Private Intelligence Terminal
What MERIDIAN is, what it reads, how it grades evidence, and what it refuses to claim. Ten answers, in the order people ask them.
MERIDIAN is a private intelligence terminal. It monitors 40+ open sources around the clock — instruments, institutions, and the narrative layer — and fuses them into one daily briefing across six domains: geopolitics, economics, society, technology, health and regulation. It is a hypothesis tracker, not a news reader: each briefing commits to probabilistic forecasts with explicit resolution criteria, and every forecast that resolves is scored in public.
People who need to know what is changing before it is news, and who would rather read a probability than a headline — analysts, journalists, researchers, and operators with geographic or supply-chain exposure. It is geopolitics-first with peripheral vision across the other five domains. It is not a curated news digest and not a trading signal.
From open sources, through their public interfaces: public APIs, RSS feeds, government and institutional publications, public data portals and public channels. Nothing is bought and nothing is taken from behind a paywall or a login. The entities MERIDIAN tracks are countries, organisations and companies; it keeps no records about private individuals. By category:
Each event is stored with its source and, where the source publishes one, a link to the original. What MERIDIAN publishes is its own analysis of those sources, not a copy of them.
Yes, deliberately — and always labelled. Iranian state television, the Russian Ministry of Defence's own channel, pro-Russian OSINT aggregators and their counterparts are monitored because what a belligerent chooses to claim is itself information. Three rules keep that from becoming contamination:
Coordinated framing across same-bloc outlets is detected and strengthens the manipulation assessment. It never raises a tier.
How much independent evidence stands behind a signal, and of what kind. Convergence counts source types — sixteen of them, with feeds derived from the same phenomenon collapsing into one — so a story republished by a dozen outlets is not twelve pieces of evidence. Evidence class then sets the ceiling:
A HIGH or CRITICAL signal that goes 48 hours without fresh physical corroboration drops one tier automatically; market prices do not count for that test. Note that the same five words label a second scale: on a forecast, the tier is the probability band committed to, not the convergence.
The Brier score measures forecast accuracy: 0.000 is perfect, lower is better. Guessing 50% on everything scores 0.250 — so anything meaningfully below that beats a coin flip. It rewards being confident and right, and punishes being confident and wrong.
The current score is read live from the public track record rather than printed here, because a number frozen at build time starts lying the next time a forecast resolves. Read the sample size beside it first: calibration figures mature around the ninety-day mark and are provisional before that. When a correction makes the score worse, the worse score is published with the reason — see the live forecast track record.
Every forecast is written with a resolution criterion — who confirms what, by when — checkable against wire reporting or a named data series. Vague, undated, sourceless, compound or conditional criteria are rejected before the forecast is stored, and a rejected forecast survives into the briefing as a watch item with the reason it failed.
Resolution is performed by a separate model that is blind to the probability, tier, direction and original wording: it sees only the criterion, entity, horizon and evidence, and must quote the decisive line. State media may corroborate a verdict but never decide one alone. The score is arithmetic, not judgment. Numeric thresholds skip the model entirely and resolve against stored observations of a wired data series, where a negative verdict additionally requires that the series had coverage over the window — missing data must never look like a non-event.
A resolution auto-approves only when the auditor was highly confident, its quoted evidence was matched back to a real stored event, and a 24-hour veto window has passed. Medium and low confidence verdicts, and every void, wait for a human. Approvals are attributed and reversible for 14 days.
The briefing still publishes, at the same time, every day. What changes is what it says: absence of signal is information, so quiet domains are reported as quiet and a coverage-honesty section states what the system cannot see. Thin evidence produces an explicitly low-confidence read rather than a promotion of the day's least boring item.
Open questions that did not move are re-quoted at their existing price — an unchanged price is still a data point. Situations the system will not put a number on stay watch items: named, tracked, unscored. Only the midday and evening passes are conditional; and if the model chain fails outright, a deterministic report is assembled from the same evidence with no model in the loop. Publishing nothing is not an available outcome.
Nothing, during the beta — no card, no trial clock. You enter an email address on the landing page and an access key comes back by email; the key opens the terminal in a browser (the daily briefing, the signals underneath it, the six-domain scan, the divergence read, the forecast board and the track record). There is no password and no account to set up.
The beta is dashboard-first: the briefing lives on the web, and the only mail MERIDIAN sends you is the one carrying your key. What is stored about you, and what is not, is on the privacy page.