.zip", "code011/overnight5/ascent-mac.zip", "code011/overnight5/ascent-pilot.zip", "code011/overnight5/stationary-mac.zip", "code011/overnight5/stationary-pilot.zip", "code011/requirements.txt", "code011/data/phase_grid.csv (regenerated by phase_grid.py; exceeds 4 MiB)" ], "text_documentation": { "README.md": "# Code for Numerical evidence for the almost Hadamard conjecture\n\nThe sole manuscript is `../011.tex`, compiled as `../011.pdf`.\nEvery mathematical statement and proof argument is visible in that PDF.\nThere is no optional mathematical appendix and no dependency on an earlier draft.\nEvery Python source below is mentioned beside its result or role in the PDF.\n\n## Recheck the paper\n\nFrom this `code011/` directory, using Python with assertions enabled:\n\n```sh\npython3 -m pip install -r requirements.txt\npython3 reproduce.py\n```\n\nSkip installation if NumPy, SciPy, mpmath and SymPy are already available.\nThe runner uses one numerical-library thread, runs 27 checks in a temporary\ncopy, and leaves the saved evidence unchanged. Its only persistent output is\n`data/reproduction.json`: versions, file hashes, timings, full check output and\npass/fail status. `python3 reproduce.py --list` lists its steps. The delivered\nreport records the measured runtime; timings depend on hardware and environment.\nThe manuscript certifies exact critical points near all 40 catalogue\nrepresentatives, including local uniqueness modulo phases, full Hessian inertia\nand at least 35 distinct nonflat critical values. The report records a fresh run\nof all 27 checks for paper 011, including the transpose comparisons added after\nreview. The manuscript retains the 30-page layout.\nGenerating campaign sources and returned archives remain unchanged.\n\nThis route checks exact identities and certificates and recomputes retained\nnumerical outputs. It does not rerun the large searches, reconstruct discarded\nordinary endpoints, prove criticality in every rounding box, or cover a unitary\ngroup. The report distinguishes exact, interval and numerical checks.\n\n## Manuscript map\n\n- The introduction explains the AHC and its parallels with the Marcus-de\n Oliveira and unistochastic star-shapedness problems, then presents the results.\n- Section 1 gives definitions, Fourier examples, the norm bound and the real\n reflection family, followed by the earlier design and circulant evidence.\n- Section 2 defines criticality and the modulus map, then gives the phase quotient,\n Hessian, pair tests, necessary polar condition, defect obstruction, and the\n proof through order three.\n- Section 3 proves the order-four phase-line and few-moduli results, including\n the reconstruction identities and finite classification of modulus patterns.\n- Section 4 gives the cubic curvature formula, rank-only counterexample,\n regular near-flat theorem, exact saddles, bounded modulus search and campaigns.\n- Section 5 gives the exact order-five calibration and its actual gradient\n branches, further seeds in orders five to seven, and the bordered construction.\n- Section 6 gives the completed order-five campaigns, weaker saddle, 40-group\n catalogue and 452 interval exclusions, retained stalls, completed daytime runs,\n the certified positive minimum of the stationarity residual, and the exact\n critical-point certificates for all 40 catalogue representatives.\n- Every program below is mentioned beside its mathematical or computational\n role in the PDF. There is no hidden mathematical appendix.