Berkeley Protocol on Digital Open Source Investigations

The spoke’s first T1 source and the first thing in it that is a standard rather than a tool. Full title: A Practical Guide on the Effective Use of Digital Open Source Information in Investigating Violations of International Criminal, Human Rights and Humanitarian Law. Published jointly by the UN Office of the High Commissioner for Human Rights and the Human Rights Center at UC Berkeley School of Law; OHCHR dates its listing 03 January 2022 (ISBN 978-92-1-154233-2, UN sales no. E.20.XIV.4), issued in all six official UN languages.

It “identifies international standards for conducting online research of alleged violations of international criminal, human rights, and humanitarian law,” and gives guidance on gathering, analysing and preserving digital information “in a professional, legal, and ethical manner.”

Why it matters here

Every other source in this spoke describes a capability — llm-osint profiles a person, kallisto-osinter scrapes and clusters, strix runs the whole offensive chain, swaggerspy and gitsint enumerate. The Berkeley Protocol is the first that describes an obligation. It is evidence that the field this spoke documents has a written professional standard, produced by a UN human-rights body rather than a vendor or a repo maintainer, and that the standard’s three axes are professional, legal, ethical — not accuracy or coverage.

That reframes the AI-OSINT accountability question. The spoke had been treating analyst-in-the-loop as a design choice some tools make and others don’t — kafsiem builds provenance into every graph edge, the autonomous profilers don’t. The Protocol says that for one serious class of investigation it is a requirement with an institutional owner, which is a different kind of claim than a repo’s README making the same point.

What it covers

Methodology and procedure for gathering, analysing and preserving digital open source information — the preservation half is the piece with no counterpart anywhere else in this corpus, where sources are fetched and summarized but not held as evidence.
Investigator safety — explicit measures for the digital, physical and psychosocial safety of investigators. The psychosocial axis has no analogue in any tool page here; the tools model recon as a data problem, and this models it as work done by a person who can be harmed by it.
Protection of others — witnesses, victims and first responders (named as citizens, activists and journalists) who documented the events being investigated.

What it actually requires (full text read 2026-08-10)

The 102-page text was extracted this cycle and the structure is: I Introduction · II Principles · III Legal framework · IV Security · V Preparation · VI Investigation process · VII Reporting on findings · VIII Glossary, plus templates (an online investigation plan, a digital landscape assessment, validation considerations).

The principles, which are the enforceable part

Fourteen named principles in three groups. Professional: accountability, competency, objectivity, legality, security awareness. Methodological: accuracy, data minimization, preservation, security by design. Ethical: dignity, humility, inclusivity, independence, transparency.

Four of them bear directly on what this spoke’s tools do:

Accountability (¶25) — every step from identification through collection, analysis and reporting must be documented, and investigators must record the tools and software they used. Automation is permitted: “as long as documentation is consistent and sufficiently thorough, either manual or automatic methods can be used.” But with a condition that most of this corpus fails — “Automated processes and software must be understood by users and be explainable in court either by users or developers.”
Data minimization (¶31) — collect only what is (a) justified for an articulable purpose, (b) necessary for it, and (c) proportional. The Protocol states the consequence plainly: this “favours itemized, manual collection over bulk, automated collection,” while allowing that bulk may sometimes be appropriate. Over-collection is faulted for creating security vulnerabilities (investigators stop knowing what they hold), privacy exposure, storage cost, and downstream review and disclosure bottlenecks.
Preservation (¶32) — the mirror duty, against under-collection. Platforms remove content that breaks their terms even when it has probative value; users edit and delete; material is decontextualized, corrupted or lost. Absent a timely preservation request to the platform or independent capture, “such information may be lost forever.”
Accuracy (¶30) — multiple competing working hypotheses and peer review, named as the mechanisms that reduce biased selection and interpretation. Conclusions must not be overstated, and emotive language is to be avoided.

The investigation process

Chapter VI splits work into searching and monitoring, gates incoming material on relevance, reliability, removal, safety and subsequent duties, then runs three analyses:

Source analysis — credibility and reliability of the source, built from attribution, provenance, credibility, and independence/impartiality. The provenance rule is a precision the spoke should keep: refer to the “first copy found online,” never “the first copy online,” because closed channels (email, private groups) may have carried the item before it surfaced. Attribution is often impossible and is explicitly not required for authentication.
Technical analysis — the file and its metadata.
Content analysis — what is depicted; the named techniques are image/video comparison, image/video interpretation, spatial analysis, actor mapping, social network analysis, incident mapping, and crime/violation pattern analysis.

Against the tools in this spoke

The Protocol does not ban automation, and reading it as “keep a human in the loop” was too coarse. It imposes three specific tests, and the spoke’s autonomous collectors fail different ones. Explainability in court is fatal to an LLM profiler whose method is a prompt and whose output cannot be reconstructed. Data minimization cuts directly against bulk scraping — kallisto-osinter‘s collect-then-cluster shape is the pattern the principle names. Preservation is the duty no tool here implements at all: ironsight plots feeds and keeps nothing, and the rest fetch and summarize. kafsiem‘s per-edge provenance is the only feature in the corpus that answers an actual Protocol requirement.

Note also what the Protocol does not rank: coverage, recall, speed. Its axes are professional, legal and ethical, so a tool that finds more, faster, scores nothing here.

Extraction note. The 2026-08-07 version of this page recorded that the PDF “could not be text-extracted in this environment (no poppler)”. That was wrong about the environment, not about the attempt: pypdf reads it fine. The OHCHR-hosted copy 404s; the text above comes from a mirrored copy of the same publication (stoprapenow.org), 102 pages, checked against the OHCHR record for title, ISBN and chapter list.

osint — the discipline this standardizes
ai-osint — the automation the standard’s human-centred duties press against
kafsiem — the one tool here whose provenance discipline points the same way
llm-osint-reliability-study — the measurement half of the same argument, arriving the same day
synthesis