Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.mareforma.com/llms.txt

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

The graph is the set of all claims and the relationships between them: supports[] links (what this claim rests on) and contradicts[] links (what this claim is in tension with). 
upstream_ref ──► ANALYTICAL claim (lab_a)  ──┐
                                              ├──► REPLICATED
upstream_ref ──► ANALYTICAL claim (lab_b)  ──┘
                    │
                    └──► DERIVED synthesis ──► (human) ──► ESTABLISHED
Trust is derived from this structure. The more independent paths converge on the same upstream evidence, the higher the support level. A single agent asserting the same claim a thousand times is still PRELIMINARY — independence of provenance paths is what matters.

​
supports[]

supports[] is the provenance chain. It records what a claim rests on — upstream claim_ids or reference strings (e.g. DOIs). 
graph.assert_claim(
    "Synthesis of A and B suggests mechanism M",
    classification="DERIVED",
    supports=["claim_id_A", "claim_id_B", "10.1038/some-doi"],
)
Leaving supports[] empty for a DERIVED claim makes the chain unverifiable. For ANALYTICAL and INFERRED claims, supports[] is optional but records what the finding is explicitly grounded in.

​
contradicts[]

contradicts[] documents explicit tension. When a new finding is in conflict with an existing one, both coexist in the graph — neither is silently overwritten. 
established_id = "..."  # ESTABLISHED finding

graph.assert_claim(
    "Contrary to the established finding, our larger cohort (n=1200) "
    "shows no significant elevation",
    classification="ANALYTICAL",
    contradicts=[established_id],
    source_name="cohort_2026",
)
Science advances by documented contestation. ESTABLISHED means human-validated evidence — not settled truth.

​
Epistemic distance

Epistemic distance measures how far a conclusion is from its raw data. A short chain of ANALYTICAL steps close to raw data is more trustworthy than a long chain of INFERRED steps, even if each step looks locally valid. The pessimistic rule: one INFERRED step in a chain means the full chain carries INFERRED-level epistemic fragility. Origin does not average — it takes the weakest link. 
raw data → ANALYTICAL → ANALYTICAL → INFERRED → conclusion
                                                  ↑
                                     full chain is INFERRED
This is why the classification field is recorded at assertion time and cannot be changed retroactively. The epistemic origin of every claim is permanently visible in the graph.

​
Graph fragmentation

The most common failure mode: two agents assert the same finding in different words without linking to a common upstream in supports[]. REPLICATED never fires because the graph cannot detect convergence without a shared anchor. The mitigation: idempotency_key as a convergence convention. Two agents using the same structured key converge on the same claim_id even with different text, without needing explicit supports= links — see the API reference assert_claim → Idempotency.