Editorial process

1. Monitor

Monitoring is scheduled editorial work, not an automated feed. Each published module is reviewed on its stated cadence. All six published modules are reviewed every week. Each receives a dated weekly brief; quiet weeks explicitly record when no material change was identified. Each question is reviewed on the Weekly Brief cycle and can be updated between briefs when material evidence appears. Nothing on the site watches the world by itself: a record changes when an editor reviews it.

The inputs to a review are, in rough order of weight:

• primary official sources — statements, documents, data releases — from the parties and institutions involved;
• the standing sources in the registry, each with a stated role and limitation;
• the observables each open Watch question names — what the previous review said it was waiting for;
• the information gaps recorded on standing records, which are questions the last review could not close;
• new reporting from established media and specialist research institutions, and high-quality local or regional reporting where it is the only competent coverage;
• datasets from international organisations, agencies and market sources.

2. Triage

Triage is the editorial decision point. Every candidate development from a review is asked the same questions, and the answers decide where — and whether — it goes:

• Is it material, or is it the week's noise?
• Does it change the standing system — an actor's posture, a relationship, a force structure, a supply system, a market exposure, a civilian condition?
• Does a canonical record need updating, or a new one creating?
• Does it belong in the next Weekly Brief, and as what — an item with sources and a grade, or context?
• Does it need an Analytical Note, because the mechanism behind it needs explaining rather than summarising?
• Does it move a Watch question — meet an observable, change a condition, resolve or supersede the question?
• Is the evidence sufficient to publish at all, and at what confidence?

The categories are not exclusive: one development can be an event record, a profile update, a brief item and a Watch review at once. The honest outcome for a quiet lane is "no material change", and it is recorded as such rather than filled.

3. Research

• Primary sources are preferred, and a consequential or contested claim is cross-checked against independent reporting. A single report is not sufficient basis for a published claim.
• Chronology is preserved: the date a thing happened and the date it was reported are recorded separately, and nothing later than the review's cutoff is written as having happened.
• Disagreement is preserved. Where credible accounts conflict, both are recorded with what each rests on; Vigil does not pick a side to make the record tidy.
• Event, attribution and assessment are kept structurally apart — in the fields the record model provides for each — so that a contested attribution cannot take an established event down with it.
• Information gaps are identified and written down. What the evidence does not establish is a finding, and it is what the next review starts from.

4. Update the standing system

This is the step that distinguishes Vigil from a news digest. Where the underlying system has genuinely changed, the canonical record changes — before or alongside the publication that reports it, never instead of it. A Brief or a Note that described a change the standing record did not carry would be an isolated fact, detached from the system it belongs to, and the next reader would find two versions of the conflict.

The records a review may touch include actors, events, relationships, locations, terrain, logistics, campaigns, impact chains, economic and humanitarian analysis, formations in the Force Atlas, equipment, the module's Profile objects and its Watch questions. Only the fields the evidence supports are changed; a review date advances only when the material was actually reviewed to the standard that date implies; and a published Weekly Brief is never rewritten to match later knowledge — it is a dated snapshot, and a later correction goes through the corrections log.

5. Draft

A review produces up to four kinds of output, each with its own shape:

• Weekly Brief — the scheduled review product. Each item states its reporting window, the specific reports behind it, its registered sources and its confidence grade; a quiet review lists the standing files examined and records that none moved.
• Analytical Note — the triggered deeper analysis. A Note states the question or mechanism, what changed and what did not, what the evidence does not establish, the canonical records it connects to, its uncertainty, and the observables that would move the picture.
• Standing update — a change to a canonical record, with its own sources, dates and grade, and where applicable a recorded change linking back to the brief that reported it.
• Watch review — a dated ledger entry on a standing question, with one of the five review outcomes and the brief, event or Note it rests on.

Drafting happens against the site's record model, which requires sources, grades and dates on the fields that carry claims. A draft that lacks them does not build, let alone publish.

6. Review

The full change set is read before anything is published — the new records and the changed ones, not only the prose. The review asks, for each claim and each record:

• Does the evidence support it, at the grade given? Does the source say what the claim says it says?
• Are the dates right — event, report, assessment, review — and kept apart?
• Is the claim state honest: is a reported role shown as reported, a disputed attribution as disputed?
• Where does this contradict a standing record, and which of the two is wrong?
• What are the information gaps, and are they stated?
• Is the record internally consistent with the rest of the module — the same actor, the same name, the same date everywhere it appears?
• What does the map or the imagery imply that the evidence does not?
• Is anything operationally useful — a position, a strength, a route, a vulnerability — and if so, why is it here?
• Do the connected records resolve, and do they still describe the same system?

The question that has become the most useful is the plainest: what does this not establish? A Note answers it in its own structure, and every review is expected to answer it before the work moves on.

7. Validate

Before publication, automated checks run over the built site. They check structure and consistency: that every source reference resolves to the registry; that every internal link and fragment resolves; that dates are authored by the record that renders them and none is in the future; that records validate against their schema and connected records exist; that every image carries its provenance and no refused image ships; that identity marks resolve consistently across surfaces; that Watch objects keep their lifecycle, ledger and basis rules; that the search index and the sitemap describe the site that was actually built; and that no surface has acquired language Vigil has decided not to use.

Automation checks structure and consistency, not factual truth. A passing check means the record is well formed and its references hold; it does not mean the claim is right. That judgment is made in review, by a person, against the evidence.

8. Publish

Nothing publishes automatically. Drafting, validation and the tools that assist them can prepare a change; publication requires a deliberate editorial decision and is a manual act. There is no scheduled publishing, no automated release when a check passes, and no system — human-assisting or otherwise — that puts content on the site without that decision. Nothing goes live merely because an automated process says it is ready.

9. Maintain

Publication is not the end of a record's life:

• Standing records carry their own review dates and continue to be re-checked on their own cycle, not only when a development names them.
• Watch questions stay open until their answer arrives or a better question replaces them; a closed question keeps its page and its ledger.
• New evidence can raise or lower a confidence grade, advance a condition, or supersede an assessment; the superseded version stays traceable.
• Quiet review cycles are recorded as findings, so silence can be told from inattention.
• Analytical Notes are dated publications and become historical analytical records; a Note is not rewritten when the picture moves, the standing records and Watch are.
• Material errors are corrected in place and logged in public.

10. AI in the editorial process

AI tools assist several stages of the work described above: searching and organising material, comparing sources and accounts, drafting text and structured records, translation, writing and running code, and running the validation checks. Where they sit is bounded by four rules:

• A model has no editorial authority. It does not decide what is material, what grade a claim carries, or what is published.
• Model output is not evidence. A statement produced by a model is treated as a draft to be checked, and a source it names is treated as a claim to be verified — including that the source exists and says what is claimed.
• Every publishable claim must resolve to actual sources or canonical evidence in the registry, exactly as a claim drafted by hand must. The record model does not know or care who drafted a field, and it enforces the same requirements on all of them.
• No model publishes. Publication is the manual act described above, and responsibility for what is published remains with Vigil.

What AI output means epistemically — why it is never a source, and how its errors are treated — is stated in the methodology. The high-level disclosure is on the About page.

11. Corrections and accountability

The workflow is designed so that revision is a normal part of the record rather than an exception to it. Material errors are corrected in place and logged, with a date and the page affected, in the corrections log, which is never purged. What counts as a material error, and how a superseded assessment stays traceable, is defined in the methodology; the sources every claim rests on are in the source registry; the vocabulary is in the glossary.

Spotted an error? Write to [email protected]with the page URL and, if you have it, a source. Substantiated corrections are credited in the log unless you ask otherwise.