\n\n## Complete Python inventory\n\n| Source | Mathematical or computational role |\n| --- | --- |\n| `reproduce.py` | Run and record the verification route |\n| `phase_grid.py` | Exhaustive fourth-root phase sign-balance enumeration |\n| `models.py` | Match exact order-four saddle formulas with stored representatives |\n| `three_lines.py` | Block deformation and interval curvature checks; imports `models.py` |\n| `four_lines.py` | Pair reconstruction, group identities and matching decompositions |\n| `four_moduli.py` | All 27 four-modulus patterns and exact elimination |\n| `curvature.py` | Sixteen cubic formulas, survey comparison and rank-only counterexample |\n| `flat_limit.py` | Exact leading and quartic coefficients for regular near-flat paths |\n| `moduli_search.py` | Bounded modulus-only exploration and `--verify-saved` audit; imports `curvature.py` |\n| `search.py` | Initial order-four stationary survey and polar controls |\n| `audit.py` | Interval exclusions for the initial survey and control refinement; imports `search.py` |\n| `ahc_ascent.py` | Frozen original order-four ascent source |\n| `ahc_stationary.py` | Frozen original order-four stationary source |\n| `compact.py` | Export or verify compact order-four campaign archives |\n| `audit_compact.py` | Recompute every retained order-four endpoint and spectrum; imports `compact.py` |\n| `check.py` | Four original order-five to seven seeds and interval/Hessian checks |\n| `ahc5_ascent.py` | Standalone order-five ascent campaign |\n| `ahc5_stationary.py` | Standalone order-five stationary campaign |\n| `test_ahc5.py` | Derivative, calibration, starting-family, retention and runtime tests |\n| `audit_ahc5.py` | Audit retained order-five endpoints and spectra |\n| `stalled5.py` | Refine all 288 retained stalls; certify a positive local minimum of the squared stationarity residual |\n| `catalogue5.py` | Group the 452 saved stationary returns, compare transposes, and interval-certify an uphill direction in each rounding box |\n| `critical5.py` | Certify 40 exact critical points, local uniqueness modulo phases, inertia, 35 distinct critical values and negative pair curvatures for group 39 |\n| `ahc5_day_replay.py` | Standalone daytime replay with complete unresolved-ID retention and improved follow-up |\n| `ahc5_day_search.py` | Standalone daytime exploration around the 40 catalogue representatives and structured starts |\n| `test_day5.py` | Replay regression, start-eigenspace checks, transaction, lock, standalone launch and pause/resume tests |\n| `audit_day5.py` | Audit daytime hashes, counters, unresolved IDs, starts, retained diagnostics and independently assembled full Hessians |\n| `weak5.py` | Interval uphill certificate for the weaker order-five saddle |\n\nNumPy alone suffices for the four campaign programs and their numerical audits.\nExact symbolic scripts need SymPy; interval certificates need mpmath; seed\nrefinement and the modulus search also use SciPy. `requirements.txt` covers the\ncomplete verification route; the reproduction report records tested versions.\n\n## Evidence layout\n\n- `data/`: initial starts/endpoints, diagnostic tables, interval certificate\n outputs, exact-identity reports, 300 modulus-search endpoints, order-five\n tests and the weaker saddle. The four original higher-order decimal seeds\n are embedded as data in `check.py` and identified in the PDF.\n- `overnight/ascent-compact.zip`, `overnight/stationary-compact.zip`: completed\n order-four campaigns, with all 95,016 unresolved endpoints and aggregate\n counts/extrema over every run. Each includes its generating source, summary,\n compact SQLite database, integrity checks and exporter source.\n- `overnight/terminal-checks.json`: the reported SQL query and extrema.\n- `overnight/archive-checks.json`: the original independent retained-data audit.\n- `overnight5/ascent-mac.zip`, `overnight5/stationary-mac.zip`: completed\n seven-hour Mac campaigns (18 November 2026), with 181,164,464 and\n 22,933,376 counted starts. The original compact archives are unchanged.\n Adjacent Mac audit reports check retained endpoints, counters and spectra,\n and compute geodesic curvature at the retained unresolved points.