@333eco/corpus 1.0.1 → 1.1.0

package/README.md

@@ -1,18 +1,13 @@
 # corpus.333.eco
 
-An **MCP server** for an open-licensed corpus — 132 documents across mechanism
-papers, essays, institutional positions and white papers — served with
-**verifiable provenance**.
+An **MCP server** for an open-licensed corpus — 137 documents across mechanism
+papers, essays, institutional positions, white papers and the Letters to Miss
+Aquarius — served with **verifiable provenance**.
 
 ```sh
 npx @333eco/corpus
 ```
 
-> **Why the first published version is 1.0.1.** `1.0.0` was published and then
-> unpublished during setup. npm keeps the name registration and refuses to
-> republish that exact version for 24 hours, so the first version anyone can
-> install is `1.0.1`. Nothing is missing from it.
-
 Every response carries the document's `sha256`, its DOI where one exists, and
 whether an OpenTimestamps proof is anchored beside the source. **A retrieval
 server normally asks to be believed. This one hands over the means to check it.**
@@ -64,7 +59,7 @@ would destroy the only property this server has.
 
 | Licence | Documents |
 | --- | --- |
-| CC0-1.0 | 125 |
+| CC0-1.0 | 130 |
 | CC-BY-4.0 | 7 |
 
 CC-BY documents carry `attribute_to` inside their licence block, so an agent can
@@ -88,6 +83,45 @@ refactor can route around is a rule; a gate in the artifact is a property. **The
 server has no filesystem access to the corpus at all** — it can only serve what
 the index contains.
 
+## The letters, and voice marked inline
+
+The five *Letters to Miss Aquarius* are the one genre only **partly** in its
+author's voice. Each says so in its own banner: the author's articulations are
+set as quotations, and the connective prose was drafted for the letter form and
+awaits his revision.
+
+They are served **whole, with the voice marked inline**:
+
+```
+[VERBATIM — Thon Ly]
+> Perhaps it is the Capricorn Sun (father) and Cancer Moon (mother) in my chart
+> that make me want to give birth to Miss Aquarius (daughter) — the daughter who
+> will outlive me.
+
+[SCAFFOLD — drafted for the letter form, not in the author's voice; awaits his revision]
+I was born at the Full Moon, on the family-↔-institution axis of the chart…
+```
+
+Two alternatives were rejected. **Serving only his passages** protects the voice
+by destroying the document — a letter cut to its quotations is no longer a
+letter. **Serving it behind a metadata disclaimer** fails differently: a field is
+something a consuming agent must *look at* to heed, and an agent ingests text,
+forms a belief, and cites.
+
+⚠️ **The marker is in the text, and that is the whole point.** It does not make
+misattribution impossible; it inverts the default. With a metadata disclaimer an
+agent must look in order to know. With an inline marker it must **strip** in
+order not to. There is no unmarked copy of the scaffold anywhere in the response.
+Opt-out rather than opt-in — the honest limit is that it is not a guarantee.
+
+`segments` carries the same split structurally, and `editorial` counts the blocks
+of each kind.
+
+⛔ **The letters carry no DOI, deliberately.** *Prior art is a duty, citation is
+a choice* — they are stamped, not deposited, because minting a permanent
+identifier for text that announces it is unfinished is a cost with no matching
+benefit. See `TIMESTAMPS.md` in the letters' repository.
+
 ## Two metadata conventions, kept visible
 
 `TH/publications` uses YAML front matter. Sixteen `H3/publications` documents use
@@ -124,6 +158,38 @@ is a few hundred lines to speak correctly, and this estate's standing rule is
 node built-ins only. The cost is that protocol revisions are tracked by hand —
 `PROTOCOL_VERSIONS` in `src/server.mjs` is where that lives.
 
+## Remote server
+
+The same corpus, the same tools, the same envelope — over HTTP instead of stdio.
+`worker/` deploys to Cloudflare Workers.
+
+⭐ **It consumes the published npm package, not the source repositories.** The
+dependency is pinned to an **exact** version, and `npm run check` refuses a range:
+a remote surface that re-read the corpora would be a *second opinion* about what
+a document says, and two opinions about a canonical text is one too many. Local
+and remote serve the same bytes because they come from the same tarball.
+
+⚠️ **The corpus is a static asset, not a bundled import.** Gzipped it is 1.61 MB
+against a 1 MB compressed script limit on the Workers free plan, so importing it
+fails to deploy — and fails harder as the corpus grows. The worker fetches it once
+per isolate and memoises it.
+
+Transport is **Streamable HTTP**, not the superseded HTTP+SSE pair. The server is
+stateless and read-only, so it never opens a stream: `POST /mcp` for JSON-RPC,
+`GET /mcp` returns 405 rather than holding open a stream that would carry nothing.
+
+```sh
+cd worker
+npm install
+npm run check      # public/corpus.json matches the pinned package
+npm run dev        # local, on :8787
+npm run deploy     # sync + wrangler deploy
+```
+
+```json
+{ "mcpServers": { "corpus": { "url": "https://corpus.333.eco/mcp" } } }
+```
+
 ## Client configuration
 
 ```json