Hi SambaNova community,
We’ve released TESC, a framework for deterministic cognitive control in LLMs via structured outputs (JSON+schema) and semiotic configuration. We validated with your OpenAI‑compatible function‑calling API on a small programming review/repair bench and report structure metrics.
Highlights:
- 100% tool_call + schema_valid on SambaNova (n=8), arg_coverage=1.00, extra_keys=0.0%, parse_err=0.0%.
- Programming micro-bench: improved coverage/actionability and repair pass rates.
- Semiotic Uncertainty and Dynamics with CIs.
- Code (public repo): GitHub - Amawta-labs/TESC: TESC: deterministic cognitive state control in LLMs (benchmarks + evals)
- Preprint PDF: LINK
- Blog: https://amawtalabs.com/blog/tesc-control-llms
Quick reproduce (SambaNova):
- export SAMBA_API_KEY=…
- poetry run python benchmarks/run_programming_bench_sambanova.py
- poetry run python benchmarks/eval_programming_bench.py bench_runs/programming_bench_samba/
Thanks to SambaNova for the OpenAI‑compatible function‑calling interface and developer support that enabled cross‑model
validation.