How Delentia chooses and switches models
The model is a setting, not a fixed component. This page separates what works from what used to be described far beyond the evidence.
Built and tested
- You set the model: a local model (Ollama), any model on OpenRouter, or any OpenAI-compatible endpoint. The choice is a setting you can change with one command or on the Models page.
- Per task, the runtime chooses a fast path (a short answer, few steps) or a slow path (step by step). If the router itself fails, it takes the slow path.
- Backup models take over when the main one fails (a network error, a timeout, a 5xx or 429 answer). A backup is allowed only if it exposes data no more than the main model, is allowed by the data-residency policy, and has a known price that fits the budget. Every switch is logged.
- A per-episode token and cost limit refuses a call whose worst case would go over the budget you set.
Not there / not measured
- No fixed roster of models. There is no seven-model panel, and no model is a partner.
- No cost saving and no routing latency was measured, so none is claimed.
- No evidence that the fast/slow choice improves results with a capable model. Only scripted and small local models have been used.