Configuration reference

Every run of Crible is driven by one resolved Config (src/crible/config.py). Defaults are safe: parallelisation is OFF, a token ceiling is always set, and the corroboration threshold is >= 2. This page is the reference for those knobs; for a task-oriented walkthrough see the README.

Resolution order (later wins): built-in defaults → CRIBLE_* environment variables → explicit CLI flags / keyword overrides.

Auth & provider

Setting	Default	Env var	Notes
`auth_mode`	`api_key`	`CRIBLE_AUTH_MODE`	`api_key` (pay-per-token) or `subscription` (Claude OAuth via `ant auth login`). In subscription mode any `ANTHROPIC_API_KEY` in the environment is stripped before the SDK client is built, so a stray paid key can never silently bill a subscription run.
`provider`	`anthropic`	`CRIBLE_PROVIDER`	Provider abstraction for the sovereign/cloud split.
`model`	`claude-opus-4-8`	`CRIBLE_MODEL`	Model id; also selects the `web_search` tool version.
`base_url`	`None`	`CRIBLE_BASE_URL`	Override endpoint for sovereign deployments.
`api_key_env`	`ANTHROPIC_API_KEY`	—	Name of the env var holding the key. The key itself is never hardcoded or logged.
`effort`	`medium`	`CRIBLE_EFFORT`	`low \\| medium \\| high \\| xhigh \\| max`. Multiplies thinking + tool tokens on every call (~12 calls/run).

The web_search tool version is derived from the model: modern Opus/Sonnet/Fable models use the dynamic-filtering variant (web_search_20260209, which honours allowed_domains/blocked_domains); older models fall back to the basic variant (web_search_20250305) and run without adaptive thinking or the effort parameter.

Cost & safety ceilings

Setting	Default	Env var	Notes
`token_ceiling`	`200_000`	`CRIBLE_TOKEN_CEILING`	Cumulative tokens; halts the run when reached (a between-call gate, not a per-call cap). Must be positive.
`require_specific`	`True`	—	Refuse to run an over-vague question (the specificity gate); `--force` overrides.

Orchestration

Setting	Default	Env var	Notes
`parallelism_enabled`	`False`	`CRIBLE_PARALLEL`	Explicit opt-in only — parallel subagents cost roughly 15× tokens.
`max_subagents`	`8`	`CRIBLE_MAX_SUBAGENTS`	Cap on bounded subagent threads.
`max_iterations_per_thread`	`3`	—	Bounded tool-use loop per thread.
`max_search_uses_per_thread`	`2`	—	`web_search` `max_uses` cap per pass.

Retrieval steering & evidence mix

Setting	Default	Env var	Notes
`domain_steering_enabled`	`True`	`CRIBLE_DOMAIN_STEERING`	Dual-pass allow/block steering toward trusted sources.
`evidence_mix_floor`	`2`	`CRIBLE_EVIDENCE_MIX_FLOOR`	Minimum distinct high+medium sources before a clean verdict; blogs never count as corroboration. Must be `>= 1`.
`evidence_research_extra_passes`	`1`	`CRIBLE_EVIDENCE_EXTRA_PASSES`	Bounded re-search on a floor breach, hard-capped at 1 (never a loop).
`noncrawlable_search_domains`	`["reddit.com"]`	—	Domains Anthropic's `web_search` user agent cannot crawl; kept out of `allowed_domains` (listing them 400s the request).
`blocked_search_domains`	see `config.py`	—	Known blog/affiliate/SEO domains blocked at the source in the open pass.
`query_templates`	see `config.py`	—	Deterministic, logged augmentation templates that push searches toward lived experience.

Client-side discovery

Setting	Default	Env var	Notes
`discovery_enabled`	`True`	`CRIBLE_DISCOVERY`	Surface URLs the provider's `web_search` misses (e.g. reddit).
`discovery_backend`	`duckduckgo`	`CRIBLE_DISCOVERY_BACKEND`	General search backend that surfaces reddit + fora (no API key).
`max_discovery_results`	`6`	—	Cap on discovered URLs fed into the fetch + quote-verify pipeline.

Link liveness & content grounding

Setting	Default	Env var	Notes
`verify_links`	`True`	`CRIBLE_VERIFY_LINKS`	A cited source must resolve or it is dropped; 403/429 count as live, only 404/410/unreachable are dead.
`link_check_timeout`	`6.0`	`CRIBLE_LINK_TIMEOUT`	Seconds.
`fetch_enabled`	`True`	`CRIBLE_FETCH`	Fetch the cited page from your host and verify the quote is actually on it; subsumes link-liveness when enabled.
`max_fetch_pages_per_finding`	`4`	—	Pages per candidate — fewer, but read fully.
`max_fetch_chars`	`30_000`	—	Cap on stored page text used for verification.
`fetch_prompt_chars`	`12_000`	—	Per-page text fed to the model.
`quote_match_ratio`	`0.8`	`CRIBLE_QUOTE_MATCH_RATIO`	Fraction of a quote that must match the fetched page text.

Skepticism & ranking

Setting	Default	Env var	Notes
`corroboration_threshold`	`2`	`CRIBLE_CORROBORATION_THRESHOLD`	Minimum independent corroborations. Must be `>= 1`.

Paths

Setting	Default	Env var	Notes
`source_tiers_path`	`config/source_tiers.yaml`	`CRIBLE_SOURCE_TIERS`	The single, version-controlled source-trust tier list.
`runs_dir`	`runs`	`CRIBLE_RUNS_DIR`	Each run writes a directory with `audit.jsonl`, `plan.json`, and `advice.md`.

Non-secret settings are recorded per run in the audit trail (Config.effective_settings()) for reproducibility; secret values (the API key and ANTHROPIC_AUTH_TOKEN) are redacted from any audit output.