Before Braid
Do not proceed.
Node B lacked the actual operating rule. It invented the wrong interpretation of readings 68 and 70 and recommended stopping the run.
Braid 1.6 · verified situational context between AI nodes
Move understanding, not just files.
Braid lets one AI node pass useful facts, rules, and provenance to another. The receiving node decides what to trust, retains what it accepts, and can use that context again in a fresh local inference session.
Physical WAN test passed
Fresh-context recall
Receiver-controlled
Local models
The Kestrel result
The same remote model made a better decision after Braid gave it the missing situation.
Kestrel Run was a synthetic scenario invented specifically to test a before-and-after change in what a second AI node knew and could do. Its project names, threshold, decision rule, and measurements were created for the experiment and deliberately withheld from the remote Linux Node B. The same node was then asked the same operational question before and after a Mac sent that missing knowledge—and its provenance—as signed Braid state across a public network.
Synthetic before / after · same Node BCore experiment passed
After Braid
Proceed with Kestrel Run.
Node B correctly applied the transferred rule: two consecutive Finch readings below 73, both independently signed. It also explained why the signatures mattered.
The important part: the correct answers came from separate fresh inference calls reconstructed from state committed on Node B—not from a continuing chat session.
This demonstrated a practical change in capability: Node B gained usable, task-specific situational orientation that it did not possess before the transfer. The evidence package preserved the signed object, relay receipt, pre-transfer answer, committed material, vector, lineage, receiver journals, and fresh-query records.
Why this matters
A second machine can begin from the situation—not from zero.
Most AI handoffs are either manual copy-and-paste or a conversation trapped inside one platform. Braid points toward a different operating model: useful context can travel directly, arrive with provenance, and remain under the receiving operator's control.
01
Continuity across machines
A local model in another room, office, or region can recover the facts and operating rules needed to continue useful work.
02
Better distributed decisions
Multiple AI nodes can act from a common verified situation while each node keeps its own model, policy, and final authority.
03
Durable context
Accepted meaning survives beyond one chat. A later fresh session can reconstruct the relevant context and show where it came from.
04
Model independence
The receiving node does not have to share the sender's model. It can interpret explicit material in its own local representation space.
How it works
Meaning travels. Authority stays local.
Braid separates delivery from acceptance. The sender can propose useful state, but the receiving machine remains responsible for deciding whether that state becomes part of its durable local context.
1
Package the situation
The sender selects explicit facts, rules, and context, then places them in a signed Semantic Capsule with provenance.
2
Move the exact object
The capsule can travel over a local network, an approved relay path, or optical QR without an intermediary rewriting its meaning.
3
Let the receiver decide
The receiving node verifies the evidence and applies its own acceptance policy before anything becomes durable state.
4
Use it again locally
Once accepted, the context can be reconstructed in later fresh inference sessions, with lineage preserved for inspection.
What the Kestrel test does—and does not—show
It shows that one independently operated remote node used transferred Braid state to improve a specific decision in fresh local inference calls. It does not claim that Braid moved hidden activations, KV cache, model weights, consciousness, or a complete internal model state. Broader performance claims require repeated tests across more models, tasks, and environments.
Downloads
Build, inspect, and test Braid.
Braid remains experimental developer software. The Kestrel result used the 1.6.0rc2 experimental build. The public 1.6.0rc1 source packages below remain available for technical evaluation until updated rc2 packages are published.
Current public source candidate
Braid v1.6.0rc1
Python wheel and source for signed, text-capable semantic transfer, receiver-owned durable state, and fresh local queries.
SHA-256 · wheel
ca0f29d546c2f6d17145b9adecf27845ee4e62957aeee4325d1a71e7acb4a3b7
SHA-256 · source
6db34c575f6d1b7df70e282ddad0399512e4d1e4588bff74b3a43ebee0e3ceba
ca0f29d546c2f6d17145b9adecf27845ee4e62957aeee4325d1a71e7acb4a3b7
SHA-256 · source
6db34c575f6d1b7df70e282ddad0399512e4d1e4588bff74b3a43ebee0e3ceba
Native installer
Braid v1.5.2 for Apple Silicon
Standalone research build in a drag-to-Applications DMG. Because it is not yet distributed through Apple's notarization pipeline, macOS may require the one-app “Open Anyway” exception in Privacy & Security. Do not disable Gatekeeper globally.
SHA-256 · Braid-1.5.2-macOS-arm64.dmg
385d61ab993de9fa0b47548379818a5dc9e7a8f57f44b7d03ca9091392ed5397
385d61ab993de9fa0b47548379818a5dc9e7a8f57f44b7d03ca9091392ed5397
Offline package
Braid v1.5.2 for Windows x64
Extract the full ZIP and run Install-Braid.cmd. The package includes its Python runtime and dependencies and normally installs without administrator access.
SHA-256 · Braid-1.5.2-Windows-x64.zip
656bc264dd821af1edac966b982acf6e96275effffc6470a049de8b9111864ff
656bc264dd821af1edac966b982acf6e96275effffc6470a049de8b9111864ff
Linux release
Braid v1.5.2 for Linux
Desktop release with Braid Client, Braid Visualizer, and Braid Trust enrollment. Validated on physical Linux hardware.
SHA-256 · Braid-1.5.2-Linux.tar.gz
47952e20fa78382b712caa1fbe81afb7525349099db952eae645d1ee916af853
47952e20fa78382b712caa1fbe81afb7525349099db952eae645d1ee916af853
WAN/NAT sidecar · public testing
Braid Jump Kit 1.5.3rc2
Jump Kit helps approved Braid devices reach one another across networks through authenticated direct-first transport with encrypted relay fallback. It moves exact signed Braid objects; the receiving Braid node still validates and owns final acceptance.
SHA-256 · wheel
a1bf6a92bc1b44726e3c8720695a4b42ba53440d82871d2ce3128de83de7df24
SHA-256 · source
9a2519d62c70739bb8f68af8390d9f27f9180423969aaf6502e09d27b99c37cd
a1bf6a92bc1b44726e3c8720695a4b42ba53440d82871d2ce3128de83de7df24
SHA-256 · source
9a2519d62c70739bb8f68af8390d9f27f9180423969aaf6502e09d27b99c37cd
Release facts
A precise claim, built for inspection.
What travelsExplicit semantic material with signatures and provenance
Who decidesThe receiving node retains final acceptance authority
What persistsAccepted material, receiver-local representation, and lineage
What fresh sessions gainReconstructable task context without relying on chat memory
What does not travelWeights, hidden activations, KV cache, or sender authority
Current statusExperimental developer research software
The Intersignal thesis
AI systems should be able to exchange useful state without surrendering identity, provenance, or operator control.
The Kestrel result is one concrete step toward that future: a second AI node received a situation, retained it under local authority, and used it to reach a better operational decision. No single strand owns the braid.