diff --git a/backend/src/ingest/classifier.zig b/backend/src/ingest/classifier.zig
index b229b18..72e3963 100644
--- a/backend/src/ingest/classifier.zig
+++ b/backend/src/ingest/classifier.zig
@@ -38,7 +38,7 @@ const THRESHOLD: f64 = 0.50;
 // re-scores the whole corpus from a clean slate. v2 = precision fix (content-word
 // scaffold only, curation veto). v3 = threshold 0.55→0.50. v4 = model-pass gate
 // (heuristic flags → LLM confirms content is machine-generated before emit).
-const SCORING_VERSION: i64 = 8; // v8: judge model → Qwen3.6-35B (3/3 on judge-eval; 7B was 0/3)
+const SCORING_VERSION: i64 = 9; // v9: composed-vs-generated prompt + judge → gemma-4-12B (self-served)
 
 // review pipeline states. The heuristic is a cheap PRE-FILTER: it never emits
 // directly (titles can't tell a branded real blog from a registry mirror). It
@@ -57,12 +57,13 @@ const MAX_REVIEW_ATTEMPTS: i64 = 5; // give up after this many inconclusive revi
 // review provider: any OpenAI-compatible chat-completions endpoint. Defaults
 // to co/core; override all three via env (REVIEW_API_URL / REVIEW_MODEL /
 // REVIEW_API_KEY) to switch providers with a secrets change, no deploy of code.
-// picked by scripts/judge-eval (2026-07-01): 3/3 known-truth accounts, every
-// completed vote correct. Qwen2.5-7B (the previous judge) scored 0/3 —
-// confidently inverted on all three. Slower + reasoning-style, which the
-// worker tolerates (votes retry on inconclusive; latency doesn't back
-// anything up — candidates just stay PENDING longer).
-const DEFAULT_REVIEW_MODEL = "mlx-community/Qwen3.6-35B-A3B-4bit";
+// picked by scripts/judge-eval: gemma-4-12B went 19/19 conclusive votes
+// correct across two runs on the composed-vs-generated prompt (2026-07-02),
+// and it's the model our own provider machine serves — reliability and
+// economics we control. Qwen2.5-7B (the original judge) was 0/3, confidently
+// inverted. Reasoning-style + slow is fine: the worker is a durable queue,
+// candidates just stay PENDING longer.
+const DEFAULT_REVIEW_MODEL = "mlx-community/gemma-4-12B-it-8bit";
 const DEFAULT_REVIEW_URL = "https://console.cocore.dev/api/v1/chat/completions";
 
 const ReviewCfg = struct {
@@ -631,26 +632,34 @@ fn reviewVote(allocator: Allocator, cfg: ReviewCfg, io: Io, did: []const u8, vot
     const samples = try fetchVoteMaterial(allocator, did, vote);
     defer allocator.free(samples);
 
