Assessment language
Viokit states how likely something is using the UK Professional Head of Intelligence Assessment (PHIA) probability yardstick, the scale UK Defence Intelligence uses. This page covers the scale, how Viokit picks a term for each kind of source, how analytic confidence differs from likelihood, and how Viokit grades the evidence itself (Admiralty A1–F6, or UK 3×5×2).
The yardstick
| Term | Probability |
|---|---|
| remote chance | 0–5% |
| highly unlikely | 10–20% |
| unlikely | 25–35% |
| realistic possibility | 40–50% |
| likely (or "probably") | 55–75% |
| highly likely | 80–90% |
| almost certain | 95–100% |
The gaps between bands are intentional. They keep a term from implying more precision than the judgement behind it has. Viokit writes "likely", never "probably", and nothing in the scale means "certain".
When a source gives a number that falls in a gap, Viokit uses the nearer band. A number exactly midway goes to the band nearer 50%, which is the more cautious claim: 0.525 is a realistic possibility, not likely. A value outside 0 to 1 is refused.
Wherever a person sees how strong a link is, Viokit shows the term with its band, for example “highly likely (80–90%)”. The console's Likelihood indicator and the reports do the same.
How each source gets a term
A relation is rated by the projection that asserts it, never by a surface. There are two ways to rate one, and a relation can carry both.
By the source's standing. This is the confidence field, set by what the source is for the
fact:
| Ordinal | Typical source | Term |
|---|---|---|
confirmed |
registry records (Companies House, GLEIF by LEI), DNS answers, routing data, whois delegation | almost certain |
probable |
a page's address for the one company it names | likely |
possible |
a page mention, a username probe hit, whois registrant data | realistic possibility |
unverified |
a claim with nothing of its own behind it | unassessed |
unverified gets no term. A claim with no signal behind it has not been assessed. Calling it a
"remote chance" would be a confident judgement that it is false, which is not what anyone found.
By the source's own probability. This is the likelihood field, a number from 0 to 1 set by
sources that compute one, such as fuzzy matchers and entity resolution. Viokit keeps the number
as the source gave it, but people only ever see it as a term.
When a relation has both fields, it reads at the stronger of the two terms. When several sources assert the same relation, the folded graph keeps the strongest rating and the highest probability, and each source's own value stays in the step log. Two weak assertions do not add up to a strong one.
An unrated relation shows no term at all. It is not treated as weak.
In the tree. An entity takes the strength of its strongest path back to what the tree already held, where each path is as strong as its weakest rated link. You can ask the tree to expand only entities found at least as likely as a term such as “likely”. Entities found less likely than that are recorded but not expanded, and so are entities whose links are rated but unassessed. Entities reached only through unrated links still expand.
Merge proposals. When Viokit proposes that two entities are the same, it states the proposal as an assessment, for example an officer matched on name and birth month: “likely (55–75%), moderate confidence”. The match rule's probability gives the term (no probability means 1, almost certain).
Organisation resolution, which runs after every tree, proposes organisations that share a company
number, LEI or Wikidata QID as almost certain (95–100%), high confidence, and a name-only
organisation matching exactly one registry company's name as likely (55–75%), moderate confidence. The case report prints an almost-certain proposal as one entity and a likely one as a
judgement (see Case reports).
Screens that find nothing. A sanctions screen with no match gets no term. "No match under this name" does not make listing unlikely: the party could be listed under another name. The report states what was screened and when, and says so.
Analytic confidence
Likelihood is how probable something is. Analytic confidence is how far the judgement can be relied on: how good the sources are, whether they corroborate each other, and how much is inferred. PHIA guidance keeps the two separate, so "highly likely, low confidence" is a coherent statement.
Viokit uses three levels: high, moderate and low. At present only merge proposals carry one:
- high: the shared identifier is the identity (an LEI, a company number). The rule states no probability.
- moderate: a derived key suggests identity without establishing it (name plus birth month). The rule states a probability below 1.
- A match rule can state its own level, which overrides both of these.
An analyst's decision on a proposal (accept or reject) does not change its term or confidence. A merge the report applies because it was accepted keeps the proposal's own term (“likely”, say) and says who accepted it and when.
A relation's confidence field is the source-standing ordinal above. It is not analytic
confidence, despite the name.
Grading the evidence
The yardstick words a judgement. Grading rates the evidence behind it: how far each source can be relied on, and how far each fact is borne out. Viokit uses the Admiralty system for this, the one NATO uses (STANAG 2511, AJP-2.1), and can also show the same grade in the UK policing 3×5×2 form. The two axes stay separate from the yardstick: a fact can be graded A2, "probably true from a completely reliable source", while the judgement built on it is only "likely".
