Threat-language metrics

Why this matters

Use the gauge as a relative radar for this corpus: which days sound hotter than S2’s own long-run average. It does not measure ground truth risk, sarcasm, or negation — only lexicon overlap in published text.

Threat-language density per 1,000 words: NRCLex fear hits plus a modest boost from an S2-specific threat lexicon (attack, strike, kill, …). Higher means denser conflict wording, not a clinical diagnosis.

Long-run normal is the mean of all Wire + Intel Update works in the corpus. z = (score − mean) / σ. Near normal is |z| < 0.5; elevated z ≥ 1; crisis range z ≥ 1.5.

Deep method notes: How we calculate the score · Home gauge (last 30 days)

1. Monthly average threat density

Each column is the mean score for that month across Wire and Intel works. The dashed line is the long-run mean; the solid upper guide is the crisis threshold. Spikes mark periods when S2’s wording ran hotter than its own baseline.

2. Score distribution (full corpus)

How often each score band appears. Wire (darker) vs Intel Update (lighter) stack in each bin. Most mass should sit near the mean; a long right tail is “hot” briefings.

3. Wire vs Intel Update

Daily Wires are short and often event-dense. Weekend Intel Updates are long-form analysis; their average threat density runs lower even when the topic is serious, because the word count dilutes peak scare language.

Average emotion density by product

Four NRCLex-related densities (per 1k words). Fear is the gauge series (with threat-lex boost already in stored fear). Trust / sadness / anger are pure NRC for context.

NRCLex trust density (per 1k words). Rises with language of institutions, allies, reliability, and calm framing — useful contrast when fear is high. NRCLex sadness density. Often tracks casualties, loss, and bleak framing rather than kinetic threat words. NRCLex anger density. Overlaps threat language but also political blame and hostility framing.

4. How often is a day “elevated”?

Share of all works by z-score band against the full-corpus mean and σ. This answers “is hot the exception or the rule?” for this corpus.

• Much calmer: 187 (17.8%)
• Calmer: 181 (17.2%)
• Near normal: 362 (34.4%)
• Slightly elevated: 149 (14.2%)
• Elevated: 95 (9.0%)
• Crisis range: 79 (7.5%)

5. Recent works (last ~120 charted days)

Each bar is one published work (Wire or Intel). Hover for title; click through from the table below.

6. Normal score by time of year

A month-of-year normal is the average score for that calendar month across all years in the corpus (e.g. every October pooled). Compare a new October Wire to October’s normal, not only to the full-year mean — fall runs hotter on average. Full-corpus mean for comparison: 40.18.

Month-of-year normal (all years pooled)

Each bar is the average threat-language score for that calendar month across the whole corpus. Dashed line = full-corpus mean. Use this as the “normal for this time of year.”

By season

By weekday

Publish weekday of each work. Weekend dips often reflect longer Intel Updates, not quieter geopolitics.

Annual / regular patterns

• October is the hottest month-of-year on average (normal ≈ 43.21, +3.0 vs full-corpus mean). November is the coolest (normal ≈ 37.57, -2.6).
• Some months repeat year to year with little drift (YoY σ of monthly means ≤ 2.5): January.
• Other months are less regular across years (higher YoY scatter): February, March, June, July, August, September, October, November, December — treat those month-normals as softer guides.
• By meteorological season, Spring averages highest (40.86) and Summer lowest (38.42).
• Weekdays average about 42.5 while Sat/Sun average about 31.0. Weekends are often Intel Updates (longer analysis), which dilutes peak threat density even when the topic is serious — a product mix effect as much as a 'mood' effect.
• Mondays run hot (normal ≈ 45.46), consistent with catch-up Wire coverage after the weekend.
• The month-of-year cycle amplitude is about 5.6 points peak-to-trough (14% of the full-corpus mean) — a modest but detectable annual rhythm.

Pooled month-of-year / season / weekday means from all Wire+Intel works with fear scores. Month normal = mean of all years' works in that calendar month. YoY σ = standard deviation of per-year monthly means (needs ≥2 years with ≥2 works).

7. Extremes in the full corpus

8. What the score is made of

The score is a weighted sum of five per-1,000-word rates. Four come from the NRC emotion lexicon; one (threat_lex) is S2’s own military-vocabulary density, which catches wording the general lexicon under-weights. Everything else NRCLex can measure is listed below but carries a weight of zero.

Contract freakout_v3. Before v3 the score read fear and the threat term only, with no positive term anywhere — so a uniformly calm briefing and a neutral-but-vivid one scored alike. The positive term is the anchor that fixes that.

Negative is deliberately excluded even though it is the largest negative category. Measured across a 219-document sample it correlates r = +0.73 with fear, so folding it in would mostly re-add fear. Anger (+0.63) and disgust (+0.36) overlap less and earn their weight; positive is effectively orthogonal to everything threat-shaped (+0.13 against fear), which is exactly why it is useful as an anchor.