Evidence first.
Then conclusions.

Separate signals

Attention measures coverage and platform interest. Evidence scores describe source completeness. Importance and urgency are heuristic triage labels. None is a truth probability.

Evidence scores

A single news publisher scores 30; further distinct publishers add 12, 8, 5 and 3 (maximum 58). Official channels start at 70; further publishers add 8, 6, 4 and 2 (maximum 90). Adding a report never lowers a score. Different publishers may repeat the same underlying source. Official statements are not independent verification.

Publication and indexing

News becomes eligible for indexing when it has official or multiple-publisher coverage and current-version copy in that language. A valid price alert becomes eligible after a complete same-source five-minute follow-up observation, even without model text. Nearby reports do not establish its cause. Topic archives need at least three eligible events. Markdown mirrors identify their canonical HTML page.

Grouping reports

Reports about the same coin within 36 hours are grouped when their headlines share enough distinctive words, normalised amounts or synonyms (for example hacked / exploit / drained). Different topics need a stronger match. The event keeps the most severe topic among its reports.

Importance and urgency

Importance combines topic (security, regulation, exchange, upgrade, market), evidence score, number of publishers, official source, core coin and a nearby price alert; 55 of 100 or more is high. Security reports are urgent for 12 hours; regulation, exchange events, official upgrades and reports near price alerts are flagged for 48 hours. These are triage labels, not truth judgements.

Minute-level price alerts

A scheduled job checks 1, 5 and 60-minute returns. Minimum absolute moves are 0.7%, 1.5% and 4%. With at least 20 earlier return windows, deviations from their median must also exceed four robust standard deviations (1.4826 × median absolute deviation; minimum scale 0.03 percentage points). Otherwise twice the absolute floor applies. These are uncalibrated heuristics.

Sampling and duplicate protection

At least 80% of expected samples must exist; endpoints use the same source and currency. Stale endpoints are rejected. Each coin and direction has a one-hour cooldown; at most two alerts are published per run. Missing ticks, delayed schedules and provider outages can delay or miss detection. This is minute-level sampling, not a seconds-level guarantee.

Evidence comparison and causation

Price alerts revisit monitored reports from two hours before the sampled window through one hour after it. Report timing is shown explicitly. Nearby reports are candidate context, never automatically confirmed causes. Alerts without a supported explanation remain unresolved. No minute-volume analysis is claimed.

Historical data and automatic repair

Closed hourly OHLCV is refreshed automatically, with a daily reconciliation of the last 30 days. Imports are validated for confirmation, timestamps, OHLC bounds and duplicate conflicts. Event statistics exclude candles that closed after the observation cutoff. At least 90% coverage and 168 adjacent hourly returns are required. BTC beta requires 168 matched returns from the same venue and currency. These statistics do not verify a news claim.

Priority, languages and models

Price alerts rank first; other events follow their importance and urgency. Waiting tasks gain limited age priority. English and Chinese use separate writing tasks with retries, shared facts, scores and a shared daily model budget. Data flashes publish before model text. A failed language does not block the other. Source instructions are ignored and generated citations and numbers are checked; semantic truth is not guaranteed.

Coin selection and attention data

BTC, ETH, XRP, SOL, DOGE and SUI have core coverage, plus six dynamic places. CoinGecko platform trends are not global search counts. Candidates need liquidity and repeated observations. Provider failures activate a fallback using news and recent liquidity. Google Trends is shown only when real data is configured.

Corrections and uncertainty

Source edits invalidate generated text and reopen both language tasks. Evidence changes are recorded in the timeline. No automated true/false verdict is issued without verifiable evidence. Market reactions are observations and cannot prove a news claim.