How We Work

We go upstream

We trace claims to where they actually started.

We watch behavior

A hospital closing a ward can tell you more than a press release.

We write down what would change our mind

So we cannot quietly move the goalposts later.

We close the file

Most Watches should end with nothing for the reader to do.

Read the current Watches, or open the details below.

Sources, claims and explanations

Origins versus relays

One original report copied by six outlets is one origin, not six independent confirmations. We trace links and attribution upstream, and show when that lineage is uncertain.

Claim classifications

We distinguish reported claims, official claims, observed behavior and analysis. Classification tells you what kind of evidence you are reading; it does not make a claim true. Evidence intake and human verification are separate labels. Standing means retained in the ledger, not independently confirmed. Contested means conflicting or unresolved evidence.

Official statements

Official statements are attributable claims, not independent confirmation. We record what an institution says and compare it with its actions, without inferring a hidden motive.

Information-control changes

An edited or deleted notice can change the public record. We preserve the wording and source context where possible, and label gaps. Deletion alone is not proof of concealment or of the underlying claim.

The boring explanation

We look for the ordinary explanation that could account for the signal. Seasonal restrictions, routine precautions and reporting errors must be considered alongside the concerning interpretation.

Levels, indicators and the permanent record

Escalation and de-escalation

We write down the evidence that would raise concern, reduce it, or close the Watch before it arrives. No level changes without named evidence and an indicator. A precaution does not by itself establish an outbreak.

Levels

WATCH: A real and potentially consequential situation is unresolved enough to justify continued monitoring. Reader action: none unless a specific affected group is named.
PREPARE: A specific, reversible, low-cost action is proportionate for a defined group. A precommitted indicator must fire, supported by more than a single uncorroborated rumor.
ACT: A concrete protective action with meaningful cost or disruption is proportionate. Evidence must meet high confidence or multiple independent indicators must converge. Two humans must review the recommendation.
CLOSED: The situation no longer justifies active monitoring. The page and its record remain available.

Source records

Watch 001 uses the existing newsroom ledger. Intake records show their source links, provenance, uncertainty and known gaps. Partial source snapshots and publisher-hosted images are available in the Watch 001 evidence packet. They are labelled with their preservation limits; full historical archives remain incomplete. These records have not completed human verification. Standing means retained in the ledger, not independently confirmed. Contested means conflicting or unresolved evidence. Retracted and superseded claims remain visible.

Observation dates supplied without a time remain dates. The capture timestamp is the newsroom record's timestamp; it does not assert that a remote page was archived. Receipt hashes cover stored records, not remote page bytes, and do not authenticate sources.

Permanence and review

Claims, timeline entries and corrections have permanent anchors. Updated wording may be displayed through a dated correction while the original wording, receipt and status history remain visible. Material errors receive visible corrections; retracted or superseded claims retain their original record. Each Watch displays its last review and next review due. Overdue reviews are marked. No level changes without named evidence and an indicator.

Human review, translation, AI and corrections

Human verification

A human opens and reviews the underlying source and context, and checks that the public claim accurately represents it. For languages the reviewing editor does not read fluently, the operative passage is checked against at least two independent translations. Discrepancies are resolved before the claim is marked human verified. Fluent-human review may be used when available; we do not claim it occurred when it did not. Agent review alone does not qualify.

AI use

AI supports search, translation, source comparison, monitoring, drafting and production. Human review is required for claims marked human verified and reserved editorial decisions. Some routine publishing operates inside pre-approved rules. How we use AI.

Corrections

Material errors receive visible, dated corrections. Original records remain available. Corrections and Changes explains the policy and how to report an error.