Two rigs. One private network.
Keep the computers on the same private LAN for this first test. This guide does not require a relay account or public port forwarding. Do not expose the browser UI or the receiver to the public internet.
On each rig, have standard CPython 3.10 through 3.13 installed. Python 3.14 and free-threaded builds are outside this release’s dependency target. Python must include venv, ensurepip, SSL, and SQLite support. Official Python downloads.
Install and start local Ollama. Then provision the embedding model on each rig:
ollama pull all-minilmGet Ollama · Official quickstart. Model weights are downloaded separately; they are not inside the Archangel ZIP.
A local chat/generation model is needed on the receiver for Ask field. The demonstrated run used mistral:latest. An existing suitable local model can be selected instead; a different model is another configuration to test. To reproduce the model choice, provision Mistral on the receiver:
ollama pull mistralThis can download substantial model weights. Check storage and memory before doing so. Mistral model page.
Extract the whole folder.
Download the same complete ZIP to both rigs. Extract into its own folder; do not run from inside the ZIP, overlay it onto an earlier release, or replace only the Mac command file. You should see start_braid.py beside the platform launchers.
Extract. Open Start Braid.command.
The Mac fix is already in this ZIP. It skips an unsupported default Python, finds a supported installed runtime, and prints its selection.
sh "Start Braid.command"Terminal alternative: run the command above inside the extracted folder. Python must already be installed. No global Python setting is changed. Mac security guidance.
Open the extracted folder in Terminal.
Right-click an empty space in the folder, choose Open in Terminal, and run:
sh "Start Braid.sh"Run as your normal user, without sudo. The Linux launcher uses your default python3. For an unsupported default, explicitly use an installed supported version, such as python3.12 start_braid.py.
Choose a supported Python explicitly.
Open the extracted folder in Terminal. With standard Python 3.12 installed, use:
py -3.12 start_braid.pyUse -3.13 instead when that is your installed supported version. Start Braid.cmd is also included, but uses the default runtime. This is not an Authenticode-signed installer.
Let installation finish. Look for Archangel is ready:, then open the printed local address in that computer’s browser. The port is chosen locally and can differ between machines.
Leave both service terminals open. The browser address beginning 127.0.0.1 is the local user interface, not the address to enter on the other rig.
macOS says it cannot verify the developer
This ZIP is not Developer ID signed or notarized. Verify its source and checksum. For a trusted downloaded item, follow Apple’s per-item Privacy & Security / Open Anyway guidance where available. Do not disable Gatekeeper globally or recursively strip quarantine attributes. An executable-bit error is distinct from a security warning. The Terminal launcher command above does not require the script itself to have an executable bit. Apple’s current guidance.
The card connects. The fingerprint verifies.
- On the Mac, open Add rig → This rig. Give the machine a name and choose its actual private LAN address, not
127.0.0.1. - Use the installed all-minilm embedding model. Press Prepare and wait for the route to finish calibrating. Then choose Save card.
- Do the same on Linux. Transfer each public
.braidrigfile to the other computer through a channel you trust. - On Linux, open Add rig → Other rig. Choose the Mac’s card. Compare the entire fingerprint with the value on the Mac’s own screen, check the approval box, and click Approve rig.
- Import and approve Linux’s card on the Mac in the same way. Do not skip the reverse approval.
Keep receiving paused for the next configuration step. Preparing or importing a card does not, by itself, open the receiver.
Choose the receiver’s local model.
On Linux, open Settings & tools → Refresh local model list. In Local completion model, choose or enter the exact name of your installed chat model, such as mistral:latest, then Save settings.
Do not choose all-minilm as the completion model. It is the embedding model. A chat/generation model writes the field answer. Exact-text sharing does not require a sender-side summarizer; local summarization is optional later.
Now start receiving on both rigs: use Add rig → Connect → Start receiving, or click Receiving paused in the top bar. Confirm it changes to Receiving, then Check connections.
The direct receiver defaults to port 8746. Permit it only on the intended private network when the operating system asks. A successful connection check is not an accepted or indexed message.
One fact. One deliberate send.
