ModelBrain
Certificate of inheritance · ModelBrainExhibit 1 of 4 · real public model · as of 18 Oct 2026

The card says apache-2.0.
The weights say Llama 3.2.

§1 · Your terms

A company headquartered in may ship commercially, if the graph says so.
or try

§2

Chain of title.

  1. 0

    your from_pretrained line

    OPEA/Llama-3.2V-11B-cot-int4-sym-inc

    card declares apache-2.0

  2. 1

    quantised from

    Xkev/Llama-3.2V-11B-cot

    card declares apache-2.0

  3. 2

    fine-tuned from

    meta-llama/Llama-3.2-11B-Vision-Instruct

    card declares Llama 3.2 Community License

“For the multimodal Llama 3.2 models, the licence rights are not granted to companies whose principal place of business is in the European Union.”

Clause summary · quoted at runtime from Meta's official licence text

Blocked · BerlinWith obligations · New York

One Cypher query walks every path, through every merge and quantisation, and collects each licence's clauses. 17 ms, end to end.

Open this verdict →

§3

Exhibit A: what the weights say.

A card can be wrong. Weights can't. ModelBrain reads about 2 KB of a model's actual weights, 1,024 values, and compares them in FalkorDB's vector index. Fine-tunes stay close to their parent; independently trained models land near zero.

tier 1 at 0.60

TinyLlama vs Llama 3.2 1B-0.006
Same model, different layer0.022
Qwen2.5-Coder vs Qwen2.5 (continued training)0.560
Bielik-7B vs Mistral 7B0.840

−0.1 unrelated

1.0 same weights

Derived · 0.99 to 1.0

  • Llama 3.2 1B Instruct vs base0.9918
  • A grammar fine-tune vs its base0.9995
  • Re-upload with a new card1.0000

Each dot is a real measurement · cosine similarity of sampled weights · 18 Oct 2026

§4

Schedule 1: where metadata and weights disagree.

381

models fingerprinted from their weights

169

match a parent their card does not declare

119

of those trace to a licence with conditions

As of 18 Oct 2026 (graph not reachable)

ItemModelCard declaresWeights trace toSimilarity
Read the full schedule →

Many cards explain lineage in prose; these rows are about the machine-readable metadata. Every row links to the Hub so anyone can check it.

§5

Amendment, with no notice served.

  1. 19 Nov 2024Xkev/Llama-3.2V-11B-cot is published; the card declares apache-2.0.
  2. 9 Jul 2025Discussion #12, “License compatibility”, is opened.
  3. 16 Nov 2025LICENSE_llama.txt is uploaded (commit e8c2529775). 130 days.
  4. TodayThe card still declares apache-2.0. None of its 9 children received the change.

When something changes upstream, ModelBrain walks every descendant and reopens each company decision that rested on one of them, as a review task for the person who owns it.

Replay the amendment →

§6

Countersigned by two parties, trusted by all.

Harbor Apps Inc.

Signed · fictional company · stored in the hub as a salted hash

your company

Sign here · tier 1 → 3 · strict verdict flips

A re-upload declares no parent; its fingerprints say Llama 3.2 1B. A strict company refuses inferred lineage until two independent companies confirm the edge. Then it is trusted for everyone, and nobody learns who confirmed.

Countersign in your sandbox →

§7

Obligations of the agent.

(a) Before calling from_pretrained, the agent shall call can_i_use.

(b) Before a pull request is merged, the pipeline shall run modelbrain-check.

(c) Where the weights speak, the agent shall ask trace_weights.

$ claude "add OPEA/Llama-3.2V-11B-cot-int4-sym-inc for our German product"

→ modelbrain.can_i_use(model_id, hq_country="DE")

VERDICT: BLOCKED

Path: OPEA → Xkev → unsloth mirror → meta-llama/Llama-3.2-11B-Vision-Instruct

"I didn't add the model, and app.py is unchanged. The licence screen came back BLOCKED for Kestrel."

Excerpt of a recorded Claude Code session against ModelBrain's MCP server

Connect your agent →

§8

Governing engine: FalkorDB.

(a)

Many graphs, one instance

A shared hub, one private graph per company, a sandbox per visitor cloned with GRAPH.COPY.

(b)

Vectors inside the graph

Weight fingerprints in a cosine index, searched and filtered by graph properties in the same query.

(c)

One query, weights to clauses

Sampled weights → nearest parent → trusted ancestry → licences → clauses. A few milliseconds.

(d)

Bounded traversal

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

(e)

Algorithms in Cypher

WCC unmasks re-uploads, PageRank ranks relabels, betweenness finds chokepoints, label propagation finds families.

(f)

Trust on the edges

Tiers, salted confirmations and merge shares read inside the path. Stop FalkorDB and every answer stops.

§9

Definitions.

“Weight DNA” means
1,024 values sampled from two tensors with HTTP range reads, about 2 KB, never the whole file. Same scheme for safetensors and GGUF.
“Tier” means
Trust in a lineage edge: 0 weak inference, 1 strong inference or weights, 2 declared on the card, 3 confirmed by two companies, 4 confirmed by the author. Strict mode needs 2 or more.
“Ruling” means
A company's decision about one deployment, append-only, resting on the facts it used. A change upstream reopens it; approving the review supersedes it.
“Sandbox” means
Your private copy of a fictional company. Everything you click writes there and expires after two hours.
“Hub” means
The shared graph of public facts. It never stores which company uses which model; confirmations arrive as salted hashes.
“Verdict” means
Allowed, allowed with obligations, blocked, or cannot determine. Deterministic code over FalkorDB rows; a language model may only phrase it.

§10 · Execution

Sign nothing
you haven't traced.

Check a modelclaude mcp add --transport http modelbrain https://api-production-fc09a.up.railway.app/mcp