.env.example

# Postgres connection string used by the Payload CMS (cms/.env)
DATABASE_URL=postgresql://datum:datum@127.0.0.1:5432/datum

# Secret used by Payload to sign tokens; generate with: openssl rand -hex 32
PAYLOAD_SECRET=

# Public origin used to build absolute canonical article URLs.
# Required in production; without it, canonical metadata is omitted (no localhost fallback).
SITE_URL=http://localhost:3000

# Local-dev admin password for `npm run seed` (default: datum-dev-password)
SEED_ADMIN_PASSWORD=

# LLM API keys. Each live model id picks its API provider by prefix: claude-* ->
# Anthropic, gpt-* / o-numbered / chatgpt-* -> OpenAI. With MOCK_MODE=false every
# configured model must use an API-backed provider with its key set.
ANTHROPIC_API_KEY=
OPENAI_API_KEY=

# Legacy codex/* model ids are available only for deterministic mock fixtures.

# Model choice normally lives in the admin: Governance -> Models (/admin/globals/llm-settings).
# These env values are fallbacks used only when the admin field is blank
# (default: claude-opus-5). Ids come from cms/src/lib/llmCatalog.ts; an unlisted id
# still runs but logs $0 cost with a warning.
PIPELINE_MODEL_GENERATE=
PIPELINE_MODEL_FACT_CHECK=
PIPELINE_MODEL_QUALITATIVE_REVIEW=
# Information-gain QA gate model overrides (same fallback rules as above)
PIPELINE_MODEL_CLAIM_EXTRACTION=
PIPELINE_MODEL_INFORMATION_GAIN_JUDGE=
PIPELINE_MODEL_EVIDENCE_VERIFICATION=
# Checks first-party claims against the evidence bank during QA
PIPELINE_MODEL_EVIDENCE_CHECK=
BRAND_VOICE_EXTRACT_MODEL=
# Drafts onboarding steps (audiences, positioning, evidence bank); falls back to
# BRAND_VOICE_EXTRACT_MODEL before the platform default
SETUP_ASSIST_MODEL=

# Ahrefs API key for keyword/SERP research
AHREFS_API_KEY=

# Country code for SERP research and corpus-snapshot scoping (default: us)
AHREFS_COUNTRY=

# The domain articles are published for. Fallback only: the Workspace global
# (/admin/globals/workspace-profile) wins when it is filled in.
TARGET_DOMAIN=example.com

# Competitor domains, comma-separated. Fallback for the Workspace global too.
COMPETITOR_DOMAINS=competitor-a.com,competitor-b.com

# When true, pipeline stages return canned responses instead of calling paid APIs
MOCK_MODE=true

# --- Information-gain policy -------------------------------------------------
# The intended surface for these is the admin: Governance -> Information-gain
# policy (/admin/globals/information-gain-policy). Resolution is admin field ->
# env var -> platform default, so anything set here is a fallback used only
# while the matching admin field is blank; it is for CI and for pinning a value
# where there is no admin to click. A value that is missing, unparseable, or out
# of range for its kind falls through to the default rather than throwing.
#
# Ratios take 0-1, counts take a non-negative integer, booleans take
# true/false/1/0/yes/no. Defaults are shown; every threshold and what breaching
# it produces is documented in docs/information-gain.md.
#
# Breach -> REVISE
#INFORMATION_GAIN_MIN_CONSENSUS_COVERAGE=0.75
#INFORMATION_GAIN_MIN_VERIFIED_NOVEL_CLAIMS=1
# Breach -> BLOCK
#INFORMATION_GAIN_MIN_VERIFICATION_RATIO=0.9
#INFORMATION_GAIN_MIN_NOVEL_FACTUAL_INTEGRITY=0.9
#INFORMATION_GAIN_MIN_NUMERIC_TEMPORAL_INTEGRITY=0.95
#INFORMATION_GAIN_REQUIRE_EXACT_VALUE_MATCH=true
#INFORMATION_GAIN_REQUIRE_EVIDENCE_LINEAGE=true
#INFORMATION_GAIN_BLOCK_FIRST_PARTY_MEASUREMENTS=true
# Breach -> HUMAN_REVIEW
#INFORMATION_GAIN_MAX_CONTRADICTION_PROBABILITY=0.25
#INFORMATION_GAIN_MATERIAL_NOVELTY_THRESHOLD=0.55
#INFORMATION_GAIN_MAX_INTERNAL_DUPLICATION_RATE=0.35
#
# Note: these thresholds are uncalibrated policy dials, not measurements, and
# they are only half the picture. The other half is the evidence-sources table
# (/admin/collections/evidence-sources): an unclassified domain's source quality
# is capped at 0.75, so a materially novel factual (0.90) or numeric (0.95)
# claim cited only to unclassified domains is blocked no matter what these say.