On the Mac, click Share summary. Select only the Linux laptop. For Subject, enter Archangel first handoff.
Under What should the other rig know?, paste:
Project Cobalt's launch color is amber.Do not paste that sentence on Linux. Its copy should arrive through Braid.
Click Use exact text. Confirm the sentence in Review is unchanged and Linux is the only selected recipient. Then click Send summary once and wait.
On the Mac, open Sent → Archangel first handoff. Look for Accepted & indexed. Expand Receipt & provenance to inspect the details.
On Linux, open Inbox and the new handoff from the Mac. Check the exact sentence and the receipt stages Received → Accepted → Indexed. The inbox entry may be titled with the sentence: Subject is a sender-local label, not part of the signed semantic text.
| Outcome | What to do now |
|---|---|
| Accepted & indexed | Proceed to receiver-local retrieval below. |
| Accepted · not indexed | Stop and inspect the receipt. Acceptance is not confirmation of indexing; do not resend. |
| Receipt not confirmed | Stop. The receiver may already have committed it. Inspect both machines; do not use New handoff as a retry workaround. |
| Not accepted / Not sent | Read the precise rejection or preparation error. Correct it before any deliberate retry. |
Find the fact. Then ask the field.
After confirmation of indexing, click Shared field at the bottom of the Linux sidebar. Do not use the inbox’s “Find a summary” search or an unrelated chat app.
In What would you like to know?, enter:
What is Project Cobalt's launch color?Click Find relevant state first. Confirm that the retrieved result contains the Cobalt sentence. Then click Ask field with the same question.
The answer should identify amber. Expand Exact result & evidence to inspect the result, retrieval, model identifiers, and guard details. The answer’s wording can vary; an answer alone is not a substitute for checking its source evidence.
This checklist is only a local browser aid. It sends no data, is not saved, and does not certify a run.
Stop here for the first pass. Record screenshots and the exact errors, if any. Reverse-direction, restart, conflict, Windows, and WAN tests are separate next steps; do not fold them into a claim about this first handoff.
Keep the good state. Find the snag.
0 indexed items / ERR_FIELD_NO_RELEVANT_ATOMS
The current field has no relevant indexed material for that question. First check whether Send summary was actually pressed. Then inspect the sender’s Sent receipt and the receiver’s Inbox. Do not infer a transport or model failure from the later query error alone. Do not repeatedly send, reinstall, re-pair, or click Index retained state before understanding the receipt.
Python version is outside 3.10–3.13
On a Mac, use the included corrected Start Braid.command, not a command file copied from the original a5 ZIP. It searches standard installed locations. On Linux, choose an installed supported executable explicitly, such as python3.12 start_braid.py. On Windows, use py -3.12 start_braid.py or another installed supported minor. Those commands do not install Python. Install a supported runtime first if none exists.
The Mac launcher cannot find a supported complete runtime
Install standard CPython 3.12 or 3.13 from a trusted source, or explicitly choose an existing interpreter using BRAID_PYTHON. A read-only check is sh "Start Braid.command" --runtime-info. A missing venv or ensurepip component may require your operating system’s matching Python environment package. Do not change system aliases or launch Braid as root.
No completion model / only an embedding model is listed
Start local Ollama and check ollama list. Provision a local chat/generation model, refresh the model list in Settings & tools, and select its exact name. Pause receiving before saving settings, then resume. The retrieval test can be inspected separately; all-minilm is not a replacement for a completion model.
Receipt not confirmed
An absent acknowledgment does not prove the message was absent at the receiver. Do not resend or choose New handoff to get a green status. Inspect the receiving inbox and the sender’s detailed receipt. Successful, in-flight, and uncertain recipients are locked for that handoff; that is intentional.
A connection worked, then the network address changed
Use Settings & tools → Network address → Update address for an actually assigned private address. Export the refreshed card and approve it on the peer. This preserves identity and history. Do not rotate keys or enable a public/wildcard bind. Overlay address support does not create a VPN or prove a WAN connection.