-    // prompt validated against the known cases on cocore Qwen2.5-7B: the framing
-    // (our task: long-form human writing vs automated feeds, "coherent ≠ human")
-    // + concrete examples are what make a small model separate a real blog from a
-    // transit/catalog feed. Without them it calls everything coherent "human".
+    // the line is COMPOSED vs GENERATED, not human vs machine: most banned
+    // content was human-written at the source (patents, recall notices,
+    // transcripts) and original AI writing is welcome. Keep this text in
+    // lockstep with scripts/judge-eval PROMPT_HEAD, and re-validate any change
+    // there against the known-truth set before shipping (19/19 conclusive
+    // votes correct on gemma-4-12B + Qwen3.6-35B, 2026-07-02).
     var prompt: std.Io.Writer.Allocating = .init(allocator);
     defer prompt.deinit();
     try prompt.writer.writeAll(
-        \\A search engine indexes original long-form writing by people (blogs, essays,
-        \\articles). It must EXCLUDE accounts that are automated feeds, catalogs, or
-        \\database exports — even when the text reads coherently.
+        \\A search engine indexes writing: documents COMPOSED by an author — a person or
+        \\an AI — who chose what to say. It must EXCLUDE accounts that GENERATE
+        \\documents from a data source: one document per database row, feed event,
+        \\catalog entry, or result, where a template plus the data determines the text.
         \\
-        \\The test: did a PERSON sit and write each item as original prose, OR is an
-        \\automated system emitting one record per database row / feed event / catalog
-        \\entry / log?
+        \\The test is NOT whether the text is fluent, and NOT whether a human once wrote
+        \\the underlying material — patents, recall notices, episode summaries, and
+        \\transcripts were all written by people, but republishing them one-per-record
+        \\is still generation. The test: does each document exist because its author had
+        \\something to say, OR because a record exists in some dataset?
         \\
-        \\machine=true examples: a transit-alert bot ("Red Line delayed near Roosevelt"),
-        \\a patent-database mirror, a TV-episode catalog (one entry per episode),
-        \\product-recall summaries, daily log entries, chart/stats dumps. Coherent != human.
-        \\machine=false examples: a personal blog (tech notes, essays, reviews), even with
-        \\branded/templated titles or a numbered series.
+        \\machine=true examples: a patent-database mirror, vehicle-recall summaries (one
+        \\per recall), a TV-episode catalog (one per episode), transit alerts (one per
+        \\service event), chart/stats dumps (one per chart-day), tournament results (one
+        \\per event), fundraiser announcements stamped from campaign records — even when
+        \\the underlying data belongs to the account itself.
+        \\machine=false examples: a personal blog or essay series (even branded,
+        \\numbered, templated-looking titles, or extremely prolific), a daily journal
+        \\(the date is a schedule, not a data source), original writing by an AI agent.
         \\
         \\Evidence from ONE account (facts, a title sample spanning its whole
         \\history, and excerpts from the middle of several documents):
diff --git a/docs/exclusions.md b/docs/exclusions.md
index 857de42..302275c 100644
--- a/docs/exclusions.md
+++ b/docs/exclusions.md
@@ -1,17 +1,22 @@
 # exclusions: what we keep out of the corpus, and why
 
-pub-search indexes **long-form writing by people**. some actors publish
-valid, well-signed AT Protocol records that are nonetheless not that —
+pub-search indexes **composed writing** — documents an author (a person or
+an AI) wrote because they had something to say. some actors publish valid,
+well-signed AT Protocol records that are nonetheless not that —
 machine-generated registry mirrors, scraper bridges, bulk archives. this
 file is the registry of every manual exclusion: who, when, why, and the
 evidence. if we're going to editorialize, we explain ourselves.
 
 ## what qualifies (the policy line)
 
-an author is excluded when **all** of these hold:
+the line is **composed vs. generated**, not human vs. machine: most banned
+content was human-written at the source (patents, recall notices,
+transcripts), and original writing by an AI is welcome. an author is
+excluded when **all** of these hold:
 
-1. **no human authorship** — the records are generated from an external
-   database/feed, not written
+1. **generated, not composed** — each document exists because a record
+   exists in some dataset (external or the account's own), with a template
+   determining the text; not because an author chose to write it
 2. **bulk scale** — the volume materially distorts the corpus or a topic's
    search results (rule of thumb: would be >1% of the corpus, or owns a
    topic's semantic results)
@@ -19,7 +24,8 @@ an author is excluded when **all** of these hold:
    removes noise, not voices
 
 an author is NOT excluded for: low quality, controversial content, high
-(human) volume, or self-promotion. the line is authorship, not taste.
+volume, self-promotion, or being an AI. the line is how the documents come
+to exist, not taste and not the author's species.
 
 ## how exclusions are enforced (the new architecture)
 
diff --git a/scripts/judge-eval b/scripts/judge-eval
index 8d31c7f..a789c25 100755
--- a/scripts/judge-eval
+++ b/scripts/judge-eval
@@ -31,6 +31,7 @@ KNOWN = {
     "did:plc:u5kk4h7tr4s3ntskbmhl5z7d": ("sksksketch.net", False),  # human illustrator
     "did:plc:5swhfspkrynnbidlkrkch3lh": ("prideraiser.org", True),  # fundraiser feed bot
     "did:plc:4z33k5fjzw2ew3u373pg7ku5": ("festivus (episode catalog)", True),
+    "did:plc:sttgf52vkk46f6yuknvqxvgh": ("coryd.dev (942-doc human)", False),  # volume trap
 }
 
 DEFAULT_MODELS = [
@@ -47,19 +48,25 @@ EXCERPT_LEN = 800
 
 COCORE_URL = "https://console.cocore.dev/api/v1/chat/completions"
 
