Braid’s Local AI Test Chooses Provenance Over Vector Portability

Intersignal is making a narrow but consequential choice about portable AI context: when two local models differ, do not force one model’s numerical representation into the other’s space. In a company-run Braid v1.5.2 test on two Apple Silicon Macs, the sender transmitted signed semantic material and provenance; the receiver verified it, then generated its own native representation with a different embedding model. The result is a practical argument for receiver control, not yet a broad proof of cross-model semantic equivalence.

One handoff, two very different rules

The test put all-minilm:latest on the source machine and nomic-embed-text:latest on the receiving machine. The source produced a 384-dimensional representation; after verification, the receiver created a 768-dimensional Nomic representation. Both ran locally through Ollama on separate Apple Silicon Macs.

That distinction is the heart of Braid’s design. Intersignal says a transfer carries explicit semantic material, a source-model digest, provenance, and a cryptographically bound source representation. When sender and receiver use different models, the receiver checks those records and re-embeds the authenticated material locally rather than attempting a direct conversion between the models’ representations.

That makes Braid less a universal translator between model spaces than a signed handoff format. The sender supplies evidence about the context it is sending. The receiving system retains authority over whether to accept it and, when necessary, how to represent it using its own local model.

The fast path preserves coordinates; the bridge preserves evidence

The matching-model path is the simpler case. If both ends have the same embedding model and identical model digest, Braid can retain the authenticated 384-dimensional representation. It avoids re-embedding because the test treats the two nodes as sharing the same representation space.

The heterogeneous path makes a more conservative promise. It does not claim that a 384-dimensional vector can simply become a 768-dimensional one. Instead, it retains the signed source material and lets the receiving model construct the representation it uses locally. Intersignal says the cross-model transaction passed Braid’s Phase B finality step with bridge_semantic_invention set to false, separating sender-signed evidence from state derived by the receiver.

This is an important boundary, but also the central limit of the result. A cryptographic envelope can establish the origin and integrity of the transferred material; it does not by itself demonstrate that different models will interpret that material equivalently. The announced test covers two named embedding models, two Apple Silicon machines, and the company’s own testing conditions.

The QR reel tests transport, not semantic understanding

Intersignal also tested a more theatrical transport route: it divided a signed .brad object into 30 QR frames, displayed them on one Mac, captured them with a second Mac’s camera, and reconstructed the object byte-for-byte. The optical test is relevant for disconnected or air-gapped settings because it moves the same signed object without requiring a normal network connection.

But byte-for-byte reconstruction answers a different question from cross-model usefulness. It shows that the object survived a camera-mediated trip intact. The earlier verification and local re-embedding design is what governs how a heterogeneous receiver handles the contents after arrival.

What developers can inspect today

Braid v1.5.2 exposes the ingredients behind that receiver-controlled model rather than presenting a single opaque transfer operation. Its listed capabilities cover signatures, source-model binding, freshness and replay checks, local receiver finality, heterogeneous Semantic Capsules, the exact-space fast path, LAN transport, and visible QR transport.

• Signature and provenance records are intended to let the receiver verify what arrived and its stated source-model binding.
• Freshness and replay checks address whether an otherwise valid object should be treated as current or as a repeated transfer.
• Receiver-controlled Ollama integrations keep the embedding model, and any optional completion model, on the user’s own hardware.

The availability and maturity signals point in different directions. Intersignal says Braid v1.5.2 for macOS is available from its site, while the project labels Braid experimental developer research software. Its current Mac build is not distributed through Apple’s Developer ID notarization system, which means users may encounter a Gatekeeper warning.

A prototype claims a useful boundary, not interoperability at large

The strongest reading of the demonstration is not that Braid has solved portable AI memory across arbitrary systems. It is that Intersignal has demonstrated a specific workflow in which local machines with different embedding models can pass signed material, verify it, and let the recipient create a local representation instead of accepting the sender’s coordinates as native.