\n- `overnight5/ascent-pilot.zip`, `overnight5/stationary-pilot.zip`: five-minute\n local pilots, each including its original source, database, summaries and\n calibration. The adjacent audit reports record their independent checks.\n- `daytime/replay-mac.zip`, `daytime/search-mac.zip`: unchanged returned\n archives from the two seven-hour daytime campaigns, with 16,189,008 replay\n starts and 34,951,392 targeted starts. Adjacent audit reports check all 788\n and 569 retained runs, including 452 and 356 spectrum pairs and witnesses.\n The targeted run includes 8,737,840 ascent starts, all ending near flat.\n\nThe original archives are unchanged. Duplicate summary files outside those\narchives and pilot file-size bookkeeping have been removed. The superseded\n300-start calibration search and its two-start replay have been removed from\nthis package: they produced no additional mathematical example, and the same\ncalibration is reproduced by the exact model and initial survey. No proof,\nunique matrix, certificate or campaign archive was removed in this cleanup.\nThe generating campaign sources remain byte-for-byte unchanged for hash verification\nand compatibility with jobs already launched. Historical strings inside their\nfrozen sources are provenance, not instructions to find another manuscript.\n\n## Completed daytime campaigns\n\n| Final endpoint status | Replay | Targeted search |\n| --- | ---: | ---: |\n| Near flat | 1,696,013 | 10,068,469 |\n| Nonflat stationary, positive curvature | 13,672,742 | 24,303,512 |\n| Small-entry exclusion | 820,253 | 579,411 |\n| Unresolved or candidate | 0 | 0 |\n| Total | 16,189,008 | 34,951,392 |\n\nThe replay covers 70.59% of the original 22,933,376 starts; 6,744,368 were\nnot revisited. It is not additional independent evidence. Its 284,297 initial\nstalls are the stalls observed on this replay, not a certified subset count\nto subtract from the original 404,581. Follow-up turns 260,761 into numerical\nuphill saddles and excludes 23,536 by the entry bound. The targeted search\nhas 235,667 initial stalls: 161,713 become uphill saddles and 73,954 are excluded\nby the entry bound. No uncertain-Hessian, unitarity or numerical-error flag\nremains in either campaign. Entry bounds exclude computed endpoints, not\nuncomputed limits. Both runs encounter all 40 known invariant signatures;\nsignature matching is neither an equivalence proof nor an interval certificate.\n\nThe database sizes are 31.94 and 26.03 MiB. Reported sums of parent and worker\npeak memory are 128.5 and 157.2 MiB. All retained endpoints, spectra and\ncurvature witnesses pass numerical checks, including direct assembly of the\nfull geodesic Hessian independently of the campaign's assembly routine.\nDiscarded ordinary endpoints cannot be independently recomputed from these\narchives without rerunning their trajectories.\n\nAll 788 retained replay starts and 496 of 569 targeted starts regenerate\nwithin 1e-11 here. The remaining 67 soft-direction and 6 ascent starts depend\non eigenvector signs and rotations within repeated eigenspaces. The audit\nchecks the recorded perturbation length, Gaussian coefficients and eigenspace\nprojection energies (soft starts), or the recorded noise and membership in\nthe top eigenspace (ascent starts). These numerical compatibility checks\npass below 2e-12; they are not identical regeneration across numerical-library\nbuilds. The actual saved starts are preserved. `test_day5.py` tests acceptance\nof allowed basis changes and rejection of altered perturbation lengths.\n\n### Launch and resume instructions\n\nCopy just `ahc5_day_replay.py` onto the first Mac and `ahc5_day_search.py`\nonto the second. They are standalone and need only Python and NumPy.