Active project
TS-Benchmarks
An audit-first falsification harness for graph relaxation, hard reasoning, matched model comparisons, and compute experiments.
What problem it addresses
A large family of prototypes can generate anecdotes faster than it generates comparable evidence.
How it works
Define fixed generators, baselines, acceptance thresholds, receipts, replay paths, and visible failure categories before interpreting results.
What exists now
Useful falsification infrastructure that needs to become the common evaluation route for Ten-SON, TSQ, and verifier integration.
Evidence / receipts
The repository includes graph generators, graph relaxation tests, scale-free failure evidence, hard-reasoning work, matched TensionLM campaigns, and GPU/Triton microbenchmarks.
Limits
Coverage is still fragmented. A benchmark harness only helps when active projects adopt it and pre-register what would count as success or failure.
Next milestone
Standardise experiment manifests and acceptance criteria across the active project stack.
How to run / inspect
Install / open
git clone https://github.com/BoggersTheFish/TS-Benchmarks && cd TS-BenchmarksInspect
Start with the generators and explicit failures, then compare project claims against fixed seeds and matched baselines.
Links
Focus tags
Related receipts