Evidence tied
to the source.
The current G suites, application-model results and historical scale records are distinct evidence sets. Figures here are reported by the owner’s release account, not engine tests rerun for this publication.
14,935 test IDs.
Two tested environments.
| Environment | Passed | Skipped | Additional |
|---|---|---|---|
| Mac ARM64 / Python 3.14 | 14,894 | 41 | 34 passing subtests |
| Existing Brev Linux ARM64 / Python 3.10 | 14,920 | 15 | 1 warning; 34 passing subtests |
Do not add the host totals and call them unique tests. Clean offline installed-package checks exercised all 82 profiles, both tested MCP protocol eras and fresh answer checking, retaining declines and interruptions.
Different counts.
Different meanings.
| Measure | Reported count |
|---|---|
| Problem-domain families | 13 |
| Profiles | 82 |
| Entry functions | 65 |
| Python exports | 133 |
| CLI operations | 38 |
| Stdio MCP tools | 9 |
| Runtime/data files | 260 |
Airline planning: 1,201,500.
For the original flugpl model, a separate proof established matching objective and global bound of 1,201,500 with gap zero. Its proof-work allowance was 50 million, above the default five million. Those are model units, not airline savings.
Read all eight application outcomes221 locations.
28 proved model batches.
The LA experiment reports checked global optima for all 28 declared batches. Modeled red-exposure proxy: 17.067 → 9.101 seconds, a 46.67% reduction against an assumed equal split. Campaign wall time: 30.74 seconds.
Not measured LA congestion reduction: 2013–2016 counts, assumed cycles and independent locations. Queues, capacities, turns, pedestrians, offsets and spillback are absent. External repartitioning enabled proof completion; it is not an automatic-router fix.
Large input.
Explicit result scope.
An explicit one-billion-variable, three-billion-edge graph used 72 GB of original arrays and approximately 17.39 GB peak CUDA allocation. Initial solve including validation/recount: about 406 seconds. Separate replay: about 73 seconds. Continuation improved the checked objective by approximately 4.425%; global optimality remains unproved.
The 200-billion records distinguish sparse namespaces, implicit structured proofs and a materialized 25 GB packed state. The temporary state was deleted after validation; evidence is retained. One physical GH200 was used, not physical multi-GPU scaling.
Selection is implemented.
A universal learned gain is not.
The retained routing study measured 186 actual dispatch outcomes. In one comparison, probing-only completed 39/42, while the then-packaged learned default completed 37/42 and was slower. The general learned promotion was not made. A separate linear-capability selector also did not show held-out improvement.
Garmo Solvers Router’s separate positive portfolio result must not overwrite these UNISOLVE records.
Keep the old records
under their old identities.
Earlier 78-profile, candidate-D and 0.13 descriptions remain historical. Current G figures are not retroactively assigned to those builds. Optional numerical/SMT/CUDA integrations retain their dependencies and trust boundaries. The source does not establish universal source-language conformance, physical multi-GPU speedup or a hosted multiuser security sandbox.
Review runtime scope