Snake Lab Prompts

Qwen starts with golden’s configuration and chooses which settings to change before running a simulation. Each prompt opens with the available decision. For value choices, choose-snake-setting tells Qwen how to weigh the evidence.

Prompt Purpose
ConfigurationNextStep Lists available keys and current values. Gives Qwen the option to run once the draft is unique.
ConfigurationChoice Names the submission tool and supplies current values, reference scores, and data paths. Learning rate, epsilon, and distance rewards can include a chart.
EpsilonNiceChoice Adds a compact definition of the custom EpsilonNice exploration rule when choosing its probability and steps.
InvalidValue Explains why Ax3l rejected a value.
NoDupes Identifies the duplicate proposal and asks Qwen to try different values for the same setting.

Use select_parameter(parameter) with a listed key. Each setting can be chosen once per draft, in any order. submit_single_value and submit_pair_values save changes using the limits in their tool descriptions. Ax3l binds the draft and setting identifiers. Duplicate proposals stay in the same conversation while untested values remain for that setting.

select_parameter("seed") immediately saves the next unused seed once per draft, without a separate seed confirmation conversation or tool call. After the run decision, golden runs first on the new seed and sets the baseline even if its score is lower. If other settings changed, Ax3l tests them on the new seed and requires a higher score to promote the draft.

The internal seven-minute decision timer spans setting selection and value choice. Prompts and tool feedback do not mention the timer or its expiry actions. Experiment Status continues to display the countdown. Saving a unique change resets the timer. A prompt with a chart adds two minutes to the remaining decision budget once; reads and rejected values keep that extended deadline. After the configured grace period (seven minutes by default), Ax3l restarts the conversation if there is no saved change or runs the saved unique draft.

data_tool(url) reads one small page. Use /draft for saved and inherited values, / for the run menu, and direct paths such as /latest/scores or /golden/activity for details. Navigation uses English. Values are JSON and missing statistics are null. Previously saved observations remain available through a run’s notes page. kb_tool(url) serves short local guides, starting at /.

A read continues the conversation. An accepted submission finishes the choice. Prompt events preserve the exact text, JSON evidence, and PNG that Qwen received.