Source reliability, A to F
| Grade | Meaning |
|---|---|
| A | Completely reliable: no doubt about authenticity, trustworthiness or competency |
| B | Usually reliable: minor doubts, a history of mostly valid information |
| C | Fairly reliable: doubts, but has given valid information before |
| D | Not usually reliable: significant doubts |
| E | Unreliable: a history of invalid information |
| F | Reliability cannot be judged |
Every source declares its grade with a one-line reason. A source that declares none is F. The defaults, by family:
| Grade | Sources | Why |
|---|---|---|
| A | Companies House, HM Land Registry, The Gazette, Find Case Law, the UK and US sanctions lists, Electoral Commission, Parliament's registers, FCA, Charity Commission, CRO, RDW, Rechtspraak, DVLA, DVSA, SEC EDGAR, FEC, USAspending, Contracts Finder | official registers and lists: the record is the legal fact, though what is filed is not checked |
| B | GLEIF, Wikidata, OpenCorporates, OpenSanctions, Open Ownership, CourtListener, postcodes.io, Team Cymru, ICIJ Offshore Leaks | curated, a step removed from the register. ICIJ is B rather than A or C: journalists verified and curated the records, but they are leaked, dated documents, not a register |
| C | a company's own website (page contacts), DNS answers, certificate transparency logs, the web archive, The Guardian, passive subdomain sources, active scans | faithful about what was published or answered, but the publisher or target controls the content |
| D | whois registrant data, sherlock username hits, GDELT keyword matches, theHarvester | self-declared or weak: a username or a mention is often someone else |
An analyst can grade a source differently for one case, giving a grade and a reason. It is recorded with who and when, the latest decision stands, and a decision with no grade reverts to the source's own. Everything below uses the grade in force.
Information credibility, 1 to 6
| Grade | Meaning | How Viokit assigns it (first rule that applies) |
|---|---|---|
| 5 | Improbable | an A source contradicts it |
| 4 | Doubtful | another graded source contradicts it |
| 1 | Confirmed by other sources | two or more independent publishers graded A to D assert it |
| 2 | Probably true | one A or B source asserts it, not rated possible or unverified |
| 3 | Possibly true | one C source asserts it, or it is a possible link |
| 4 | Doubtful | only D or E sources assert it |
| 6 | Truth cannot be judged | nothing graded asserts it |
Corroboration counts publishers, not sources. Companies House's company profile and its officer list are one publisher, and asking the same source twice is not corroboration. A relation is corroborated by any relation of the same type between the same two entities, whatever its id. An identifier is corroborated by the same kind and value on the same entity, ignoring case and spacing.
A contradiction needs a fact that holds one value at a time, such as a company's registered office, its status, or a date of birth. Two publishers giving one subject different values over overlapping dates contradict each other. One publisher's own history (an office it says moved in 2015) is not a contradiction. Dates of birth are compared at the precision both give, so 1960-05 agrees with 1960-05-12.
The grade shown is the letter of the most reliable source asserting the fact, followed by the fact's credibility. For example, "A1" is a register entry that another publisher confirms, and "D4" is a whois registrant name and nothing else.
UK 3×5×2
The College of Policing's Authorised Professional Practice (Intelligence management: intelligence report) grades the source 1 to 3, the intelligence A to E, and gives a handling code. Viokit derives it from the Admiralty pair:
| 3×5×2 | Meaning | From |
|---|---|---|
| 1 / 2 / 3 | reliable / untested / not reliable | A or B / C or F / D or E |
| A | known directly to the source | an A source with credibility 1 or 2: a register knows its own record |
| B | known indirectly but corroborated | credibility 1 otherwise |
| C | known indirectly | credibility 2 to 4 otherwise |
| D | not known | credibility 6 |
| E | suspected to be false | credibility 5 |
| P / C | lawful sharing permitted / conditions apply | always P: Viokit holds open-source records only |
Where grades appear
- The console lists each source with its grade, reason and any override, grades every relation and identifier, and lets an analyst override a source's grade.
- The console shows a chip such as
B2beside a link's likelihood and beside each fact, with the meaning on hover ("Usually reliable source · probably true — one curated record (GLEIF)"). The case overview has a legend and a setting for Admiralty or UK 3×5×2, saved with the view settings. "Corroborated" counts entities that two or more publishers assert. - The case report cites the grade after each source in a judgement, for example
(Companies House [A1], high confidence). It gives each source's reliability and reason under Sources, and ends with "How evidence is graded".
- Exports carry the grade, including the link label in Maltego exports ([A1]).
References: "Intelligence source and information reliability", Wikipedia (the A–F and 1–6 wording, after US Army FM 2-22.3 appendix B and NATO AJP-2.1); College of Policing, APP Intelligence management, "Intelligence report" (https://www.app.college.police.uk/app/intelligence-management/intelligence-report).