Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mareforma.mintlify.app/llms.txt

Use this file to discover all available pages before exploring further.

​
The problem

AI scientists are being deployed on real research problems — including drug targets, clinical hypotheses, and materials discovery — before any infrastructure exists to answer a basic question: which of their findings can be trusted? The problem is not that AI scientists get things wrong. Every scientist gets things wrong. The problem is that the current generation of agent systems provides no principled way to distinguish:
  • a finding backed by data from one backed by LLM prior knowledge
  • genuine independent replication from two agents repeating each other
  • an established consensus from a single speculative assertion
Without that structure, every output looks like a result.

​
What other tools address

Tracing and observability tools record what the agent did: which tools were called, what was returned, how long it took. This is useful for debugging. It is not epistemic infrastructure. Knowing that an agent called a retrieval tool does not tell you whether the retrieved finding was independently replicated or whether the data pipeline actually ran. Structured outputs and validation schemas catch format errors. They do not catch epistemic errors: a correctly structured claim can still be wrong, ungrounded, or a duplicate of something already known. Vector stores and retrieval give agents access to prior knowledge. They do not record which of that knowledge has been independently validated, which is contested, or which paths the current claim was derived from.

​
What Mareforma does differently

Mareforma is not an observability tool. It does not record execution traces. It records epistemic objects (claims with provenance) and derives trust from the structure of the graph, not from the agents that populate it. Trust from topology, not introspection. Two independent agents reaching the same conclusion through different data paths is a stronger signal than one agent reporting high confidence. Mareforma detects this automatically. Origin as epistemic honesty. The distinction between ANALYTICAL (data ran) and INFERRED (LLM reasoning) is recorded permanently at assertion time. An agent that asserts ANALYTICAL on a silent pipeline failure is making a claim the graph will preserve, and that a reviewer can challenge. Accumulation, not evaporation. Findings persist across runs. When a new agent asserts something, it can query what is already established and build on it with DERIVED classification. Knowledge compounds instead of resetting with every run. Documented contestation. When a new finding contradicts an established one, both coexist in the graph with an explicit contradicts[] link. Science advances by documented tension, not by one side being silently overwritten. Human oversight at the right layer. Agents accumulate at machine speed. Humans review at the ESTABLISHED gate: not per step, but per finding that has earned independent replication. The graph does the accumulation; the human does the validation.

​
Who this is for

Autonomous AI scientists: agents running multi-step research pipelines on real scientific problems. Mareforma gives them a shared memory where findings accumulate and can be queried before any new assertion is made. Labs building AI science infrastructure: teams that need to know which of their agents’ outputs are grounded in data, which are replicated, and which are speculative. Mareforma provides the epistemic substrate without requiring changes to the agents themselves. Multi-agent research systems: where multiple agents contribute to a shared knowledge base and convergence needs to be detected automatically across runs, agents, and data sources.

​
What Mareforma is not

Mareforma is not a guarantee of truth. The graph records what agents assert and what the provenance structure supports. It does not verify that assertions are correct. A REPLICATED finding can still be wrong if both agents were reasoning from the same false prior. Mareforma is not a replacement for scientific methodology. It is infrastructure that makes methodology legible: it records the difference between “two agents happened to say the same thing” and “two agents reached the same conclusion through independent data paths.” The distinction matters enormously. Mareforma makes it computable.