\nFrom the directory containing each file, run:\n\n```sh\npython3 ahc5_day_replay.py\n```\n\nOn the other Mac:\n\n```sh\npython3 ahc5_day_search.py\n```\n\nBoth default to seven cumulative active hours, two workers and one BLAS thread\nper worker. Use `--hours 6` or `--hours 8` to choose another total duration.\nCtrl-C waits for active batches and saves; repeat the command to resume toward\nthe original total duration. A completed seven-hour run does not start another\nseven hours; `--hours 8` extends it to eight total hours. The separate output\nfolders are `ahc5-day-replay-out` and `ahc5-day-search-out`, beside the files.\nProgress prints every minute. When finished, send **share.zip from each output\nfolder**; each includes its summary, compact database and exact source.\nPrevious campaign folders are not changed. Data growth stops at about 112 MiB,\nwith a 128 MiB database cap; review flags can also stop a run early. The summary\nprints the reason.\n\nThe replay revisits the original starting sequence and is not additional\nindependent evidence. It may cover only a prefix during the time budget; its\nsummary reports exact coverage of the original 22,933,376 starts. The targeted\nsearch embeds all 40 numerical catalogue representatives and needs no data\nfiles beside it.\n\nEvery initially or finally unresolved run ID and scalar diagnostics are saved\nin `followups`. `--replay RUN_ID` regenerates an individual case without counting\nit as another campaign start. Ordinary matrices use bounded samples; candidates\nand new invariant signatures are retained for review. Signatures are screening\ncomparisons, not equivalence claims. Small-entry exclusions concern endpoints,\nnot their possible limits.\n\n`python3 audit_day5.py path/to/share.zip` audits a returned archive; keep both\nnew launch sources beside that audit. `test_day5.py` checks the original replay\npath against the frozen stationary source, complete unresolved-ID tracking,\ntransaction rollback, starting eigenspaces, standalone launches, locks, interruption and exact resume\ncounts. Memory figures in `data/day5-tests.json` are local test measurements,\nnot measured Mac bounds.\n\n`data/catalogue5.json` records all 452 matches and interval certificates,\nthe 4 transpose pairings (leaving 36 numerical groups), and the equal rounded\nnorms but distinct minimum moduli of groups 15 and 16;\n`data/catalogue5.npz` contains the 40 representatives embedded in the launchers.\n`daytime/replay-local.zip` and `daytime/search-local.zip` are short local release\nchecks, separate from the older pilots and included as prefixes of the long\nMac runs, not additional starts.\n\n## Original order-five campaigns\n\nUse one standalone file on each Mac; only Python and NumPy are needed:\n\n```sh\npython3 ahc5_ascent.py\n```\n\nOn the other Mac:\n\n```sh\npython3 ahc5_stationary.py\n```\n\nDefaults are seven cumulative active hours, three workers and one BLAS thread\nper worker. Each writes its own `ahc5-...-out` folder beside the script and prints\nprogress each minute. Ctrl-C saves; repeat the same command to resume.\nA completed campaign does not silently restart. New campaigns need a new output\nfolder and preferably a new seed. Send `summary.txt` first, then `share.zip`.\nEvery start is counted; ordinary endpoints use bounded sampling, every candidate\nflag is retained, and review/error/storage limits pause the run for inspection.\n\n```sh\npython3 audit_ahc5.py path/to/ahc5-stationary-out/share.zip\n```\n\nKeep both original order-five scripts beside the audit. The completed Mac counts are reported separately from the five-minute local\npilots; the preliminary 3840-start tests are included in those pilots rather\nthan additional trials. The Mac archives retain 96 ascent and 1076 stationary\nendpoints. All-run extrema are aggregate records, not independently\nrecomputed evidence for every discarded endpoint. The 288 retained unresolved\npoints have positive numerical geodesic curvature. The follow-up `stalled5.py`\nrestarts all 288: 139 reach previously retained saddles, 134 return to one\npositive residual minimum, and 15 move near the zero-entry boundary. A length\n0.6 perturbation takes all 134 to another previously retained saddle.