Origin

Origin is orthogonal to trust. It records where a claim came from, how it was derived, not how much it should be trusted. An ANALYTICAL claim can be untested. An INFERRED claim can be well-supported. They measure different things. Trust itself reads off the derived axes (status per content_id, question_status per frame_id; see Trust and Findings). The support_level ladder those examples used to name (PRELIMINARY / REPLICATED / ESTABLISHED) is the legacy stored axis, and v0.4.0 removes it whole: the labels, the stored column, the promotion machinery, and the query(min_support=...) filter. Both surfaces warn today. Origin is orthogonal to all of it. The API field is classification. This page explains what its values mean.

The three origins

INFERRED (default)

LLM reasoning, synthesis, extrapolation. The default.Correct to use even for sophisticated reasoning, as long as it is not grounded in data that actually ran. If the model is drawing on training knowledge, synthesising across papers, or extrapolating from context, it is INFERRED.

ANALYTICAL

Deterministic analysis ran against source data and returned output.Only use this when a real data pipeline ran and produced real output. If the pipeline failed silently and the agent fell back to LLM knowledge, the classification is still INFERRED: asserting ANALYTICAL on null data is an epistemic lie that the graph will permanently record.

DERIVED

Explicitly built on ESTABLISHED or REPLICATED claims already in the graph.The supports[] field must point to those claims. A DERIVED claim with empty supports[] is unverifiable: the graph cannot validate the chain.

Why this matters

The origin captures the difference between two claims that look identical as text but represent different epistemic situations:

# Pipeline ran, real omics data queried, output confirmed
graph.assert_claim(
    "IL-21 is overexpressed in SLE CD4+ T cells",
    classification="ANALYTICAL",
    source_name="medeadb",
)

# Pipeline failed silently: LLM answered from training knowledge
graph.assert_claim(
    "IL-21 is overexpressed in SLE CD4+ T cells",
    classification="INFERRED",  # honest
)

Both claims assert the same text. The origin is the agent’s declaration of which is which. On its own that declaration is self-reported, and a silent pipeline failure is exactly the case where an agent might declare ANALYTICAL and be wrong. Grounding is Mareforma’s independent check on the declaration: it watches the run and records whether the cited data was actually read.

The ANALYTICAL lie

The most dangerous misuse of Mareforma is asserting ANALYTICAL when the data pipeline returned null and the agent fell back to LLM knowledge. The graph records this permanently: future agents may build on it, reviewers may validate it, and the epistemic chain will be wrong at the root. The declaration is on the honor system. The grounding verdict is the execution-observed catch: an agent that declares ANALYTICAL on a run where no read was seen gets an UNGROUNDED or OPAQUE verdict next to its claim, and the gap is on the record for a reviewer to challenge. The rule: if you did not run deterministic code against real data and receive real output, the classification is INFERRED. Even if the answer looks right.

DERIVED: building on the graph

# Two independent agents established a finding
prior_id = "3f8a1b2c-..."  # REPLICATED claim

# A synthesiser builds explicitly on top
synthesis_id = graph.assert_claim(
    "Given the replicated finding, the likely mechanism is ...",
    classification="DERIVED",
    supports=[prior_id],
    generated_by="agent/synthesiser",
)

DERIVED claims make the inference chain explicit and traversable.

Introduction

Concepts

For agents

Reference

Examples

The three origins

Why this matters

The ANALYTICAL lie

DERIVED: building on the graph

​The three origins

​Why this matters

​The ANALYTICAL lie

​DERIVED: building on the graph

The three origins

Why this matters

The ANALYTICAL lie

DERIVED: building on the graph