Configuration reference

This page is the comprehensive reference for the system’s configuration surface. For the operational guide on which knobs to use when, see cost and performance and review modes.

Configuration priority order (highest to lowest): CLI args → environment variables → defaults.

Environment variables

Variable	Type	Default	Used by	Purpose
ELIFE_CORPUS_DIR	path	none	write, verify-refs, run	Where claim files are read/written
ELIFE_EXTRACT_OUTPUT	path	./out	extract	Where draft tables and intermediates land
ELIFE_EXTRACT_MODEL_RESULTS	string	claude-sonnet-4-6	extract, run, evaluate	Override Results-reader agent’s model
ELIFE_EXTRACT_MODEL_CAPTION	string	claude-sonnet-4-6	as above	Override Caption-reader agent’s model
ELIFE_EXTRACT_MODEL_STRUCTURE	string	claude-sonnet-4-6	as above	Override Structure-reader agent’s model
ELIFE_EXTRACT_MODEL_RECONCILE	string	claude-opus-4-6	as above	Override reconciliation + reviewer model
VERTEX_PROJECT_ID	string	cr-mainen	all subcommands using LLMs	GCP project for Vertex AI Claude
VERTEX_REGION	string	europe-west1	all subcommands using LLMs	Vertex region
GOOGLE_APPLICATION_CREDENTIALS	path	none	all subcommands using Vertex	Service account JSON for Vertex auth
ANTHROPIC_API_KEY	string	none	all subcommands using LLMs (when Vertex isn’t configured)	Direct Anthropic API auth

CLI flags by subcommand

extract flags

Flag	Type	Default	Purpose
--doi	string	(required)	eLife paper DOI
--corpus-dir	path	env or required	Corpus root
--output-dir	path	env or ./out	Where draft JSON lands
--paper-slug	string	derived	Override slug
--prompt-variant	string	default	Named directory under prompts/<variant>/
--reconcile-strategy	enum	confidence-tagged	confidence-tagged / union / intersection-only / majority-vote
--max-claims	int	unlimited	Hard cap on per-paper claim count
--no-retry-on-thin	flag	retry enabled	Disable retry when agent output looks thin
--model-results	string	env or claude-sonnet-4-6	Override Results-reader model
--model-caption	string	env or default	Override Caption-reader model
--model-structure	string	env or default	Override Structure-reader model
--model-reconcile	string	env or claude-opus-4-6	Override reconciliation model
--vertex-project	string	env or cr-mainen	GCP project
--vertex-region	string	env or europe-west1	Vertex region

write flags

Flag	Type	Default	Purpose
--draft	path	(required)	Path to draft JSON from extract
--corpus-dir	path	env or required	Where claim files land
--review-mode	enum	interactive	interactive / external / auto-approve / dry-run

verify-refs flags

Flag	Type	Default	Purpose
--paper	string	none (sweeps all)	Single paper slug
--corpus-dir	path	env or required	Corpus root
--dry-run	flag	write enabled	Show resolutions without writing back

run flags

(Same as extract, plus implies --review-mode auto-approve for the write phase.)

evaluate flags

Flag	Type	Default	Purpose
--reference-dir	path	(required)	Curated reference corpus root
--work-dir	path	(required)	Where per-paper artifacts and scorecards land
--paper	string	mutually exclusive with —papers / —all	Single paper slug
--papers	string	comma-separated list of paper slugs
--all	flag	mutually exclusive with above	Evaluate every paper-dir under reference-dir
--review-mode	enum	external	external / auto-approve
--skip-existing	flag	re-runs	Skip papers with existing scorecard.json
--prompts-dir	path	package-local	Override prompts directory
--prompt-variant	string	default	Named prompt variant
(model flags)	as in extract	—	Per-agent model overrides

The Config dataclass

All configuration resolution lives in elife_extract/config.py. The Config dataclass holds the resolved values:

@dataclass
class Config:
    model_results: str = DEFAULT_MODEL_RESULTS
    model_caption: str = DEFAULT_MODEL_CAPTION
    model_structure: str = DEFAULT_MODEL_STRUCTURE
    model_reconcile: str = DEFAULT_MODEL_RECONCILE

    vertex_project: str = DEFAULT_VERTEX_PROJECT
    vertex_region: str = DEFAULT_VERTEX_REGION

    corpus_dir: Path | None = None
    prompts_dir: Path | None = None
    output_dir: Path | None = None

    prompt_variant: str = DEFAULT_PROMPT_VARIANT
    reconcile_strategy: str = "confidence-tagged"
    review_mode: str = "interactive"

    max_claims: int | None = None
    retry_on_thin: bool = True

    @classmethod
    def from_args(cls, args) -> "Config":
        """Build a Config from argparse Namespace, falling back to env then defaults."""
        ...

Each subcommand handler calls Config.from_args(args) to resolve the configuration; the resulting Config is passed to the per-step modules.

Default values

DEFAULT_MODEL_RESULTS = "claude-sonnet-4-6"
DEFAULT_MODEL_CAPTION = "claude-sonnet-4-6"
DEFAULT_MODEL_STRUCTURE = "claude-sonnet-4-6"
DEFAULT_MODEL_RECONCILE = "claude-opus-4-6"

DEFAULT_VERTEX_PROJECT = "cr-mainen"
DEFAULT_VERTEX_REGION = "europe-west1"

DEFAULT_PROMPT_VARIANT = "default"

These defaults reflect the empirically-validated configuration the 10-paper sweep used. Change at your own risk; re-run evaluate to confirm any default substitution doesn’t regress.

Implementation references

• elife_extract/config.py — the Config dataclass and resolution logic
• elife_extract/cli.py — argparse definitions and per-subcommand argument lists

Next steps

• Cost and performance — operational guide on which configuration choices matter
• Review modes — when to use each — the most consequential configuration choice
• Adding a prompt variant — how the --prompt-variant flag works