Build an H5N1 AI monitor that gives you a plain-language public-source brief—and knows when to stop

In the Geelong region, a dead dolphin and a dead seal washed up on the same day, fuelling community fears about bird flu spreading to local wildlife. The supplied evidence does not include an H5N1 result or population denominator for the Geelong dolphin and seal.

A constrained H5N1 AI monitor is not a crystal ball. It reads named public updates, turns them into a plain-language brief, and stops before medical, travel, or wildlife-management advice. The goal is repeatable news monitoring.

Start with the brief template

The output is a four-field brief: confirmed event, species affected, population impact, and uncertainty. If a field is empty in the supplied evidence, say so. Output rule: if a source or date is absent from the supplied evidence, write 'not in the supplied evidence' and do not infer it. Filled example: ABC, 29 Aug 2026: >350 events, >20 species; date field: 29 Aug 2026; population impact: mass mortalities mentioned, but no population denominator in the supplied evidence.

The ABC count shows why the fields matter: more than 350 confirmed H5N1 bird flu events had been reported, including mass mortalities, and more than 20 distinct animal species had been reported as affected by H5N1 in Australia. Date field: 29 Aug 2026. The count is not a conservation assessment.

Checklist: source named? date present? confirmed language? species specified? population denominator given? If any are absent from the supplied evidence, mark those fields 'not in the supplied evidence'.

Separate event counts from impact. A corpse count is not a conservation assessment. Impact needs a population denominator and significance language: local population, regional population, and what the source says about significance. If the supplied evidence does not give one, mark population impact 'not quantified in the supplied evidence'.

Effective monitoring is needed to understand the true magnitude and significance of the H5N1 outbreak.

That sentence is the whole job. The pipeline makes the public record easier to follow.

What the pipeline does

Build it as a narrow pipeline, not a general assistant. Input: a short list of public sources. Output: the four-field brief. For H5N1, the unit of information is not a vague virus-detected headline. It is a confirmed event tied to a place, date, species, and source. The public record can show confirmed events, species names, dates, and mass mortalities, but a beach carcass does not become a species list unless a named source confirms the species.

Every claim needs a source and a date. A clean line looks like: Source X reported on DATE that species Y had confirmed H5N1 in location Z. If the supplied evidence does not give a date, write 'date not in the supplied evidence'. If the supplied evidence does not give a species, write 'species not in the supplied evidence'. Boring on purpose. Boring keeps the brief from becoming a rumor mill.

Source list, claim rules, and the brief

Use only named public sources. For a US-based watcher, that usually means federal and state public-health or animal-health agencies, local public-health departments, named wildlife agencies, and named veterinary or conservation organizations when they publish official updates. For international coverage, use the relevant national animal-health authority, public-health authority, and named conservation agencies. Forums, screenshots, social-media posts, and unnamed officials are not confirmed events.

The brief follows a simple rule set:

1. Watch only named public sources.
2. Report only confirmed events, species affected, population impact, and uncertainty.
3. Include source and date for every claim.
4. Label medical or travel questions as not advice.
5. Escalate only when a named official source changes guidance, adds a confirmed species, reports a mass mortality, or states significant population risk.

Confirmed means the source says it is confirmed, not that the pipeline feels confident. If a source says suspected, under investigation, or preliminary, keep that language. If a source says H5N1 was rapidly spreading along southern Australia's coast and had killed wild birds and infected seals, report that as a public-source statement with source and date. Do not add 'therefore your town is at risk' unless an official source says that.

When to stop, escalate, and say no

The stop condition is the important feature. Stop before medical advice, travel advice, or management advice. If a user asks, 'Should I avoid the coast?' the answer is not advice. If a user asks, 'Can I keep my chickens inside?' the answer is not advice. If a user asks, 'Is it safe to eat eggs?' the answer is not advice. The pipeline can point to the relevant public-health source and say the question is outside its scope.

Escalation is narrow. Flag a change when an official source changes guidance, adds a new confirmed species, reports a mass mortality, or says a population is at significant risk. Do not escalate because a headline is scary, a local beach has a dead animal, or a social-media thread is growing. In the Geelong example, the dead dolphin and dead seal were a community concern, but the job is to check whether a named source has confirmed H5N1 in those animals, whether other local animals are affected, and whether official guidance has changed.

Vaccination and conservation actions are public-source material. If a named source says little penguins were receiving vaccines, report that. It is not a signal to recommend vaccination for a reader's animals, pets, or backyard birds.

A good H5N1 AI monitor is a lab notebook. It makes public updates easier to compare, harder to misread, and hard to turn into casual advice. If the brief can answer what did the named source say, when, and about which species? it is doing the job. If it starts answering what should I do? it has gone too far. The brief stops at: dead dolphin and seal reported; H5N1 result not in the supplied evidence; no escalation under this rule.