Datasets:
id stringclasses 7
values | kind stringclasses 5
values | certificate_json stringclasses 7
values | provenance stringclasses 1
value | training_admission bool 1
class |
|---|---|---|---|---|
finite-reference-0 | heat | {"claim":{"sum":"1/1","weights":["1/1","0/1","0/1","0/1","0/1"]},"inputs":{"r":"0/1"},"kind":"heat","schema":"grh.workbench.certificate.v1","scope":"Exact five-point stencil coefficients; not floating-point execution or a continuum PDE proof."} | constructed_exact_mathematical_fixture | false |
finite-reference-1 | heat | {"claim":{"sum":"1/1","weights":["1/2","1/8","1/8","1/8","1/8"]},"inputs":{"r":"1/8"},"kind":"heat","schema":"grh.workbench.certificate.v1","scope":"Exact five-point stencil coefficients; not floating-point execution or a continuum PDE proof."} | constructed_exact_mathematical_fixture | false |
finite-reference-2 | heat | {"claim":{"sum":"1/1","weights":["0/1","1/4","1/4","1/4","1/4"]},"inputs":{"r":"1/4"},"kind":"heat","schema":"grh.workbench.certificate.v1","scope":"Exact five-point stencil coefficients; not floating-point execution or a continuum PDE proof."} | constructed_exact_mathematical_fixture | false |
finite-reference-3 | rank | {"claim":{"p":"1/4096","resolution":"1/4096"},"inputs":{"draws":4095,"exceedances":0},"kind":"rank","schema":"grh.workbench.certificate.v1","scope":"Exact finite Monte Carlo rank; probability validity additionally requires exchangeability of the complete frozen search statistic."} | constructed_exact_mathematical_fixture | false |
finite-reference-4 | risk | {"claim":{"slack":"7/625","upper_bound":"97/2500"},"inputs":{"alpha":"1/20","drift":"1/100","per_test":"1/10000","tests":288},"kind":"risk","schema":"grh.workbench.certificate.v1","scope":"Exact union-bound budget with a supplied null-model discrepancy allowance; individual event bounds and discrepancy are assumptions.... | constructed_exact_mathematical_fixture | false |
finite-reference-5 | gain | {"claim":{"row_slack":["0/1","0/1"]},"inputs":{"matrix":[["1/4","3/2"],["3/32","1/4"]],"rate":"5/8","weights":["4/1","1/1"]},"kind":"gain","schema":"grh.workbench.certificate.v1","scope":"Exact weighted comparison-matrix contraction; the model-to-matrix comparison remains an assumption."} | constructed_exact_mathematical_fixture | false |
finite-reference-6 | mutation | {"claim":{"matrix":[[0,-1],[1,0]],"seed":["3/1","5/1"]},"inputs":{"matrix":[[0,1],[-1,0]],"seed":["2/1","5/1"],"vertex":0},"kind":"mutation","schema":"grh.workbench.certificate.v1","scope":"Exact skew-symmetric exchange-matrix and positive rational seed mutation; not a categorical or spectral equivalence certificate."} | constructed_exact_mathematical_fixture | false |
Certified Surface Computation — physical graph workbench
Associated project: https://github.com/amelie-iska/GRH-research Interactive surface: https://amelie-iska.github.io/GRH-research/workbench/surface.html Volume: Certified Surface Computation
One program, inspectable realizations
This release connects dessins, integral orders, gentle relations, BGG structures, graph-to-graph computation and quantum realizations through explicit maps and independent evidence. It includes the expanded 61-page volume; the exact authored research-source archive; complete finite graph-table and Boolean-circuit Lean proofs; finite spectral, stochastic-channel and half-density constructions; a linked 18-qubit interactive explorer; and actual measured-data coding experiments.
The source archive preserves 475 named claim occurrences in 23 documents at pinned
commits. Its review register distinguishes preserved text from individually
reviewed statements and formalized results. Twelve named source occurrences have
explicit review decisions; 463 retain source-port status. The separate requested
repository name amelie-iska/TropicalGT-Program returned 404 through both public
and selected-connector access. No substitute repository is assumed.
Measured records and trained components
The original 1,648 NOAA-derived graph windows remain unchanged: 1,015 train, 316 validation and 317 test. The new speech and gravitational-wave configurations each contain 2,400 windows per split, adding 14,400 measured-signal graph windows. Every new row is checked against raw source context, target, direction, quantizer and group support. There are no duplicate target identities or cross-split raw support overlaps. Repeated physical values are not discarded as fake duplicates.
