What this instrument cannot tell you

Every reading on this site carries a section on what the data does not support. Those sections keep repeating the same handful of limits, because the limits are properties of the instrument rather than of any single day’s finding. They belong here, once, so that each reading can carry only what is specific to it.

Counts are capped before they are counted

Each entity and relation carries at most four evidence items. That four is an export limit, not a census. A relation resting on “four headlines” may rest on four hundred; the file cannot distinguish the two cases, and neither can a reader. Where a reading says a claim rests on a narrow base, it means the base shown is narrow.

Co-occurrence records a pairing of words, not a relationship

Evidence for a relation is selected by finding headlines that carry both names. A headline naming Iran and the United States in one sentence is evidence that a publisher wrote both names together — nothing more. It is not evidence that the two interacted, and the direction of any interaction is not recoverable from the pairing.

Region tags over-grip

Articles are assigned to a theatre by tag. Measured over fifteen days, the TAIWAN_STRAIT tag classified 165 articles, of which 106 fell into “other”; sampled items from that residue included a Singapore recycling policy, K-pop coverage and a school dispute in Cape Town. Counts on a thinly-covered theatre therefore measure the classifier as much as the theatre.

Absent sensing is not absent activity

The physical layers — thermal detections and radar contacts — only see where they are pointed. A radar scene tasked over one strait says nothing about any other body of water; a thermal pass that recorded nothing east of a given longitude did not look there. Absent tasking, absent contacts prove nothing. Any reading that treats an empty layer as quiet is misreading it.

Reason lines are inherited, not written

Where a relation carries a stated reason, that line is carried over from the upstream scoring. It is not a summary of the evidence attached below it, and the two can disagree — one axis is filed under air-defence incursions while none of its four evidence items describes one.

Scores are outputs, not findings

Two entities carrying the same score are not thereby symmetrical, comparable or equivalent. A score is what the scoring produced. Where two numbers match, that is a fact about the scale, and any meaning read into the match has been added by the reader.

Narrative labels are machine syntheses

A narrative label is generated over headlines. Its article count and trend are measured directly; the label above them is not, and stays a rung below corroborated for that reason. A narrative flagged steady describes steady coverage — not a steady world.

The same figure can sit under two entities

A single reported number can enter the file twice, attached to two different entities with two different geographies, and nothing in the file settles whether that is one event reported twice or two events. Deduplication is by link, not by claim.

Where these limits came from

None of them was written in advance. Each was found by drafting a reading and then checking it: the export cap and the word-pairing surfaced while working the Iran–US axis, the over-gripping region tag and the inherited reason line while working China–Taiwan, the machine-synthesised label while working European energy transmission, matching scores while working the eastern axis, and the doubled figure while working Tehran.

That is why they sit on one page rather than in each reading. Stated five times over, side by side, they add up to a claim none of them makes on its own — that the instrument measures a great deal and knows very little. Stated once, they say what they are: the known edges of a working instrument, named by the people operating it.