About and method

The tension index is a public engine that scans open sources and answers, per topic, one simple question: how likely is a specific event in the next 7 and 30 days, and when is the most likely window. The first topic is Iran, with two questions: a strike on Iran, and an Iranian attack on Israel.

Three things set it apart from similar indexes: full transparency (every signal is shown with its source, reliability, evidence level, corroborations and contribution to the number), a public accuracy record (an append-only forecast log and a Brier score), and calm alerts.

How the number is computed

1. Collection. Every 30 minutes the engine fetches new items from a fixed list of sources: the news feeds of news sites and official bodies. The source list and its fetch status are at the bottom of the signals page.
2. Filtering and cleanup. Duplicates are detected and dropped. Only items containing relevant keywords are sent for classification.
3. Classification. A language model reads each item once and decides: is it relevant to one of the questions, in which direction (raising or lowering the chance), how strongly (1 to 3), and at what evidence level: rumor, single outlet, corroborated (several independent sources) or official (a statement by a government or military body). It also gives each item a "claim key" that identifies the underlying event.
4. Summing. Each question has a baseline: the historical chance of such an event in a 30-day window, derived from documented events. Each signal gets a contribution: direction times magnitude times an evidence factor times the source's reliability, with a small bonus when several independent sources report the same claim. Only the strongest signal per claim is counted, so one story repeated by twenty outlets cannot dominate the number.
5. Decay. Signals lose weight over time: the half-life is three days for the 7-day horizon and ten days for the 30-day horizon. A signal older than four half-lives is not counted at all.
6. Context indicators. Data such as the open-market rial rate or military flight activity enter with a very small weight, as context rather than news.
7. Probability. The sum is added to the baseline on a log-odds scale, clamped so it can never move too far, and translated into a probability between 2% and 97%.
8. The most likely window. Each positive signal spreads weight around the time it is expected to matter. The result is a distribution over five windows inside the 30 days: days 1 to 3, 4 to 7, 8 to 14, 15 to 21 and 22 to 30.
9. The explanation. A stronger language model writes, for each question, a few calm sentences about what changed since the previous update and why. If it fails, the previous explanation stays and the number is computed algorithmically.

What does not enter the number

• Manual judgment. The baseline is never changed by hand mid-stream; if it changes, the criteria version is bumped and the change is documented.
• Classified information. There is none. Only what was published openly.

The accuracy record

Every run that changes something writes a row to the forecast log, and at least one row every six hours even without a change. No row can be edited or deleted: the database blocks it. When a forecast's validity window ends, it is scored: if a qualifying event happened inside the window, the outcome is 1; otherwise 0. An event is recorded as "happened" only with confirmation from at least two major outlets or an official statement, and the evidence is linked. The Brier score is the mean squared difference between the probability and the outcome: lower is better, and 0.25 is the score of a constant 50% guess.

Alerts

There are two separate channels, each with its own level. Changes: an alert when the 30-day estimate changes by more than the threshold you chose and the change persists for two consecutive runs. News: an alert about a report according to its weight, from critical only to every relevant report. You can also set quiet hours, let critical news break them, choose a daily digest instead of separate alerts, and the language. Set up alerts.

FAQ

Is this intelligence?

No. It is an analytical estimate computed from open sources only. The system has no access to classified information, and it is not a substitute for Home Front Command instructions.

Why did the number change?

Each run adds new signals and decays old ones. The signals page lists every signal with its contribution, and the topic page has a sentence about what changed and why.

How much can I trust the number?

Not on its own. It is an estimated probability, not a certain forecast. The accuracy record shows publicly how good the forecasts have been so far.

What counts as the event happening?

The criteria are written in advance for each question and carry a version number. They appear on the topic page (under "What counts as happened") and on the accuracy page.

Who is behind this?

A private, independent project with no connection to any government or security body. The site collects no personal data beyond what the privacy policy describes.