Speech uses six utterances each from three distinct LibriSpeech speakers, with speaker-separated training/validation/test. The native PCM16 representation stays at 16 kHz. All 18 selected audio identifiers have matching chapter-specific transcripts; transcripts are provenance, not predictor input. This local re-split is not an official LibriSpeech ASR benchmark.
The GWOSC control trains on H1 0–8 seconds, validates on H1 8–12 seconds and tests on L1 24–28 seconds relative to GPS 1126259446. The known GW150914 event interval is excluded from fitting. Its int32 coding view uses 1e-24 strain units, with quantization error recorded. One detector-transfer interval is not a population of independently held-out gravitational-wave events.
Sixteen original native Mora-3 encoder component runs and eight new runs are reported, with actual pinned TokenGT/tropical modules, width 16, one layer, seeds 17/29 and 80 optimizer updates. Checkpoints, probability arrays and all outcomes are retained under their source-specific rights. These are component studies, separate from the full approximately 80M Mora-3 model and the existing campaign. Retrieval is a trained-embedding adapter, not the entire native forest head.
Complete coding costs and controls
evidence/physical-controls/ contains actual arithmetic-coded packets with frozen
models, exact residuals, segment boundaries, initial samples and termination.
Independent decoding recovers every sample of the declared integer representation.
Actual packet BPB is 6.4216 for held-out speech PCM16 and 7.6132 for the strain
fixed-point view. General-purpose zlib/lzma baselines perform better in these
whole-stream comparisons; those outcomes are preserved. Sampled-window neural
likelihoods use a separate exposure and representation contract.
The count baseline wins the speech symbol-likelihood comparison. The native encoders beat the selected count baseline on the chosen detector-transfer interval, with substantial seed variation. A simple known-event correlation control does not isolate the event from its off-source windows. These results test specific models and protocols; compressibility is not itself evidence of intentionality. Known speech supplies communicative provenance independently of its code length. Kolmogorov complexity is not computed; code lengths are operational proxies.
Formal and numerical evidence
The release audit records 107 regression methods under normal Python and python -O, 48 new Lean declarations in six modules, fresh checks of the prior four risk and six BGG statements, and five axiom-free finite payload replays. Accepted formal logs have no warnings, no sorryAx and no unexpected foundational axioms.
Nine surface-controller configurations agree across independent NumPy, browser and actual cuStateVec execution. The nonlinear triad has 15 state/Jacobian/energy checks; its periodic half-density discretization has second-order temporal convergence and six actual four-qubit cuStateVec matrix-application checks. The linked browser interface passes 30 local checks, including all five local kernel contracts, URL restoration, mobile layouts and stale-result rejection. GPU simulations and constructed mathematical fixtures are identified separately from measured physical records and quantum-hardware experiments.
Reproduction and history
reproduction/ is the current workbench source. reproduction-project-dependencies/
contains the exact small project dependencies used by its tests. native-mora3/
contains the two source-pinned model modules. Use materialize_reproduction.py to
lay these out in a fresh directory; it does not install packages or read credentials.
See REPRODUCE.md and reproduction/VALIDATION.md for commands and prerequisites.
proof-port/ contains originals, exact statement/proof records and the review
ledger. documents/source-proof-port.zip is a compact version of that archive.
The earlier 37-page PDF, measured results and prior evidence remain unchanged;
revision 84aa1e7241e67c9b85e31d2e45400d3f285c001d is preserved. Its original card,
catalog and manifest are also recorded under history/84aa1e7/.
The continuum Yang–Mills and GRH/Hilbert–Pólya routes are stated as precise conditional transport criteria. Their analytic/arithmetic premises are not manufactured from finite examples. The finite order, spectrum, channel and half-density results have their own proofs and executed checks.
Attribution
NOAA GML/PSL, SILSO Royal Observatory of Belgium, PhysioNet MIT-BIH, Project Gutenberg, LIGO Scientific Collaboration/Virgo Collaboration (GWOSC GW150914, DOI 10.7935/K5MW2F23), and Panayotov, Chen, Povey and Khudanpur (LibriSpeech, ICASSP 2015, OpenSLR12). Consult LICENSES.md and source manifests for exact terms.
- Downloads last month
- 107