Likelihood words
Only a named editor can attach a likelihood to a path or a forecast, and only as one of these seven fixed words — never a percentage with decimals. The printed range is vocabulary, not a computed number.
| Word | Printed range | In plain words |
|---|---|---|
| Remote | 1-5% | less than 1 in 20 |
| Highly unlikely | 5-20% | between 1 in 20 and 1 in 5 |
| Unlikely | 20-45% | between 1 in 5 and just under even |
| Roughly even | 45-55% | about as likely as not |
| Likely | 55-80% | better than even, under 4 in 5 |
| Highly likely | 80-95% | between 4 in 5 and 19 in 20 |
| Almost certain | 95-99% | more than 19 in 20 |
Signpost states
- Quiet
- The rule's condition is not met; nothing unusual in the reading.
- Watching
- Above the watch threshold but not yet lit — worth following.
- Lit
- The rule's lit condition held long enough to clear single-day noise (hysteresis, §6.4).
- No data
- The underlying series had no evaluable observation — never shown as calm.
- Stale
- Past the series' own max_age_days without a fresh reading.
- No baseline
- Fewer than 14 usable historical days to compare against — too young to score.
The baseline
For every evaluation day, the engine compares today’s reading to the median of the same series over the trailing 8-35 days (excluding gaps, outages, and low-coverage days), scaled by a robust spread (MAD), a Poisson floor for sparse counts, and a registry-set minimum — never a Gaussian assumption. Measured tail behaviour on Zaviye’s own series is fat: 3-sigma days occur on 1.6-2.8% of days, not the 0.3% a normal distribution would predict. That score is never converted into a probability or a “confidence.”
Known regime breaks and outages
- 2026-08-13 – 2026-09-09 wp2repair.% (regime, filed)
wp2 repair-pass live test break
Signal freshness by adapter
| Adapter | Series | Last observation | Days behind |
|---|---|---|---|
| market_daily | 9 | 2026-09-16 | 1 |
| news_keywords | 41 | 2026-09-16 | 1 |
| derived | 1 | 2026-09-16 | 1 |
Known limitations
- This measures press coverage and market prices, not events — a real development under-covered by outlets in Zaviye’s archive will read as quiet.
- Coverage is uneven across the ~26 tracked languages; a concept with only English/Arabic terms will under-count Persian, Turkish, or Hebrew coverage of the same story.
- A topic-taxonomy change on 2026-06-05 means cluster-topic series (snapshot_topics) are not comparable before that date.
- Pegged currencies (the Saudi riyal, UAE dirham, and others) only move on real peg stress — a calm chart is the peg holding, not calm conditions generally.
- A three-day outage (2026-08-22 to 24) and several partial-coverage days are marked and excluded from baselines rather than silently averaged in.