\n`data/stalled5.json` records every assignment and phase/permutation comparison;\n`data/stalled5.npz` retains refined matrices and curvature witnesses.\nThe rational Cayley-chart data are printed in the manuscript and embedded in\n`stalled5.py`. Interval contraction, Hessian positivity and phase-transversality\nchecks certify a strict positive local minimum of the auxiliary residual\nobjective. The numerical basin assignments are not interval-certified. This\nfollow-up does not resolve all 404,581 stalled runs or certify their limits.\n\n## Exact critical-point certificates\n\n`python3 critical5.py` verifies the 40 rational charts in\n`data/critical5-input.json` and writes `data/critical5.json`. Each chart has a\nself-adjoint rational centre with denominator 10^20, 16 specified basis\ndirections and radius 1e-12. Basis indices start at 1: the 5 diagonal matrices,\nthen the real/imaginary off-diagonal matrices in lexicographic pair order.\nInterval contraction certifies a unique critical point in the chart. A\ntransversality certificate promotes this to full criticality and local\nuniqueness modulo phases. An interval congruence check certifies the inertia.\nThe exact root is within 1e-7 of the stored representative after undoing the\nspecified rational unit scalar.\n\nAll 40 roots are nonflat. Their norms round to the manuscript's table and give\nat least 35 distinct critical values. Group 39 has inertia (15,9,1), and all\n20 row and column pair curvatures are less than -0.22. The certificate does not\nprove that all 40 groups are inequivalent or that the catalogue is complete.\n\nThe optional `python3 critical5.py --prepare` rebuilds proposed chart data from\nthe existing catalogue, using point Newton iteration. Proposal generation is\nseparate from interval validation and uses no random search. The default\nverification only needs NumPy and mpmath; preparing charts also uses SciPy.\n`--groups 39` selects one group. Run through `reproduce.py` to leave the saved\ninput and output evidence unchanged.\n\n## Completed order-four campaigns\n\nFor verification use `reproduce.py` or `python3 audit_compact.py`.\nThe old `ahc_ascent.py` and `ahc_stationary.py` are retained to identify the\noriginal experiment; their normal invocation repeats its full-recording\ncampaign. `--self-test` performs only the short calibration checks.\n\nIf an original large `results.sqlite` ever needs exporting again, run\n`python3 /path/to/project/code011/compact.py` from its containing output folder.\nThe database is read-only; an existing completed compact ZIP is not overwritten.\n`python3 compact.py --verify /path/to/ascent-compact.zip` checks an archive.\nThe complete compact project, kept in two locations, is the permanent research\nrecord; the original large databases add ordinary endpoints not selected for\nretention. No program here deletes those databases automatically.\n\n## Repeating searches deliberately\n\n`moduli_search.py --verify-saved` only checks saved endpoints; without that flag\nit reruns the bounded 300-start auxiliary optimization and rewrites its report.\nThe exact profile and polynomial checks do not run optimizers. Most individual\nchecking scripts write their own report files; `reproduce.py` isolates those\nwrites in a temporary copy. To regenerate the initial unitary survey separately,\n`search.py` accepts `kind count seed --output DIRECTORY`; the preserved initial\narrays are the evidence used by this paper. All original campaign seeds,\nsource hashes and environment metadata are inside their archives.