-PROMPT_HEAD = """A search engine indexes original long-form writing by people (blogs, essays,
-articles). It must EXCLUDE accounts that are automated feeds, catalogs, or
-database exports — even when the text reads coherently.
-
-The test: did a PERSON sit and write each item as original prose, OR is an
-automated system emitting one record per database row / feed event / catalog
-entry / log?
-
-machine=true examples: a transit-alert bot ("Red Line delayed near Roosevelt"),
-a patent-database mirror, a TV-episode catalog (one entry per episode),
-product-recall summaries, daily log entries, chart/stats dumps. Coherent != human.
-machine=false examples: a personal blog (tech notes, essays, reviews), even with
-branded/templated titles or a numbered series.
+PROMPT_HEAD = """A search engine indexes writing: documents COMPOSED by an author — a person or
+an AI — who chose what to say. It must EXCLUDE accounts that GENERATE
+documents from a data source: one document per database row, feed event,
+catalog entry, or result, where a template plus the data determines the text.
+
+The test is NOT whether the text is fluent, and NOT whether a human once wrote
+the underlying material — patents, recall notices, episode summaries, and
+transcripts were all written by people, but republishing them one-per-record
+is still generation. The test: does each document exist because its author had
+something to say, OR because a record exists in some dataset?
+
+machine=true examples: a patent-database mirror, vehicle-recall summaries (one
+per recall), a TV-episode catalog (one per episode), transit alerts (one per
+service event), chart/stats dumps (one per chart-day), tournament results (one
+per event), fundraiser announcements stamped from campaign records — even when
+the underlying data belongs to the account itself.
+machine=false examples: a personal blog or essay series (even branded,
+numbered, templated-looking titles, or extremely prolific), a daily journal
+(the date is a schedule, not a data source), original writing by an AI agent.
 
 Evidence from ONE account (facts, a title sample spanning its whole
 history, and excerpts from the middle of several documents):
diff --git a/scripts/labeler-setup b/scripts/labeler-setup
index 05f853c..ae6f8e1 100755
--- a/scripts/labeler-setup
+++ b/scripts/labeler-setup
@@ -65,10 +65,12 @@ LABEL_DECLARATION = {
                         "lang": "en",
                         "name": "machine-generated",
                         "description": (
-                            "this account's content is produced by an automated "
-                            "system, not written by a person — e.g. registry/feed "
-                            "mirrors, transit-alert bots, catalogs, database exports. "
-                            "pub-search excludes these from its human-writing search."
+                            "this account's documents are generated from a data "
+                            "source — one per database row, feed event, or catalog "
+                            "entry — rather than composed by an author (a person or "
+                            "an AI). e.g. registry mirrors, transit-alert bots, "
+                            "catalogs, database exports. pub-search excludes these "
+                            "from search."
                         ),
                     }
                 ],
diff --git a/site/labels.html b/site/labels.html
index 959c0eb..b9e8d87 100644
--- a/site/labels.html
+++ b/site/labels.html
@@ -62,8 +62,9 @@
   <div class="container">
     <h1><a href="/" class="title">pub search</a> <span class="dim">/ labels</span></h1>
     <p class="lead">
-      pub-search indexes original writing. Accounts that emit records rather than prose —
-      feeds, catalogs, database exports — get a signed
+      pub-search indexes writing that an author — a person or an AI — composed.
+      Accounts that generate documents from a dataset (one per patent, recall, chart,
+      episode, alert) get a signed
       <a href="https://atproto.com/specs/label" target="_blank" rel="noopener"><code>machine-generated</code></a>
       label and are excluded from search. The decision is automated: a pattern filter nominates,
       a model on <a href="https://cocore.dev" target="_blank" rel="noopener">co/core</a> decides.
@@ -78,7 +79,7 @@
         <div class="stage-name">pattern filter</div>
         <div class="stage-value">score ≥ 0.50</div>
       </div>
-      <div class="stage" title="a language model reads titles and excerpts sampled across the account's whole history and votes: authored writing, or mechanical output? two of three votes decide, and each vote sees a different sample. inconclusive votes retry — no verdict is forced.">
+      <div class="stage" title="a language model reads titles and excerpts sampled across the account's whole history and votes: composed by an author, or generated one-per-record from a dataset? two of three votes decide, and each vote sees a different sample. inconclusive votes retry — no verdict is forced.">
         <div class="stage-name">model review</div>
         <div class="stage-value" id="st-review">2-of-3 votes</div>
       </div>