Every thought an agent saves in FAVA Trails begins as a draft, and the Trust Gate decides whether it deserves to become shared, permanent memory. The gate reviews each candidate by asking whether a future agent, with no context about the conversation that produced it, would find it useful, and it rejects secrets, vague claims, transient state, and instructions disguised as facts. At the time, review used a hosted model by default: promotion candidates traveled through OpenRouter to Google Gemini, which meant every candidate crossed an external inference boundary and cost a small fee before it could be promoted.
With v0.6.0, the review could happen on your own machine. The Trust Gate could talk to any OpenAI-compatible endpoint, such as Unsloth Studio or vLLM, through the same provider-neutral interface it used for hosted models. With a local reviewer, promotion candidates stay on the host you configure and no per-thought inference fee applies. Whether text leaves your machine still depends on the endpoint you configure: with a hosted reviewer, the candidate has already been sent by the time a rejection comes back. Which reviewer a given installation uses is decided by its own configuration and the release you installed, so check the setup instructions linked above.
What the July 2026 experiment compared
We ran an Unsloth-served, quantized Qwen 3.6 27B model over 49 historical promotion decisions that carried earlier Google Gemini verdicts, and compared its verdicts with Gemini’s. The comparison asks whether two models agreed. It does not establish that either was right about the facts in those records, and 49 cases from one project say little about other corpora or other models. In this run the local model’s disagreements all went in one direction: it rejected some candidates that Gemini had approved, and it never approved one that Gemini had rejected. Those are historical comparison outcomes between two reviewers, not an error rate for either. Under the current lifecycle, a rejected candidate is frozen with the status rejected and is excluded from approved recall. A revision is a new, linked draft rather than an edit of the rejected one. The raw numbers, the second-model audit of each disagreement and the limitations are in the recorded comparison on GitHub.
The full story on Substack
Why the direction of the errors matters more than the agreement rate, what the nine disagreements with Gemini revealed, and where a quantized model struggles.
Read the benchmark write-up →Where setup lives now
The configuration shown in this experiment was the per-machine setup of the v0.6 era, and it is no longer maintained here. The setup instructions own the current reviewer configuration, including credentials, timeouts and per-trail overrides. A release that includes the local-reviewer bootstrap lets fava-trails bootstrap write a local reviewer setup for new installations, and the release history says which release that is. Check what your installed version does with fava-trails version, and read the effective reviewer destination from fava-trails doctor before you rely on it. The local review walkthrough shows a local reviewer rejecting two drafts, approving a third and recalling it. For the wider picture, start at the FAVA Trails home page.