How it works.

Triage Engine is an early-warning indicator, not a forecast. It reads public sources for your company, groups what it finds into incidents, scores how unusual the last day looks, alerts the people you choose, and has a person at AgendaX approve any drafted response before it reaches you. This page sets out exactly how, including what it can't do yet, so you can judge it for yourself.

1. Read
2. Filter
3. Judge tone
4. Group
5. Score
6. Alert
7. Draft
8. Approve

1. What it reads

A scan runs for every monitored company every two hours. Each scan reads:

• 11 Nigerian news sites: TechCabal, Nairametrics, Punch, Vanguard, Premium Times, Channels TV, ThisDay, Business Day, Guardian Nigeria, Techpoint Africa, The Cable.
• Google News, searched for your company in your country's edition.
• Forums: Nairaland, Reddit and Hacker News.
• YouTube, at most once every 6 hours per company, to stay inside the free daily search allowance. The search adds up to two of your keywords, so they shape which videos come back.
• Regulator news, checked every four to six hours and shown for the regulators that cover your industry: Central Bank of Nigeria (CBN), Securities and Exchange Commission (SEC), National Insurance Commission (NAICOM), Nigerian Communications Commission (NCC), NAFDAC, Nigerian Postal Service (NIPOST), Nigerian Shippers' Council, Federal Competition and Consumer Protection Commission (FCCPC), Nigeria Data Protection Commission (NDPC). An item is kept only if it names a regulator and uses policy language such as circular, fine, licence or recall.

It does not yet read X, TikTok, Instagram, Facebook or WhatsApp. Whether each source worked on the last scan is tracked, and a source that keeps failing shows on your board and alerts us.

2. What counts as a mention

• Your company's name has to appear as a whole word. Names of three letters or fewer also need one of your chosen keywords, because short names are often ordinary words or abbreviations.
• Keywords you choose, such as a product name or "refund", raise a mention's priority. They never hide a mention that names you.
• Obvious noise is dropped: Reddit bot threads, very short Reddit posts with no risk word in them, and entertainment videos such as trailers, music videos and Shorts.
• A link already on your board isn't added again. The same article found through Google News and through the outlet's own site can appear twice; grouping usually puts both in one incident.
• A person at AgendaX can mark a mention as irrelevant. It leaves your board, your score, alerts and summaries, and is counted, so we can measure how much noise gets through.

3. How tone is judged

• Google's Gemini reads each new headline from your company's point of view, told to read Nigerian English, Pidgin, slang and sarcasm the way a Nigerian reader would. It sees only the public headline, the source's name and your company's name.
• If Gemini isn't available, a standard sentiment model scores the headline instead, with an added list of words it would otherwise miss, from "wahala" and "419" to "outage" and "downtime".
• Tone is judged from the headline, not the full article.
• You or AgendaX can correct any label, and the score then uses the corrected one. How often labels are corrected is our running measure of tone accuracy.

4. Grouping mentions into incidents

Ten posts about one outage are one incident, not ten. A new mention joins an incident from the last seven days when the two share at least two meaningful words and a quarter of their wording overlaps, once your company's own name and filler words are set aside.

The Incidents figure on your board counts each story once, and a mention that belongs to a larger incident shows how big it is. The score still counts every negative mention for volume and share, and uses incidents for corroboration.

5. The risk score

The score runs from 0 to 100 and looks at the last 24 hours. It has four parts:

• Fixed lines and your own normal. Acceleration compares your company with its own recent past. Volume and share use the same lines for every company, so a brand that is always discussed will usually score higher on those two parts than one that rarely is.
• No negative mentions, score 0. If nothing negative was found in the last 24 hours, the score is 0.
• Warming up. Acceleration counts once there are 3 days with a completed scan in the week before the last 24 hours, and the comparison is marked ready at 5. The average is never taken as less than one negative mention a day, so a single first complaint can't look like a spike.
• When it happened. A mention is timed by when it was published. If a source gives no date, the time we found it is used, and the board says so.
• No stale readings. If no scan has completed in 24 hours, the board shows no score rather than an out-of-date one.
• Every score explains itself. "Why this score?" on your board lists the reasons behind the number.

1. Baseline0 to 24
2. Watch25 to 49
3. Elevated50 to 74
4. Critical75 to 100

6. Alerts

• The score is checked after every scan. When it reaches your company's alert line, 60 unless we agree another, your alert contacts get an email with the score and its reasons.
• After an alert there is a 6-hour quiet period, so a running story doesn't fill your inbox. A move into a higher tier breaks the quiet period.
• Every Monday morning your alert contacts get a summary of the week.
• AgendaX gets a copy of every alert, and is told separately when a scan fails or a source stops working.

7. Drafts and human review

• When an incident needs an answer, we draft a holding statement, along with who to involve and a playbook. The draft follows the kind of incident it looks like: fraud or security, regulatory, product safety, payments and refunds, service outage, customer complaints.
• Gemini writes the draft when it is available; otherwise a prepared template for that kind of incident is filled in, so a draft is always ready.
• A person at AgendaX reviews it, edits it if needed, and approves it before it reaches you. Our internal case notes and working drafts are never shown on your board.
• Triage Engine never publishes anything. You decide whether, where and when to use a statement.

8. What it can't do yet

• There is no accuracy figure yet. We will publish one only after backtesting the score against known past incidents. Our backtest tool replays past Google News coverage through the same scoring the live service uses. It can't replay forums or video, so its results show only part of what the live service would see.
• The weights are judgement, not training. The four parts and their sizes were set by hand from how crises tend to unfold. They will change as pilot data comes in, and this page will say when they do.
• Coverage has gaps. No social platforms yet; some news sites and Reddit block automated reading from time to time; Google News returns a limited number of results for busy weeks.
• Headlines can mislead. Tone is read from headlines, and sarcasm and mixed languages are hard. That is why you or AgendaX can correct any label, and the score then uses the corrected one.
• It is an indicator, not a probability. A score of 80 means "look now", not "80% chance of a crisis".

9. How we measure ourselves in a pilot

• Detection lag: how long after publication a mention reached your board, by source.
• Tone accuracy: how often a person changed the automatic label.
• Noise: the share of mentions removed as irrelevant.
• Alerts: how many were sent and how many failed to send.

We share these with our pilot clients. Questions about the method are welcome: hello@agendax.com.ng.