How the reading is built

The reading is produced in a chain where machines do the routine work and a person checks every significant rise before it is published.

The chain, step by step

  1. Fast scan every 5 minutes

    Official press releases and feeds from reliable news media are read. A keyword filter picks out potentially relevant items; only those are shown to a light AI model for triage.

  2. A rule decides on an early run

    A fixed, written rule – not the AI – decides whether to start a full analysis immediately. The fast scan never changes the reading.

  3. Collection

    The full analysis gathers all feeds for the last 36 hours and runs 25 standard searches. Every fetched page is stored with a timestamp and checksum.

  4. Coding into a closed codebook

    A capable AI model reads the material and fills in a form: event type, modifiers and evidence with quotes. It does not assess the situation or compute the number.

  5. Validation

    Code checks that every code is allowed, the source is on the allow-list and the quote appears word for word in the fetched page. Anything else is rejected.

  6. Computation

    The model appends accepted events to the ledger and computes the reading, band and warning level. The ledger is append-only: corrections are new rows.

  7. Human review

    If the reading would move a lot, a new gate would open or the event is serious, a second model codes the same material independently and the operator confirms the result before publication.

  8. Publication and alerts

    The site is rebuilt as static pages and published. Alerts go out only from a confirmed reading.

The role of AI: a coder, not an analyst

AI models do one narrow job in FWSI: they read sources and place what they find into a predefined codebook. They never

When in doubt the model is instructed to choose the lower level. Rises are decided by a rule, a second model and a person. The coding instructions and the response schema are public.

Evidence rules

Nothing serious is published unchecked

When the computation detects a significant change – the reading would move five points or more, a new gate or tripwire would open, or the event is serious – publication stops. A second, stronger AI model codes the same material independently and the results are compared field by field. The operator sees both codings and the source pages with the quotes highlighted, and decides: confirm, correct downwards, or reject. Until then the status page says "new information under review".

The morning schedule

TimeStep
05:30Collection: feeds for the last 36 hours and 25 standard searches.
05:45Coding and evidence checks.
05:55Events into the ledger; human review if needed.
06:00The reading is computed.
06:057-day outlook and briefing in Finnish and English.
06:15The site is published.
06:20Briefings and notifications to subscribers.
06:30–08:00Morning message at the subscriber's chosen time.

When the warning level rises, full analyses run more often: at level 1 also at 18:00, at level 2 every six hours, at level 3 hourly and at level 4 every 15 minutes. The fast scan runs every five minutes at every level.

If something fails

If AI services or sources are unavailable, the model computes the reading from the existing ledger – decay is deterministic – and the site says there is no new source material. The fast scan carries on with keyword rules, and the operator can enter events by hand with the same checks. The site is static, so it stays readable even when the rest of the server is down.