OSINT pipelines break in three compounding ways before a finished subject profile reaches an analyst.
A subject appears as "Chen Wei" in a Reuters article, "陳偉" in a Hong Kong filing, and "Чэнь Вэй" in a Russian wire. A string-match pipeline returns three identity records. The analyst de-duplicates manually — and misses the fourth variant in a Farsi-language source entirely.
A subject profile lists a date of birth and a corporate affiliation. Neither carries a source reference. A reviewer cannot determine whether the DOB came from a reliable registry or a poorly-translated news article. The attribute propagates through the system, compounding the error.
Two documents place the same subject at different addresses. A de-duplication pipeline picks one by recency and discards the other. The conflicting evidence disappears from the profile. An analyst making a targeting decision has no way to know the record is contested.
OSINT sources arrive in every script — news wires, regulatory databases, court filings, social-media exports. Kloom resolves entity mentions across Arabic, Cyrillic, Hebrew, CJK, and Latin without normalising names at index time. Cross-script retrieval is a query-time operation that preserves original forms and avoids transliteration ambiguity.
Each subject profile is a cluster of typed attributes extracted from source documents. Every attribute carries its source reference, extraction confidence, and temporal scope. The PKG does not collapse uncertainty: conflicting attributes coexist as probabilistic edges until an analyst resolves them with an explicit decision.
When two sources disagree on an attribute — different addresses, different dates of birth — Kloom surfaces both with their evidence weights rather than silently picking one. The profile carries a certainty score per attribute; a reviewer sees the conflict inline, not buried in an audit log.
An investigation opened today can be reproduced exactly six months from now against the same evidence graph and cluster state. Session pinning addresses the evidence record by version, not by snapshot copy: no duplicate storage, no drift, and a complete chain of custody from query to decision.
Kloom extracts typed attributes from each ingested document and anchors them to their source. The PKG stores every claim as a probabilistic edge — not as an overwriting update. When a second document corroborates a claim, the evidence weight increases. When it contradicts, the conflict is preserved and surfaced.
A single-source attribute carries a lower certainty score than one corroborated by independent documents. That distinction is carried through to every downstream consumer — screening tool, analyst workbench, or regulatory report.
“We persist the source — not just the verdict.”
Cross-script parse–retrieve–rerank. Resolves OSINT mentions in the script they arrived in — no index-time transliteration.
standalone libraryAppend-only record of every ingested document. Each attribute traces to its source passage, with confidence and character offsets preserved end-to-end.
Continuous reshaping as OSINT evidence arrives. Cluster boundaries update without destructive merges — every split and merge recomposable.
Session pinning against a versioned evidence record. Reproduce any investigation exactly against the graph state at decision time.
A demo runs against a sample OSINT corpus, resolves subject mentions across scripts, and returns a sourced profile your analysts can trace from attribute to document. Thirty minutes, with engineering in the room.