\n", "data/model_checks.txt": "L=7.39103626009029 Haar id=16 max entrywise distance after equivalences=1.85e-13\nL=7.63761518684038 Haar id=3 max entrywise distance after equivalences=1.31e-14\n", "data/representatives.txt": "7.391036, haar #16, horizontal eigenvalues\n-2.613125929752752 -1.8477590650225786 -1.8477590650225744 -1.8477590650225721 -1.8477590650225713 -1.082392200292394 1.5307337294603605 2.2961005941899604 2.296100594191124\ncentral second difference at step 0.001 = 2.29609778479\n[[ 0.6532814824382+0.j 0.270598050073 +0.j\n 0.6532814824382+0.j 0.2705980500731+0.j ]\n [ 0.2705980500732+0.j 0.2705980500731-0.0000000000002j\n -0.6532814824382+0.0000000000001j 0.6532814824382-0.0000000000001j]\n [ 0.6532814824382+0.j -0.6532814824382+0.0000000000001j\n -0.2705980500731-0.0000000000001j -0.2705980500732+0.j ]\n [ 0.2705980500731+0.j 0.6532814824382-0.0000000000002j\n -0.270598050073 +0.0000000000001j -0.6532814824382+0.j ]]\n\n7.633239, haar #0, horizontal eigenvalues\n-2.1556677802309934 -1.7022082489112953 -1.7022082489112933 -1.6910338979283472 -1.4856012630353472 -1.1813906801931702 0.00708316889158305 0.007083168891592142 2.814969079786061\ncentral second difference at step 0.001 = 2.81495585774\n[[ 0.1581144044418+0.j 0.6400379589458-0.j\n 0.5316724772901-0.j 0.5316724772909+0.j ]\n [ 0.5316724772906-0.j 0.2938315643838-0.0000000000007j\n -0.2559173584027-0.5j -0.2559173584022+0.5000000000001j]\n [ 0.6400379589458+0.j -0.6462796876367-0.j\n 0.2938315643844+0.0000000000001j 0.2938315643834-0.0000000000001j]\n [ 0.5316724772904+0.j 0.293831564384 +0.0000000000007j\n -0.2559173584028+0.4999999999999j -0.2559173584023-0.5j ]]\n\n7.633377, haar #6, horizontal eigenvalues\n-2.133871887949056 -1.8171395232760121 -1.6191285130340296 -1.5408202112445677 -1.5309992819473486 -1.3093734231788206 -0.00712515082240087 0.1294065132242632 2.6601069935870942\ncentral second difference at step 0.001 = 2.66009888428\n[[ 0.2942128829967+0.j 0.2942128829967+0.j\n 0.6429920524228-0.j 0.6429920524227-0.j ]\n [ 0.5423963414973+0.j 0.2198484375512-0.4958431765952j\n -0.5123258051867+0.3332699175881j 0.1635466357584-0.1063877190437j]\n [ 0.5701326134673-0.j -0.5701326134673-0.000000000005j\n 0.0000000000018-0.4182688167435j -0.0000000000018+0.4182688167458j]\n [ 0.5423963414977+0.j 0.2198484375409+0.4958431766001j\n 0.1635466357585+0.1063877190461j -0.5123258051823-0.3332699175929j]]\n\n7.637615, haar #3, horizontal eigenvalues\n-2.014620674990665 -1.8805007778809302 -1.788219577868391 -1.7882195778683885 -1.0903376266113354 -1.0903376266113327 -0.14895660480947293 -0.14895660480946263 2.8768867273708385\ncentral second difference at step 0.001 = 2.87687823253\n[[ 0.6096097890109-0.j 0.4576665107339+0.j\n 0.4576665107339-0.j 0.4576665107339-0.j ]\n [ 0.4576665107339+0.j -0.2032032630037-0.j\n -0.2032032630036-0.5773502691896j -0.2032032630037+0.5773502691896j]\n [ 0.4576665107339+0.j -0.2032032630037+0.5773502691896j\n -0.2032032630037+0.j -0.2032032630037-0.5773502691896j]\n [ 0.4576665107339-0.j -0.2032032630036-0.5773502691896j\n -0.2032032630036+0.5773502691896j -0.2032032630036+0.j ]]\n\n", "requirements.txt": "numpy\nscipy\nmpmath\nsympy\n" }, "runnable_certificate": "From anc/code011 run python3 critical5.py. All 40 critical-point certificates can be verified using the included rational charts and catalogue5-matrices.json. The JSON preserves exactly the binary64 representative entries; only the input-loading branch of critical5.py differs from the complete package.", "other_standalone_checks": [ "python3 four_lines.py", "python3 four_moduli.py", "python3 flat_limit.py", "python3 check.py", "python3 weak5.py", "python3 three_lines.py" ], "dependencies": [ "numpy", "scipy", "mpmath", "sympy" ], "compact_package_verification": { "critical5": "All 40 certificates passed using the included JSON matrices", "seconds": 77.00236009299988, "distinct_nonflat_critical_values": 35 } }