The pipeline, and every dial on it.

No person curates the board. Every row is generated by a three-step funnel, and the human decision starts at the board, not before it:

1. The watch-list decides what's monitored (the only "selection", set once in config): every morning Google News is searched for the last 7 days of "NiCE CXone" OR Genesys OR Five9 OR Talkdesk OR Verint OR Zendesk OR "Salesforce Service Cloud" OR "Amazon Connect" OR "Zoom Contact Center" — the list lives in entities.json → ingest_query_competitors (see § 04). Typically 25–40 articles return.
2. Code classifies & scores each article, deterministically: a category from a fixed dictionary (AI agents · Contact center · Enterprise AI · Voice AI · Customer experience · Analyst recognition · Workforce management · Competitive landscape · Market moves), which brands are mentioned, sentiment, and a base score 0–70. No AI yet — the same article always scores the same.
3. Claude clusters & names: related articles merge into one opportunity (three Sopra Steria deployment stories → one cluster), and Claude writes the topic line you see on the board, the why it matters, trigger type and format suggestions — plus a bounded −15…+30 score nudge.

Final score = base + nudge. Score ≥ 40 → it becomes a board row. Below 40 → stored for trend maths, never shown. Your team's job begins here: pick rows, request briefs, publish.

Each signal gets a deterministic base score 0–70 computed in code — topical relevance, brand/competitor presence, engagement, recency. The same signal always scores the same. Claude only adds a bounded −15…+30 nudge for what code can't judge (newsworthiness, search demand, format gaps). Day-over-day score movement is therefore real market movement, not AI mood. Thresholds: ≥40 lands on the board · ≥60 makes the digest · ≥80 is flagged Immediate.

An editorial gate keeps the board clean: stock-price, earnings and layoff stories are auto-marked Intel-Only — tracked for awareness, never turned into content briefs.

⚠ These are the configured schedules. They only fire while the host machine is running — see § 06.

Persistence moved from Turso — a community database node we had to maintain ourselves — to Supabase, managed cloud Postgres accessed through n8n's native integration over REST. All 352 historical rows migrated and verified, schema and row-level security applied to every table, and all five production workflows ported and live-tested the same day.

Why it matters: zero custom components remain, which is the property that makes the planned move to the company's online n8n instance a configuration exercise rather than an engineering project.

The Opportunity Board sheet, sorted by score. New rows land after 07:00 — on days the host is running. "Immediate" status means act this week.

Send the one-liner in § 03.1, or ask whoever operates the pipeline to. That's the entire ask — the system schedules its own follow-up and verdicts it three weeks later.

One JSON edit plus one command chain (§ 04, chain ①). Minutes, not days. The same applies to removing noisy aliases — and it's worth doing, since the feed currently pulls in several unrelated companies that happen to share a competitor's name.

SEMrush is already live. Search Console access was granted on 30 Jul and the daily feed is the next thing to build. Neither touches the scoring core — see § 05.

Yes, field for field — Supabase is the system of record; the Sheet and Doc are human-readable copies. The Supabase write is the protected one; Sheet and Doc writes fail soft with a logged warning. Caveat: the mirror is one-way. Hand-edits in the Sheet are not synced back, and the pipeline reads only Supabase. Use the Sheet to read and annotate, not to steer.

Claude usage on the daily scoring run and per brief, bounded at 25 signals per run with briefs on demand. Tavily search on briefs and reviews. SEMrush comes out of the existing subscription. Supabase is on the free tier — worth upgrading before the online move. The trend report is deliberately zero-AI.

No. Code computes the base score deterministically; the AI adds a bounded adjustment and drafts prose. Dedup ledgers, thresholds, verdicts and the editorial gate are all deterministic rules you can read and change.

The failure fires the error workflow: it's written durably to the ops_errors table and posted to Teams once the real channel is wired. Nothing fails silently — that's a design rule, including soft failures like a Doc-sharing problem.

Three things worth naming. Our own acquired brands are not in the search queries, so Cognigy news has never entered the pipeline. Conversation on Reddit and LinkedIn is not flowing yet. And why we actually lose deals — sales win/loss — has no agreed source. Each is a decision or an approval rather than an engineering problem.

Pipeline operations, config edits and new sources: the operator (Aleksandar). Content decisions on board items: the owning team named on each row. This page is regenerated from the vault — ask for a refresh anytime.