A researcher finds that a specific claim appears in 40 separate documents. Standard tools report high consensus. Every one of the 40 traces back to a single original source — a working paper that was never peer reviewed. The claim has been amplified, not corroborated. Nothing in a standard research workflow catches this. The claim makes it into the report with high apparent confidence.
The difference between corroboration and amplification is the most consequential thing you cannot see in a conventional evidence review.
The confidence formula's diversity weighting scores independent providers separately. A claim corroborated by three independent Tier 1 sources scores materially higher than the same claim cited by 40 papers tracing to one. This distinction is structural — it cannot be done by reading; it requires tracking source lineage.
Research reports what is known. Professional incentives push against explicitly naming what is not known — gaps look like incompleteness. But in investigative and accountability work, the gap is often the finding. "There is no reliable evidence on X" is a result, not a failure. Epistamate makes gaps first-class outputs with importance ratings.
A 2018 study supporting a claim about AI capability is materially different from a 2024 study on the same question. The confidence formula applies a recency decay to older evidence — not because old evidence is wrong, but because its current applicability requires explicit acknowledgment, not silent inclusion.
An investigation team working on related stories rebuilds the same factual base repeatedly. What was verified in story A is relevant to story C — but it is in a notes document nobody has systematically organised. The knowledge graph preserves verified findings across sessions without requiring a separate documentation system.
Not a sentence in a summary — a typed assertion with a status (VERIFIED / CONTESTED / WEAK / UNVERIFIED), a confidence score, a source tier, a provider count, and a citation list. Each is individually addressable. A WEAK claim is shown as WEAK. A CONTESTED claim is shown with the challenge that contested it.
Before the brief is synthesised, Phase 3 challenges every claim. The counterargument, the alternative interpretation, the weakening evidence — all run as a structured phase. Claims that don't survive lose their socratic bonus and are flagged. The reader sees what survived scrutiny, not what the model produced unchallenged.
Documents you ingest are assigned a source tier and participate in the confidence formula directly. An authoritative primary document from a Tier 1 institution raises the confidence score of claims it corroborates. This is not just context passed to the LLM — it is evidence that changes the number.
The Decision Log records the full evidence state at the moment of logging — not what is currently known, but what was known and acknowledged at that specific moment. For investigations with long timelines, this matters: the question of what was knowable at the time of a decision is answerable from the log.
We're talking to research teams, investigative units, and accountability organisations in active development. We'd like to understand the specific evidence problem before describing the solution.