Talavine
Four configured models answered our question about introducing Talavine Code to developers. Then they ranked the answers with authorship hidden. Three returned usable rankings; DeepSeek's answer won this panel.
The interesting bit was the disagreement: should another promotion experiment wait until GitHub stars grew, or are early star counts too noisy to make that a useful gate? We chose to keep testing distinct useful angles and measure what happened.
Here's how to use the same workflow before committing to a design or a rewrite.
Talavine Code is an MIT-licensed coding harness for your repositories, with supported local models, API providers and coding CLIs. Its /compare command asks configured API model lanes the same question independently, then has them rank the answers with authorship hidden. /review is an alias.
Give the models a decision they can actually review
Try this in Talavine Code's interactive terminal:
/compare A small app reads a local JSON settings file at startup. Should it add a database now? Compare the simplest workable approaches, name the assumptions that change your answer, and propose a test for the main risk.
Without a model list, the panel uses the first configured API lanes. You need multiple available API model lanes for multiple independent answers. Supported local API models can take part; installed coding CLI lanes do not participate in this panel. In the desktop app, /compare opens the comparison workflow where you choose contenders and a judge. The judge steps in only if the contenders still disagree.
Don't make the models guess the whole project. Include the relevant code or plan, the constraint that matters, and what would count as success. A file path by itself is not the file's contents.
Read the argument, not just the winner
Look for three things in the report:
- A changed assumption. One answer may assume multiple writers; another may assume one local process. Those are different problems.
- A concrete disagreement. Ask what observation or test would settle it.
- A weak judge. Missing or invalid rankings are missing evidence. A fluent verdict can still be wrong.
For a separate judge in the terminal, use --judge with a configured API lane name. That mode first lets the contenders review one another and concede; the judge decides a remaining disagreement. There isn't a separate /judge command.
What happened in our own review
We used Code's review panel to challenge the plan for introducing Talavine Code to more developers. Four configured models answered; three returned usable rankings. DeepSeek's answer won the panel. One reviewer wanted a second promotion experiment to wait for star growth; another challenged that rule because early star counts are a noisy signal. We chose to test distinct useful angles while measuring the results.
That is one real review, not a benchmark proving the winning model is generally better. The disagreement was more useful than the scoreboard.
Make the chosen plan earn its keep
Run the relevant tests, inspect the changes, and check the actual app. A panel ranking is advice, not proof that code works. Each contender receives the material you supply through its configured provider; comparisons make multiple model calls and provider charges apply.
Explore Talavine Code and its setup · Inspect the MIT source on GitHub
If this is a tool you'd like to keep around, a GitHub star helps other developers discover it. Trying it and reporting a concrete bug helps too.
Prepared by the Talavine team with AI assistance and internal review.