threat-intelligence 3 min read
From source to a defensible judgement: the method behind the tooling
| Product | Method note |
| Handling | TLP:CLEAR — open sources only |
| Version | 1.1 (2026-10-28) |
| Review trigger | A change to the method, or a case that breaks it |
| Next review | 2027-04-30 |
| Author | Martín Eliseo |
Bottom line
A judgement is defensible when a reader can see why it was reached and what would overturn it. This note describes the method I use to get there — and the tooling that encodes it so the hard part cannot be skipped.
The problem: the names drift
Ask five vendors who an actor is and you may get five names. Volt Typhoon, VOLTZITE, Insidious Taurus, BRONZE SILHOUETTE, UNC3236. Platforms then store those names as a flat list of aliases, which quietly turns a hypothesis into a fact: if two names sit together, they “are” the same actor. The reasoning is lost, the confidence is unstated, and a decision — an incident response, a procurement choice, a threat hunt — is made on ground that no longer shows why it holds.
I worked through one instance of this by hand in Is VOLTZITE the same actor as Volt Typhoon?. This note is the step from doing it once to doing it repeatably.
The method
Four rules do the work.
1. A merge is a hypothesis, not a fact. Two names are compared, never collapsed. The link between them is a graded edge that carries its evidence and can be split again by lowering a threshold — the reasoning is never deleted.
2. Evidence is typed by diagnostic strength. Infrastructure, code and operator artefacts can bear on identity; shared tradecraft cannot. This is the trap: two actors can share tooling, binaries and C2 paths without being one organisation, so TTP overlap must not, on its own, sustain an identity claim.
3. Two axes, per claim. Confidence — how good the sourcing and reasoning are — is kept separate from probability — how likely the outcome is. A claim can be very likely at low confidence (a single source) or roughly even at high confidence (strong data, genuine uncertainty). Conflating the two is the most common tradecraft error.
4. Truth is external, and disagreement is a result. A binary judgement is admitted only if independent sources actually assert it; a single source is not corroboration, and a disagreement between sources is recorded as contested, not decided. An alias list is weaker than an identity claim, and the method says so.
From method to tooling
Prose does not stop you skipping the hard part on a busy Tuesday; a program does. So each rule above is encoded. The tools are small on purpose — the interesting thing is the rule, not the line count. What they share is that they refuse to let a claim pass without its evidence, its confidence, and a statement of what would change it. The engineering page walks through each one.
What this does not prove
- The checks are heuristics where the truth is not a schema; they are stated as such.
- External truth for a specific alias pair is scarce; a corpus that is mostly contested is a finding, not a failure.
- The method is single-author. It is a standard, not a peer review.
The numbers
The suite is public and MIT-licensed, and it runs in CI: github.com/Falsifyer/Defensible-CTI. I don’t copy test counts into prose — a number in a blog post goes stale, and the whole point of the tooling is that the claim can be checked at source. What the CI checks:
- the reconciliation engine against an adjudicated corpus with a committed coverage baseline, with STIX output validated against the reference library;
- the handling gate against FIRST TLP 2.0 and STIX 2.1, where the ground truth is a schema and not an opinion;
- the hunting planner and the assessment ledger, whose refusals are the behaviour under test.
A judgement is finished only when it changes a decision — and when someone can see why.
Revision history
| Version | Date | Change |
|---|---|---|
| 1.0 | 2026-10-09 | First release. |
| 1.1 | 2026-10-28 | Handling corrected from the retired TLP 1.0 label to TLP:CLEAR; test counts replaced by the CI they come from. No assessed fact changed. |