What this measures, and what it does not
Archive 1979-01 to 2019-12 · registered knowledge cutoff 2019-12.
The archive is a monthly event-mention share under frozen actor-pair filters. It measures press attention. It does not measure conflict, risk, escalation, casualties or outcomes, and nothing on this page is a forecast. A candidate break is a break in attention. An analog is a similarity between two months of one measured series — it carries no claim that they shared circumstances and none whatever about what followed either of them.
This is also a different construct from the live 2017+ instrument and is never spliced into it or drawn beside it. Comparing a 1984 number here against a 2026 number on the front page is not a comparison of like with like.
Two channels of five publish
A channel appears here only where the registered 2017–2019 overlap audit returned a tracking verdict. A low correlation is a published negative finding, not a channel to repair, so the refusals travel with the data rather than being dropped from it.
| Channel | Published | Reason |
|---|---|---|
| pakistan_west | yes | overlap audit verdict: tracks |
| china_east | yes | overlap audit verdict: tracks |
| us_trade | no | overlap r=0.216 over 36 months; source payload verdict DOES NOT PUBLISH (negative finding) |
| gulf_energy | no | overlap r=0.153 over 36 months; source payload verdict DOES NOT PUBLISH (negative finding) |
| shipping | no | excluded at registration: no defensible historical event analog for the channel's construct |
Regime baselines
Periods are calendar decades, named by their years and nothing else. They are not political regimes, administrations or eras, and no boundary asserts that anything changed at it — naming a period after an era would be a historical claim this instrument cannot support from its own artifacts. Every statistic states the denominator it was computed over, and a period with fewer than 24 observed months reports unavailable with its count instead of a mean over the survivors.
| Channel | Series | Period | Months | Observed | Coverage | Mean | Median | p90 |
|---|---|---|---|---|---|---|---|---|
| This table is rendered from historical_intelligence.json so the values on screen are the published values and cannot drift from them. Without JavaScript, use the baselines CSV, which carries the same rows including the unavailable ones and their reasons. | ||||||||
Candidate structural breaks
The scan takes the largest absolute Welch t between the two segments either side of every admissible split, against a seeded permutation null that shuffles the values to destroy time ordering while keeping their distribution. It is a descriptive scan, not a change-point model with a stated generative assumption.
The sweep is the honest part. A single break month reported alone reads as a finding; the same month reported beside the settings that move it reads as what it is. Where the candidate moves with the minimum segment length, this page says so.
Rendered from the published payload. Without
JavaScript, the candidate breaks, their statistics, p-values and the full
sensitivity sweep are in
historical_intelligence.json
under structural_breaks.
Analog retrieval
Choose a channel and a month. The lab returns the nearest other months in the same channel by standardised distance over three registered features, and for every match it names which features it used and which were null for that pair. A null is never imputed and never replaced by zero; a pair with fewer than two usable features is refused rather than guessed. Months within six months of the query are excluded because adjacency is trivially similar.
A near neighbour is a month whose measured attention looked similar. It is not a precedent, not a cause, and says nothing about what happened next in either month.
The analog explorer needs JavaScript. Every query and its result is also published as a static row in the analogs CSV — one row per query month and rank, with the distance, the features used and the features excluded as null.
Event archetypes
An archetype is attached to a month only where a human-authored anchor was registered. Nothing here is machine-labelled. A month with no registered anchor reports unavailable and is never given an inferred one, and an anchor for a channel that does not publish is retained for the record and marked unavailable with that channel's reason.
| Month | Channel | Archetype | Available |
|---|---|---|---|
Rendered from the published
payload; see event_archetypes in
historical_intelligence.json. | |||
Provenance and downloads
The payload pins the SHA-256 of the three things it was built from — its registered contract, the source archive and the implementation — so a reader can regenerate it and get the same bytes. The permutation null is seeded and ties break on ascending month, so the same query is the same answer on every run and platform.
- historical_intelligence.json — the citable machine payload
- Regime baselines (CSV) — including unavailable rows and their reasons
- Analog retrieval (CSV) — every query month, rank, distance and feature availability
- The 1979–2019 archive itself
- Codebook entry
- Contract
- —
- Source archive
- —
- Implementation
- —
This payload is exempt from the daily freshness window on purpose: it derives from a frozen archive with a registered 2019-12 cutoff, so it cannot go stale by the passage of time. It is republished only when its contract, its source or its implementation changes.
Limitations
- This measures press attention in a monthly event-mention share, not conflict, risk, escalation or outcome.
- Two of five channels publish. The other three are refused on registered grounds and the refusal is part of the output.
- A candidate break is a property of the series under a declared descriptive scan; a different algorithm would place breaks differently, which is why the sensitivity sweep is published beside it.
- Analog distance is a similarity over three registered features of one series. It is not a claim of comparable circumstances and it is not evidence about what followed.
- The trailing-percentile feature is undefined for the first 120 months, so early-archive analogs rest on fewer features, and each result reports how many it used.