Let one AI node pass useful understanding to another.

The second node gained usable state - and the September field test showed what the receiver chose.

In the Kestrel run, a local model on a remote Linux node was asked a question about an invented operating protocol it had never seen. Before Braid, it confidently made the wrong decision. A Mac then sent the missing facts and rules through a public TLS 1.3 relay. After the receiving node accepted and committed that state, the same model answered new questions correctly in fresh inference calls.

Object digest: 5f81e8a4578f4b766644e95c08a26f8e22e82b6ff03bcad87c71b2efd49e08cb. This proves the core A-to-B contextual-hydration result from committed semantic state. It does not claim transfer of model weights, hidden activations, or KV cache; broader release-candidate negative testing remains ongoing.

In ordinary terms, a shared text pad would simply hold all five lines. Braid used the receiving computer's own semantic index to find the four lines relevant to the question, leave out the unrelated line, notice that 73 and 75 conflict, retain where each claim came from, and construct a new local prompt. The field worked even though the small answer model still needed stronger guardrails.

Behavioral result: receiver-local 384D cosine + MMR retrieval passed; numeric conflict detection passed; source preservation passed; whole-transcript replay was false. Completion quality remains model-dependent. The stricter Qwen 7B rerun still selected 73 despite receiving both 73 and 75, so Braid will next pass reconciliation constraints into generation more deterministically.

More than one LLM should not mean starting over every time.

A second model may live on another computer, specialize in another task, or operate in another physical location. Without a shared state layer, useful context often dies where it was created or has to be copied manually. Braid gives that context a portable, inspectable form, then lets the receiver retrieve the parts relevant to a new question while leaving acceptance and local use under receiver control.

One LLM can work alone. The moment you have two, continuity becomes infrastructure.

Technical users can reproduce the core mechanism now. The one-click field experience is still being packaged.

Kestrel was a controlled benchmark, not a hard-coded answer. The scenario was invented so the receiver could not know it in advance; Braid carried the missing state, and the receiver's own local model performed the later reasoning. Braid 1.6.0rc2 remains the public technical release candidate. Braid 1.7.0a2 has now passed its first physical Mac-to-Linux Semantic Field UAT, but it remains a CLI-level engineering alpha rather than a finished consumer workflow.

Join the active build - with today's capabilities stated plainly.

Braid is moving quickly, but not every planned capability or one-click workflow exists today. Membership places you inside the working feedback loop: a practical degree of direct human support, every Braid software update released during your active membership term, and access to Braid Managed Relays once they are sufficiently hardened for member use.

Selected meaning moves. Control does not.

Braid is not a shared chat window and it does not ask unrelated models to interpret an opaque vector as a universal language. The sender chooses what should become durable. Braid binds explicit semantic material to a compact machine representation and its provenance. The receiver verifies the object, applies local policy, and only then makes the state available to its own model.

macOS may block the first launch of the v1.5.2 research build.

The current downloadable DMG is not yet distributed through Apple's Developer ID notarization pipeline. Gatekeeper can therefore show that Braid was blocked. You do not need to disable macOS security.

What Braid 1.6 is - and is not

Braid 1.6.0rc2 is experimental research software for moving authenticated situational meaning into durable, queryable receiver-owned state. Native v1.5.2 packages remain available while native 1.6 installers complete physical validation.

The plain-language result sits on top of an inspectable technical history.

The homepage now leads with what Braid does. The materials below preserve how the result was tested, what changed during the Kestrel run, and where the present claim stops. Earlier native installers remain available because they document the LAN, trust, optical, and cross-OS foundation that preceded queryable state.

Useful intelligence should not die where it was created.

Models, agents, robots, workstations, and edge systems will increasingly operate as independent nodes. They need a way to pass useful situational state without collapsing into one cloud account, one proprietary memory store, or one model-specific session.

Braid is an early semantic state plane for that future: explicit meaning, compact representation, visible provenance, receiver-owned acceptance, and multiple possible paths between machines. Evidence may travel. Authority remains local.