Frequently Asked Questions

No. Tanvelo is strictly a diagnosis and recommendation engine. We never alter your weights, modify your codebase, or change your prompt templates automatically. You receive concrete diagnostic reports and exportable datasets, leaving you in 100% control.

All you need is your model's inference endpoint URL (OpenAI-compatible, Ollama, vLLM, OpenRouter, or custom HTTP endpoint) and an authorized API key if required. Zero custom SDK installation or code rewrite is needed to start benchmarking immediately.

The Tanvelo Health Score (0–100) is a weighted composite based on 5 calibrated operational vectors: Reasoning Depth (30%), Knowledge & Grounding (25%), Deterministic Coding (20%), Safety Separation (15%), and Semantic Consistency (10%). Each vector isolates specific failure modes rather than relying on noisy general-purpose prompts.

When automated benchmark runs complete, anomalous or sub-par responses are transformed into dense vector embeddings. A clustering algorithm groups semantically similar failures into impact-ranked buckets, allowing you to instantly understand recurring patterns instead of reading hundreds of raw logs.

Remediation datasets are exported in standard JSONL (JSON Lines) and CSV formats. Each entry contains the failure prompt, the categorized failure mode, and a verified gold-standard target completion ready for supervised fine-tuning or prompt guardrail testing.

Never. Your inputs, completions, and evaluation reports are strictly private to your workspace and never utilized to train public, open-source, or foundational AI models.

Yes. Tanvelo supports regression comparison. You can benchmark release candidate v2.1 against baseline v2.0 to quantify score deltas, verify that known weaknesses have been remediated, and ensure no new regressions were introduced.

Enterprise tier workspaces can deploy private runner instances or configure VPC peering to test internal models that are not exposed to the public internet.