How the Crisis Bulletin warns and checks itself

NightWatch Labs, updated 2026-10-21

The Crisis Bulletin watches spot tokens across many exchanges and writes a warning when something that usually comes before trouble shows up. It then keeps a public score of how often those warnings were followed by an exchange acting on the token. This page defines a warning, explains severities, describes how the score is computed and ends with a glossary of the terms used on the bulletin.

What a warning is

A warning is a row written by a background worker that runs on a five-minute loop. Each row names an exchange, a trading pair, a warning type, a severity from 1 to 5, the data that triggered it, and the time it was issued. It starts with the outcome pending. Later checks move it to confirmed (the feared event happened) or false alarm (the condition cleared). A warning is a statement that a pattern was observed. It is not a prediction that the token will fail.

Main warning types	What triggers it
grade_drop	The token's liquidity grade on that exchange is D or F
grade_slip	The detailed grade is C- to D- while the overall letter is not D or F
depth_collapse	Order-book depth thinned out sharply
spread_spike	The bid/ask spread widened sharply
zero_volume	No trading for several scan cycles
exchange_notice_delist, _suspend, _margin_delist, _assessment	An exchange published a notice in that category
st_tagged	The exchange applied a special-treatment tag, its own distress marker
cross_exchange_flag	One exchange declared a new state on a token, and the token is also listed elsewhere
Other types	Volume anomalies and on-chain or social-layer signals also exist; the table lists the main ones

How severity is set

Severity is an integer from 1 to 5. For exchange notices the worker classifies the title text against fixed patterns, in this order of urgency: a spot delisting is 5; a margin delisting is 4; a deposit or withdrawal suspension is 4; an assessment or monitoring-zone move is 3; a network upgrade or swap notice is 2. Pure promotional and new-listing notices are skipped. For a liquidity grade drop, an F is 5 and a D is 4; a grade slip is 2.

Warnings also carry layers: the centralised-exchange market data layer, an on-chain layer and a social layer. When one token has warnings from two layers on the same day, the warnings are raised to at least severity 4 and given a convergence score of 0.7; with three layers they go to 5 with a score of 1.0. A badge such as “12 anomalies” mixes weak and strong signals, so open the card to see its severity.

The board layout

The Crisis Bulletin reads left to right as a funnel. Detection lists raw anomalies. Tracking shows the smaller Priority Watchlist under closer live monitoring. We Called It lists confirmed outcomes that a NightWatch warning preceded. A real event can skip straight to the last column with no earlier detection, and the score below is built to stay honest about that.

What “We Called It” counts

The headline rate answers one question: of the spot tokens that had a real, exchange-declared bad outcome, what share had an earlier NightWatch warning? The unit is the exchange and pair, not the token: a token with outcomes on two exchanges counts as two. The denominator is the set of outcome events, namely a special-treatment tag, a delisting notice, a suspension notice or a confirmed delisting. The numerator is the subset for which a qualifying warning was issued before the outcome.

The definition has changed more than once; the current one, B3.1, applies from 2026-10-01. It excludes or restricts the following:

Routine wallet and network maintenance, margin-only delistings, quote-pair removals whose base still trades on that exchange, tokenized equities and stablecoins are not outcomes.
A special-treatment tag counts only if it is a new state change, not one already in force.
A notice that duplicates an earlier one is counted once.
A warning issued at or after that same exchange's own first tag or caution designation on the pair does not count. Only a call made before the exchange showed its hand does.
A warning must come at least 24 hours before the outcome. Two types, spread spikes and zero volume, count only when the same pair also had a warning of the same type on each of the two previous days.
Perpetual markets are outside the spot universe and never enter the rate.

The headline window is 90 days, with a 60-day figure beside it and a breakdown by outcome type. The numbers are computed by a worker every 30 minutes and stored; the page only reads the stored result, so a slow computation can never block the page. The rule that earlier numbers are never restated is part of the method: a new definition applies from its start date, and the old rate is not shown next to it.

How to read the rate: it mixes two questions. How often does a liquidity-driven decline show up first in order-book data, and how often does an exchange make a business decision that no market data foretold? The second is close to unknowable from data, so a rate well under 100% is expected and not by itself a sign of failure.

The lifecycle ticker

Above the board runs a ticker of token lifecycle events: scheduled delistings with a countdown, first listings, promotion to an exchange’s main board, a token appearing on a new venue, assessment-zone placements, and deposit or withdrawal suspensions. Each source query fails independently, so one broken feed does not blank the strip. A countdown shows a notice date; some scheduled dates pass with the token still trading, and a separate worker corrects those.

Reading a warning card honestly

Check the type first. A depth collapse on a thinly traded token says something different from an exchange delisting notice.
Check the severity and whether several layers agree before giving a card weight.
A countdown is a notice date, not a certainty, and the original notice link is the evidence.
A warning that disappears from view is not forgotten: its count and outcome stay in a summary row, and warnings tied to a cited announcement, or used in the We Called It numerator, are kept for longer.
Compare the rate with its definition text and start date before quoting it anywhere.

Freshness and stale grades

A grade built from a seven-day aggregate that is more than three days old is labelled stale instead of being shown as current, and the bulletin does not issue new warnings off a stale grade. A warning issued earlier, while the grade was fresh, stays a countable call.

Glossary

Term	Meaning on the bulletin
Warning	A stored observation of a pattern that has preceded trouble. Not a forecast.
Severity	An integer from 1 (weak) to 5 (strong) assigned by the type of signal and raised when several layers agree.
Pending, confirmed, false alarm	The three outcomes of a warning.
Convergence	Warnings from two or more data layers on one token on the same day.
NW Grade	The A to F liquidity letter from spread and depth. See the Guardrail method page.
ST tag	A special-treatment designation applied by an exchange to a listed pair.
Assessment zone	An exchange's monitoring area for tokens under review.
Detection, Tracking, We Called It	The three columns of the board.
Priority Watchlist	The smaller set of tokens under closer live monitoring.
WCI	The We Called It rate: outcome events with an earlier qualifying warning, divided by all outcome events in the window.
D-N	A countdown to a scheduled delisting, N days away.
Lead time	Hours between the warning and the outcome.

The bulletin reports what was observed and how often earlier observations lined up with later exchange actions. It is not a list of tokens to buy or sell.

Informational only, not investment advice. See the risk disclosure.