Confidence is an editorial triage score. It estimates how safe the draft is to review as a pre-draft, based on source quality, relevance, extraction quality, and whether the copy came from live AI or deterministic fallback. It is not a factual accuracy guarantee and it is not approval.
That badge means the draft came from deterministic fallback rules rather than a live LLM call. This is expected when `smart_connection.live_ai_enabled` is false in Rails credentials, even if an AI API key and model are configured.
Fallback drafts are useful for seeing what was scanned and for early triage, but they are not editorial-quality AI rewrites. Reviewers should compare them with the original source URL before approving or rewriting them.
If live AI is enabled later, existing fallback drafts will not automatically regenerate. New source items, or an explicit regeneration workflow, are needed to produce live-AI drafts.
Primary or highly credible source, clear relevance, enough source material, and no obvious rejection flags. Still requires editorial approval.
Credible source and likely relevant, but the reviewer should check tags, tone, summary, and original source before approval.
Usually a partial web extraction, thin source text, deterministic fallback, or one-sided framing. Do not approve without comparing against the original source.
Source is blocked/manual/unverified, not clearly relevant, low quality, or flagged by AI/rules. Treat as unsafe for pre-draft use.
When live AI is enabled, the model returns a structured confidence score and review note, then the app validates the output. The app can cap the score when the source is not clearly about outdoor sports surfaces, facilities, standards, procurement, management, sustainability, safety, or end-of-life.
When live AI is disabled, deterministic fallback assigns a conservative score from source reliability and source depth:
Important: deterministic fallback means the draft was generated by local rules, not live AI. It should always be reviewed by a human.
Active sources are eligible for weekly scans and manual scans. Paused sources stay documented but are skipped by scheduled scans.
Approvals are intentionally one-by-one. Each approval becomes an editorial decision and a feedback signal for future ranking; bulk approve would dilute that signal — both for human editorial control today and for any learned ranking model later.
Bulk reject is still available behind a one-step confirmation, because rejection signals are coarser-grained and editors need a bulk path for cleaning out obviously off-topic batches.