The Linux command cannot find Start Braid.sh
The terminal is probably outside the extracted folder. Open the folder containing Start Braid.sh and start_braid.py, right-click empty space, choose Open in Terminal, then run the command. Do not substitute a guessed Downloads path or create a new profile to work around a folder-location issue.
Only Apple touch-icon requests show 404
Those requests concern optional browser icons, not receipt or semantic-field processing. Distinguish them from failures on the actual page assets, API, receiver, or transfer. Use the handoff receipt to diagnose a send, not a missing touch-icon log entry.
For help, send the operating system, Python version, exact action, and error or receipt screenshot to hello@intersignal.ai. Do not send private keys, enrollment tokens, sensitive context, or an entire profile.
Quit cleanly. Upgrade deliberately.
A working a5 + Mac Fix 1 installation does not need reinstalling for this website launch. The published application ZIP is the same byte-for-byte bundle, not a new engine build.
For an older install, quit it first. Make a complete external backup of its profile, and extract the complete new ZIP into a separate directory. The launcher reuses the standard a4 Messenger profile when present. Its pre-Archangel snapshot covers desktop/trust data, not the whole semantic inbox or model library.
After moving from a4, save fresh Archangel rig cards and approve them in both directions. The direct Messenger path requires mutual TLS; there is no silent legacy-card downgrade. Do not run a4 and a5 against one live profile. Do not delete or copy live databases to force a connection.
Closing a browser tab does not stop the service. Use Settings & tools → Quit Archangel or Control-C in the launcher terminal. A leftover lock filename alone is not a reason to delete it.
The fix is already inside.
Complete ZIP: Braid-Archangel-1.7.0a5-Mac-Launcher-Fix1.zip
Application version: 1.7.0a5
Packaging revision: mac-launcher-fix1
Expected SHA-256 of the ZIP:
b6bf35bd10edcb687e89e21f9674db0ca58a9dc7adb37b667c9d139213ab4227Compute the checksum before extracting. On macOS:
shasum -a 256 Braid-Archangel-1.7.0a5-Mac-Launcher-Fix1.zipOn Linux:
sha256sum Braid-Archangel-1.7.0a5-Mac-Launcher-Fix1.zipOn Windows PowerShell:
Get-FileHash .\Braid-Archangel-1.7.0a5-Mac-Launcher-Fix1.zip -Algorithm SHA256Compare the complete value. A checksum checks consistency with the published bytes; it is not a code-signing certificate or publisher signature.
After extraction, the Mac discovery launcher can check the internal manifest without installing or opening a profile:
sh "Start Braid.command" --checkWith a supported interpreter on Linux/Windows, use python3.12 start_braid.py --check or py -3.12 start_braid.py --check, selecting the version you actually installed.
Evidence, without conflating the runs
The retained a5 QA report records 264 tests plus six subtests on Linux, using deterministic embedding/completion fixtures. The Mac fix adds 16 launcher tests with simulated installation layouts. The September 5 founder run adds physical MacBook Air startup and Mac-to-Linux indexed field output with real local model configuration. It does not close the rest of the platform and network matrix.
The packaged QA notes are retained historical build evidence and predate the founder run. The new physical-run note supplements them rather than rewriting the earlier record.
QR reels, without a membership.
The release retains the optical tools in its Advanced workspace. Beam QR reel displays a signed object as a sequence of QR chunks; the receiving camera workflow collects the chunks and reassembles the object for receiver-local validation. Seeing a reel or collecting its chunks is not, by itself, acceptance or semantic indexing.
This is a separate advanced workflow, not the Messenger composer's Send summary button. It does not require an Intersignal relay or a paid membership, and it does not bypass approved identities, compatible routes, or the receiver's own policies.
Camera support matters. The browser can use a native QR decoder where available, or the local optional OpenCV decoder. The core installer does not bundle that optional camera dependency. Camera permission, a supported decoder, and appropriate camera/display conditions are required; no universal phone/browser or air-gap guarantee is implied.
The September 5 Mac-to-Linux demonstration on this site used LAN, not a newly completed optical-transfer test. Keep that working LAN configuration intact while testing an optional transport separately.