configs

Configs

One YAML decides everything about a run. Every config here was regenerated from the config.yaml snapshot of the run that produced a number in the paper; its header names the paper item and the output folder, and its experiment_name (plus category for inference) is that folder's name in the released outputs, so re-running a config writes exactly where the release holds the paper's run. A config lists only what differs from the defaults in daedalus/core/config.py, whose dataclasses document every field.

configs/<benchmark>/generation/     DAEDALUS banks          (daedalus.scripts.generation)   → outputs/daedalus/
configs/<benchmark>/accumulation/   DAEDALUS-curated banks  (daedalus.scripts.accumulation) → outputs/daedalus-curated/
configs/<benchmark>/inference/      evaluation              (daedalus.scripts.inference)    → outputs/inference/
configs/appworld/inference/models/  other base agents (Table 2, Figures 1 and 8)
configs/smoke/                      one-task smoke tests   (tests/smoke_*.py)

The baselines keep their configs next to their code, in references/<method>/configs/; they write to outputs/baselines/<method>/memory/<benchmark>/ and outputs/baselines/<method>/inference/<benchmark>/. experiments-paths.md maps every table and figure to its configs and outputs.

Blocks

block	read by	main fields
top level	all	benchmark (appworld, tau2, automationbench), experiment_name, category (inference only: the paper section, i.e. the output sub-folder: main-results, cross-family, pipeline-ablation, …)
agent	all	the solver / base agent: model, max_turns, temperature, reasoning_effort
appworld, tau2, automationbench	all	task selection and environment (split, domain, step cap, τ² user simulator)
memory	inference	enabled, pool_path, heuristics_at_start (the method: whole bank at task start), or per-turn retriever + injection.policy + retrieval_mode
accumulation	accumulation, generation	the Solver loop: max_failures (Nf), num_success_to_continue (Ns), extraction_model
generation	generation	the Explorer, Judge, Surveyor (coverage_tags), guideline bank (explorer_guidelines), max_refinements (Nr), num_sessions, stage (Table 3 ablations B and C)
run	all	num_runs, parallel, max_tasks
provider_routing	all	OpenRouter upstream pinning (the Qwen and DeepSeek runs)

Two safety rules:

• Unknown keys inside a block are refused, and the error names the near miss, so a typo cannot silently fall back to a default. Unknown top-level keys are allowed: that is how a baseline carries its own block (expel:, erl:, …).
• A run folder is fingerprinted against its config. Relaunching a changed config under an existing experiment_name is refused instead of silently reusing finished work; use a new name or --force. run, logging and generation.num_sessions are not fingerprinted, so raising num_runs or resuming a generation run with more sessions is always allowed.

Paths are relative to the repository root, so pools resolve once the released outputs (illuin/daedalus-traces) are downloaded as outputs/ there. Put local, unshared configs in configs/private/ (git-ignored).