Changelog

Engine v4.5 — one story counts once

• Why. Outside reviewers made a fair point: a burst of coverage could lift the score with nothing new behind it. Two causes were found in the code and both are fixed.
• Copies no longer count as events. When 41 feeds carry the same story it used to count 41 times. Near-identical headlines are now grouped into one story; breadth still counts for a little (capped), because many outlets carrying a story is evidence, not a multiplier.
• A shock now needs corroboration. An event phrase in a story carried by fewer than three independent outlets no longer fires the overlay. One dramatic headline cannot move the score.
• Things said weigh less than things done. Coverage is split into events (struck, seized, closed, killed) and rhetoric (warns, threatens, may, could). Rhetoric carries one third of the weight, and both readings are published.
• Ongoing stories stop stacking (v4.5.1). A topic in the news for months created one shock event per day — 85 live events for a single strait at the worst. A topic already alive is now refreshed, not re-added, and no single topic may own more than 8 of the 18 overlay points.
• Effect measured before release. On a simulated burst (one story, 30 outlets) the tension share fell from 0.457 to 0.097 and shocks fired from 31 to 1; a genuine five-outlet strike still fires once at full weight; on a normal day the shares moved 1–6%, so the level and the record stay comparable. No historical value was changed.
• Now archiving. A daily corpus digest is kept from today, so future method changes can be replayed against real data — which was impossible before, as raw headlines were never stored.

Engine v3.5 — self-running, self-drawing, self-documenting

• Baseline research concluded in public: 12-month window proved war-contaminated, 5-year still war-heavy (active war ranked ~56th percentile), production moved to a 10-year window — calm 2016–2019 anchors the scale.
• Engine self-schedules full analysis at 00:10 & 12:10 UTC — timezone-proof, with automatic catch-up if a slot is ever missed.
• Daily share card auto-rendered as PNG on every full run — with a built-in zero-dependency renderer, so it works on any host, forever.
• Public API documented at /api/; dataset schema markup added (Google Dataset Search); Methodology Whitepaper v4 published.
• Legacy engines retired: the data layer now runs on entirely free, keyless sources at $0.

Engine v3 — the honest instrument

• Scoring rebuilt on the academic GPR coverage-share method: our own news corpus global monitoring volume, each signal ranked against its own history — no keyword guessing, no invented ranges.
• Fear physics (fast up, max 3 pts/day down) and a capped event-shock overlay. First live detection: the Kuwait–Iraq border drone attack, +14 within minutes.
• Honesty layer: anything unavailable is reused from last-known-good and publicly flagged "estimated"; stale data shows "Delayed" — never wrong numbers.
• Score recalibrated from the old engine's 94 to the corrected 60s — published openly rather than hidden.

Design System 1.0 — the whole site, one language

• Every page rebuilt: live threat band, instrument gauges (red always right), rising-red/falling-green convention, dark mode, mobile-first layouts.
• News Pulse restored with per-story tension scoring, categories and filters; 2026 history rebuilt with the day-wise record; article library redesigned with content byte-preserved.
• Security & speed layer: security headers, compression, smart caching, config shielding.

Analysis library & regional tracking

The analysis library grows to 46 data-driven articles; regional tension breakdowns and the conflict watchlist expand.

The permanent ledger begins

Daily scores start being appended to a public, append-only record — one row per day, never edited. The ledger is the platform's long-term evidence.

Launch

World Tension Meter goes live: the composite index, seven sub-indexes, and the first methodology whitepaper.