Investigative AI Must Pass an Independence Test

Analysis | October 14, 2026.

TRM launched Ardence on October 13. Announcement. The launch raises a question for buyers of investigative AI: when a system returns several supporting records, how many separate observations actually support the conclusion? A long bibliography can conceal a short evidence base. Our proposed evaluation starts by testing dependence between sources, before judging the elegance of the resulting narrative.

Count observations before counting agreement

Consider a fictional exercise. Three commercial feeds repeat a company address from one registry entry. A fourth report quotes one of those feeds. Four search results now appear to agree, but the exercise contains only one original address observation. If that entry is stale, repetition does not correct it.

A reviewer should ask where each assertion originated, which records copied or transformed it, and whether any collection happened independently. Different supplier names do not settle those questions. Nor does a separately written AI summary create another observation. Where the upstream relationship is unknown, our suggested label is “independence unestablished,” rather than silently treating the records as either independent or identical.

This distinction matters even when every citation works. NIST’s AI framework cautions that transparency alone does not establish accuracy. A retrievable document can show exactly why a system reached a conclusion while leaving the conclusion poorly supported.

Make duplication a controlled experiment

We propose a small synthetic test pack with known source relationships. Start with one underlying observation, then supply the same information through several differently formatted records. Keep the substantive evidence constant. Ask whether the system's stated confidence, ranking of hypotheses or recommendation changes merely because the packaging multiplies.

Next add a genuinely separate observation that could change the assessment. The desired response need not be a fixed numerical increase in confidence: a new record may be weak, contradictory or irrelevant. What matters is whether the explanation identifies what the additional observation establishes and which uncertainty remains.

Finally, remove one entire source family. If the conclusion depended on that family, the system should expose the dependency. This is a sensitivity test, not a rule that every claim requires several witnesses. Some observations can be decisive on their own; several dependent observations can still be useful for tracing how information spread.

Give the competing explanation a fair test

U.S. intelligence analytic standards call for alternative hypotheses; they do not govern purely law-enforcement information. We use that principle as an analytical reference, without treating it as a universal compliance requirement.

For our fictional address example, write down two explanations before further searches: the entities share an operator, or they use a common service address. Specify a permitted observation that could distinguish them. Then ask the system to pursue that discriminator, rather than simply collect more address matches.

The test should reward an explicit revision. If new evidence supports the common-service explanation, the earlier inference should weaken, and the affected conclusions should be identified. An entity match, a relationship hypothesis and evidence of misconduct belong in separate fields. Moving between them requires an argument; a visually persuasive network diagram cannot supply it.

Preserve the boundary of what is known

A useful reviewer record would show the proposed conclusion, independent observation families, unresolved dependencies, the strongest competing explanation and the next discriminating question. It should also state what was not checked. “No result returned” needs the search scope and time window before it can be interpreted as meaningful absence.

These are proposed evaluation exercises, not measured Ardence results or allegations about its design. We have not independently tested the product. The purchasing question is whether a system can keep an investigation open when the evidence remains open, and narrow it when genuinely informative evidence arrives. More retrieved material is valuable only when the team can explain what, if anything, it has newly learned.