How it works
Signals you can trace, words you can check
Coastal Intelligence turns open satellite, ocean-model and weather data, plus your own records, into five signals for your site. A language model then puts those signals into words for the assistant, the weekly brief and the post-mortem. This page explains each part, what it is good for and where it stops.
The five signals
Each signal says whether the coming days look favourable, marginal or unfavourable for an activity or a species, with the evidence behind it. Signals are computed by fixed rules and statistical models from the data; no language model is involved in computing them.
Operability
Whether each of the next seven days is favourable, marginal or unfavourable for each activity you name (longline work, trestle access, harbour movements), from forecast wave height, period and direction, wind and tide in daylight. Days 6 and 7 carry low confidence.
Thermal stress
Sea temperature at stock depth against the watch and stress thresholds of your species, bias-corrected with your own logger where you have one, else the nearest buoy.
Bloom risk
A regional 0 to 100 index from satellite chlorophyll-a, the sea-temperature anomaly, calm days and catchment rain. The bay itself is not verified by it: for closures follow the official bulletin.
Runoff shock
Salinity drops after heavy catchment rain, from ERA5-Land and the ECMWF forecast over the bay's own catchment and the IBI model's salinity, confirmed by Sentinel-2 where cloud allows.
Growth
A 0 to 1 growing-conditions index for the species you stock, from temperature, light and day length, plus nutrients for seaweed.
Scale matters: regional model values describe a cell of about 3 km, not your exact longline; satellite values are observed but have cloud gaps; only your own logger and records describe the site itself. Near the shore the models can miss your site altogether, and the app then samples the nearest sea point and says so.
What the assistant can and cannot do
The assistant answers questions about your site from the signals, the observations and your records.
It can
- explain why a day is favourable, marginal or unfavourable for an activity you named;
- compare the model with your logger and say how far to trust it;
- find what coincided with a past event (a heatwave, a salinity drop, your notes);
- point you to the official HAB bulletin and the closures in force.
It cannot
- say whether shellfish or seaweed can be eaten or sold: that is food safety, decided by the official monitoring and closures;
- give navigation or sea-going advice, or replace a weather warning;
- tell you whether you comply with a licence, a law or an insurance policy;
- use another operator's records, or anything it did not look up for this answer.
Every factual sentence must carry a citation to data the app looked up for that answer. An automated check removes sentences without one, scores the answer and replaces it with a short “I couldn't back that with your data” message when too little of it is sourced. Text inside your notes that tries to instruct the assistant is ignored.
How to read a citation
A citation is the id of a record in the knowledge graph, shown as a chip at the end of a sentence. Its first part tells you what kind of source it is:
obs-an observation or forecast value for one day (model, satellite, buoy or your logger);ind-a signal value (operability, thermal, bloom, runoff, growth);thr-a species threshold,act-one of your activity limits;evt-an event such as a bulletin entry, a closure or a marine heatwave;ev-your note, photo or record,anm-an anomaly in your records,skill-how the model compares with your logger.
Click a chip to open the record in the graph explorer, with where it came from and what it was computed from. In a PDF the chips become numbered footnotes with a Sources list. A sentence you cannot trace to a chip is a reason to report the answer.
You decide: human oversight
- The signals and the assistant are decision support. You decide whether to go out, harvest or deploy.
- Official closures, the Marine Institute HAB bulletin and weather warnings always prevail over anything here.
- Under every answer and brief you can give a thumbs up or down, or report this answer. Reports are read by a person, together with the automated grounding check, and harmful answers are handled as incidents.
AI notice
AI-generated from the cited data; check official bulletins before acting.
- The assistant's answers, the weekly briefs, the post-mortems and the e-mails that carry them are worded by a general-purpose large language model (OpenAI or Anthropic, whichever the service is configured with) from the data the app computed. They are labelled as AI-generated wherever they appear, including the PDF export.
- The five signals, the maps and the charts are computed without a language model.
- Before anything is sent to the model provider, names are replaced by pseudonyms and contact details, addresses and coordinates in free text are removed. Under the providers' API terms, what we send is not used to train their models.
- The service is not a safety component and not food-safety, navigation or legal advice. Our assessment under the EU AI Act, the model card and how we handle a harmful answer are documented with the source code.
- Questions about this notice: Triskelion AI Systems Limited.