Build a research assistant inside the canvas.

Double-click the file or drag it into a browser window. The canvas boots showing the Memeroot title, a drop zone, and just the Tree tab.

Firefox should work too but XSLT support is most reliable in Chromium. Safari works for everything except XSL transforms.

Open the features/ folder. Select all the XMLs and drag them all onto the canvas's drop zone at once. They load in arbitrary order.

features/companion-streams.xml
features/companion-region-signing.xml
features/companion-transforms.xml
features/companion-section-author.xml
features/companion-invoker.xml
features/companion-compose.xml
features/companion-bundle-export.xml
features/companion-console.xml
features/companion-schemas.xml
features/feature-reader-code.xml

Verify: the header shows "10 features (incl. 3 inline)" — actually 13 total once the bootstrap inline ones are counted. New tabs appear: Streams, Identities, Sign, Transforms, Endpoints, Compose, Schemas, Console, Bundle.

Look at the row of tab buttons. You should see at minimum:

Tree · Streams · Identities · Sign · Transforms · Endpoints · Compose · Schemas · Console · Bundle

If some tabs are missing, the matching feature didn't load. Drag-and-drop it individually and watch the Console tab for parse errors.

Click the Identities tab in the canvas header. You'll see an empty list with a CREATE button at the top.

This imports the tour-author's public key (no private key — public-only). When the tour's chunk signatures verify in your canvas, they'll show the author name research-assistant-tour-author instead of just the fingerprint.

Optional but recommended. Skip it and signatures still verify; you just see the fingerprint.

stylesheets/xsl-tone-terse.xml
stylesheets/xsl-tone-public.xml
stylesheets/xsl-tone-annotated.xml

Each stylesheet registers as a transform. Once dropped, the Transforms tab shows three entries: tone-terse, tone-public, tone-annotated. These become reader options for any <chunk> region.

This is the stream declaration — it announces "there's a stream called research-assistant with this description" so the Streams tab can group its chunks.

Multi-select chunk-research-1.xml through chunk-research-8.xml and drag them all onto the drop zone. They can be dropped in any order; the streams feature uses the seq attribute on each chunk to order them.

The Streams tab shows one stream entry with eight chunks listed. Each chunk should show a green tick (✓) for hash chain integrity — each chunk's curr attribute equals SHA-256 of (previous chunk's curr + this chunk's canonical content).

If you see any red ✗ marks, the chunks weren't dropped cleanly. Refresh the canvas and start Phase 2 again.

The chunk renders in the public tone — a friendly tutorial voice. To flick tones: click the radio for xsl:tone-terse (fragmented shorthand) or xsl:tone-annotated (explanatory prose with reasoning).

Read chunk 1's welcome message in whichever tone fits you.

The Author tab is on the right of the tab bar. Fill in:

source id: question-1.xml
number:    Q1
title:     research question
register:  lay
paragraph: (your question)

Click SEND TO TREE. The Tree tab shows your new question-1.xml source.

Open Identities. Type your name. Click CREATE. An ECDSA-P256 keypair generates; the private key stays in your browser's localStorage; only the public part attaches to your signatures.

The signing feature produces a <signature> element and appends it to your question's root region. The Sign tab confirms with the fingerprint of the signature.

Switch back to Tree, click the new <signature> child of your question — the signature reader verifies live and shows ✓ verified.

Stub from chunk 4:

<endpoint id="answerer" provider="anthropic" model="claude-sonnet-4-20250514" max-tokens="600">
  <system>You are a helpful research assistant. Answer concisely with one or two paragraphs. Be precise.</system>
  <prompt>Replace this with your question — or wire it to the question source in the next chunk.</prompt>
</endpoint>

This creates a new source with the endpoint as root. The Endpoints tab now shows one card.

Watch the badge change to "invoking" for a few seconds. A <response> element appears under the endpoint with the AI's answer, hash, model, and token counts.

You just made an Anthropic API call. No key was needed — the canvas runs as a claude.ai artifact and auth is handled automatically.

Drill into question-1.xml in the Tree until you find your question paragraph. The Inspector shows its address, e.g. mr://question-1.xml/0.0.0. Copy that.

Paste the wired-endpoint stub (replace the prompt-ref value):

<endpoint id="answerer-wired" provider="anthropic" model="claude-sonnet-4-20250514" max-tokens="600" prompt-ref="mr://question-1.xml/0.0.0">
  <system>You are a helpful research assistant. Answer concisely with one or two paragraphs. Be precise.</system>
</endpoint>

INVOKE. The response now answers your specific question. Edit your question region and re-invoke — the answer updates because prompt-ref is live.

The Compose tab clones the source and inserts a <fork-source> lineage element. You now have two sources: the pristine original (re-invokeable later) and the fork (your draft to edit).

Find the response element's address in your fork. Paste the critic stub from chunk 7 with that address as prompt-ref:

<endpoint id="critic" provider="anthropic" model="claude-sonnet-4-20250514" max-tokens="500" prompt-ref="mr://(your-fork-id)/(path)">
  <system>You are a careful critical reader. Identify weaknesses, missing context, and false claims. Be specific. Suggest one concrete improvement.</system>
</endpoint>

INVOKE. The critic's response appears.

Open Sign, pick your fork source, SIGN. Now your final draft is cryptographically attributed to you. Anyone with the bundle can verify both your signature and the AI response's content hash.

The Bundle tab serialises every loaded feature, every content source, every signed region, into a single file. Download it. That file is now a complete record of your session: questions, AI responses, drafts, critiques, signatures, lineage — all together, verifiable end-to-end.

XML mode is for recipients who already have a canvas. HTML mode wraps the bundle in a fresh canvas so a recipient can open the file directly.

For each new research session, the loop is: Author question → PASTE endpoint with prompt-ref → INVOKE → Fork → Critique → Sign → Bundle. The pieces compose freely. You can have many endpoints in one session, multiple identities, fork-of-forks, signed responses inside signed drafts inside signed sections.

Pass your bundle to someone with a canvas and they have your whole session, verifiable.