ModelBrain
Exhibit H · Governing engineFalkorDB

Obligations travel with the weights. The graph decides.

Licence obligations travel with the weights, through every fine-tune, merge and re-upload. ModelBrain keeps that real lineage in FalkorDB and lets the graph decide. A language model may phrase an explanation; it never decides.

Kinds of graph, one instance
3
Hops of ancestry, max
6
Trust tiers per edge
5
Agents
5

§1

One FalkorDB instance, three kinds of graph.

A company's confirmation reaches the hub only as a salted hash. Two different hashes raise an edge to tier 3 for everyone; nobody learns who confirmed.

GraphHolds
hubPublic facts. Models, DERIVED_FROM edges with trust tiers, licences, clauses, file fingerprints and weight DNA vectors. Never stores which company uses which model.
tenant_*One per company. People, products, deployments, append-only rulings and the facts they rest on, events, tasks and playbooks. Kestrel and Harbor are fictional.
sbx_*One per visitor. A GRAPH.COPY of a fictional company. Everything you click writes here, expires after two hours, and never reaches the demo companies.

§2

What only the graph engine gives us.

Stop FalkorDB and nothing answers: verdicts return 503, never a cached answer.

  1. (a)

    Many graphs in one instance, GRAPH.COPY

    A shared hub, one private graph per company, and a sandbox per visitor cloned in one command. A company's rulings never touch the hub.

  2. (b)

    Vector index inside the graph

    1,024 sampled weight values per model, searched with db.idx.vector.queryNodes and filtered by graph properties in the same query.

  3. (c)

    Vector search + traversal + clauses in one query

    From a model's weights to its nearest parent, up the trusted ancestry, to the licence clauses: one Cypher statement, a few milliseconds.

  4. (d)

    Bounded variable-length traversal

    Every ancestry path up to 6 hops, through merges, at a minimum trust tier; every descendant for the reverse fan-out.

  5. (e)

    Algorithms in Cypher

    WCC unmasks re-uploads by shared fingerprints; PageRank ranks relabel hubs; betweenness finds chokepoints; label propagation finds families; shortest path proves a derivation or its absence.

  6. (f)

    Properties on edges, read inside paths

    Trust tiers, salted confirmer hashes, weight similarity and merge shares live on the edge; strict mode and merge exposure are enforced in the traversal itself.

§3

Trust tiers on every edge.

Default verdicts use tier 1 and above. Strict mode uses tier 2 and above and answers CANNOT DETERMINE when a parent is only inferred.

TierEvidence for the edge
0Weak inference: config fingerprint only
1Strong inference: identical file hash, tokenizer + config, or weights that match
2Declared on the model card
3Confirmed by two independent companies (salted hashes)
4Confirmed by the model's author

§4

Agents and handoffs.

Agents hand work to people through Task nodes. Approving a reopened review writes a new ruling that supersedes the old one.

AgentDuty
ScoutIngests a model and its ancestors from the Hub, keyed by canonical id.
ResolverInfers hidden parents from file hashes (WCC) and from the weights (vector search).
AuditorDecides a verdict from the graph and writes an append-only ruling.
WatcherRe-reads watched models, diffs them, and reopens rulings down the lineage.
PublisherTurns a company's confirmation into a salted hash on the shared edge.

§5

Read the queries.

Automated screening, not legal advice.