How the analysis on this site is produced, graded and corrected. It is stated so a reader can judge the products against a declared standard rather than trust the writing.
Scope
Open-source cyber threat intelligence: actor profiles, structured assessments and the tradecraft behind them. Every product states its boundary — what is observed, what is inferred, and what could not be verified.
Two-axis grading
Every claim carries a source and is graded on two independent axes, per claim, not per section:
- Confidence — how good the sourcing and the reasoning are: High (corroborated by two or more independent sources, where independent means separately collected — two reports that trace back to one origin are one voice, not two), Moderate (one credible source, or corroboration with gaps), Low (single-source, self-claimed, or heavily inferential).
- Probability — how likely the assessed outcome is, where the claim is estimative: nearly certain, very likely, likely, roughly even chance, unlikely, very unlikely. The terms and their bands follow the ODNI analytic standards (ICD 203 / ICD 208), so a stated range means the same thing to every reader.
The two are separate. A claim can be very likely at low confidence (a single source) or roughly even at high confidence (strong data, genuine uncertainty). Conflating them is the most common tradecraft error; keeping them apart is the point.
The ledger table in each product is generated from its claim registry (_data/claims/), not
written by hand: scripts/render_claims.py renders it and CI fails if the table is out of date. What
a reader sees is what the ceiling computed.
Tradecraft standards
| Standard | Requirement here |
|---|---|
| Sourcing | Every factual claim traceable to an identifiable public source. |
| Confidence | Each judgement carries an explicit confidence, and a probability where it is estimative. |
| Alternatives | The competing explanations and why they were set aside — see the ACH assessment. |
| Assumptions | Stated openly, not smuggled into the conclusion. |
| Gaps | What is not known, and why. |
| Review | Stated below. |
| Correction | Published judgements can be revised, visibly. |
Review — stated honestly
These products are single-author. They are not presented as peer-reviewed, because there is no second qualified analyst signing them. What exists instead is a self-check against the standards above, and an open invitation to challenge. If an independent review happens, it will be recorded here rather than implied.
Versioning and corrections
- Each product carries a version, a date and a review trigger in its header.
- Changes are recorded in the product’s Revision history; no assessed fact is changed silently.
- A correction that changes a judgement is published as a new version, not as an edit in place.
- A product carries a next review date, and
scripts/sitegate.pyfails the build once it has passed: an assessment that has outlived its review is not a product any more. - Every claim a reader would act on carries at least one review trigger: a documentary, technical or behavioural condition, the artefact that would satisfy it, and — for technical and behavioural triggers — an expiry date. An expired trigger fails the build. A weekly job checks the ATT&CK collection against the pinned version and opens an issue when it moves.
Handling
Products are TLP:CLEAR unless marked otherwise — open sources, shareable without
restriction. That is the TLP 2.0 name; the retired TLP 1.0 label TLP:WHITE is not used here.
The marking is checked rather than asserted: every product’s handling row is run through
sharegate in CI, and the build fails on a missing marking, conflicting markings or a
retired one.
Rule provenance
Detection content on this site is described as opportunities, not as validated rules: no rule is published without a stated provenance and a way to check that it still holds.
Related
All analysis is open-source. Views are my own.