Configuration reference

This page lists every key you can set in a study's config.yaml. Every key has a default, so you can omit the file entirely. For how to write a study, see Author a study.

adda reads exactly the keys on this page. An unknown top-level key stops the run at start with an error that suggests the closest known key, so a typo cannot pass silently. Put your own settings under study:, which adda never reads.

Top-level keys

key	meaning	default
model	The language model to use.	The backend's default.
backend	claude, ollama, openrouter, or vllm.	claude
budget	Soft wall-clock limit, as "HH:MM:SS" or seconds. It is a nudge, not a hard stop.	none
eval_budget	Soft cap on how many real evaluations the run may spend.	none
budget_usd	Hard cost ceiling. The run halts when spend reaches it, and you can raise it and resume. Inactive on a backend with no per-call cost data, such as Ollama.	none
required_deliverables	Extra files that must exist before the run can finish.	none
evaluator	How a design gets scored. See How designs get evaluated.	Honor system.
objective	Which output is optimized, in which direction, and which 0/1 column marks a design feasible. See Declaring the objective.	Undeclared.
funnel	The 0/1 outputs that the viewer's stage funnel counts, in order. See Declaring the objective.	none
training_data	A precomputed pool used only as training data, with no live oracle. Use it for a surrogate-only study.	none
base_url	The model server's endpoint, on a backend that has one. See Use a different model or backend.	The backend's default.
nodes	Per-agent tools, model, backend, and base_url. See Customize agents and tools.	Each agent's own.
llm_slurm	Serve the model on a Slurm GPU allocation. See Use a different model or backend.	Off.
mem_cap	Hard memory cap per delegation, in bytes. When absent, the cap is the Slurm job's memory allocation, else a built-in default.	See meaning.
review_statement	Review PROBLEM_STATEMENT.md before the run. The review is advisory only. false skips it.	true
runtime	Run knobs: debug capture, timeouts, retry, and limits. See the Runtime reference.	All defaulted.
study	Settings for your own scripts. adda never reads inside it.	none

For the backend and model details, including how to set them per agent instead of for the whole run, see Use a different model or backend.

The runtime: block has its own page: the Runtime reference.

Required deliverables

pipeline.ipynb is always required. List any other file the study must hand back under required_deliverables, by bare filename:

required_deliverables:
  - design.json

adda refuses Done() until each listed file exists in the study directory. The agent writes it with WriteDeliverable, which takes a bare filename and writes into the study directory. A path with a directory, such as temp/submission.json, cannot be declared. If a study's PROBLEM_STATEMENT.md asks for a file by name, list that file here. A file that you do not list is not checked, and the run can finish without it.