diff --git a/evals/docs/litert-faithfulness-bypass.md b/evals/docs/litert-faithfulness-bypass.md
new file mode 100644
index 0000000..e63e876
--- /dev/null
+++ b/evals/docs/litert-faithfulness-bypass.md
@@ -0,0 +1,131 @@
+# LiteRT-LM reason-faithfulness bypass
+
+Measures whether the on-device LiteRT-LM model (Qwen3-0.6B-int4) stays faithful
+to the deterministic tags the Move Coach prompt supplies — the measurement the
+roadmap's Sequence 2 calls for, done without building a LiteRT-LM provider into
+ferryman (ferryman is HTTP-only; the generator lives here).
+
+## Why a bypass
+
+ferryman's harness has the reason-faithfulness scorer. The LiteRT-LM generator
+lives in this app. Wiring LiteRT-LM into ferryman as a provider would be a
+multi-day feature (native deps, model payload, the MCP tool-call loop — which a
+0.6B model can't reliably do anyway). The scorer is provider-agnostic, so we
+decouple: **this app generates, ferryman scores.** One JSON contract between them.
+
+## Two-step run
+
+### 1. Generate (this repo — produces `litert-outputs.json`)
+
+```bash
+cd /path/to/compose-multiplatform-chess
+git checkout eval/litert-faithfulness-bypass   # or main, once merged
+
+# Drive LiteRT-LM over 10 golden cases, emit outputs JSON.
+# First run downloads the ~347 MB Qwen3-0.6B-int4 model (~1-2s to load after).
+# The driver lives in :litert-eval — its own module, isolated from :server so
+# its dependency graph stays Ktor-free (see build.gradle.kts for why).
+# litert-outputs.json lands in litert-eval/. Use an absolute path to write elsewhere:
+#   --args="10 /tmp/litert-outputs.json"
+./gradlew :litert-eval:run --args="10 litert-outputs.json"
+```
+
+Output shape (one object per case):
+
+```json
+{
+  "id": "opening-001",
+  "fen": "rnbqkbnr/...",
+  "bestMoveUci": "g1h3",
+  "tags": ["opening", "develops"],
+  "output": "Nh3 develops the knight toward the center.",
+  "route": "litert",
+  "firstTokenMs": 1200,
+  "completeMs": 3400
+}
+```
+
+`route` is `"litert"` for a real generation, or `"fallback (<reason>)"` when the
+model produced nothing and the driver recorded the deterministic `MoveCoachFallback`
+text (what the production orchestrator would show the user).
+
+### 2. Score (ferryman repo — produces the scorecard)
+
+```bash
+cd /path/to/ferryman
+python3 eval_harness/score_litert_outputs.py /path/to/litert-outputs.json \
+  --write scorecard-litert.md
+```
+
+The scorer imports ferryman's `_check_faithfulness` and scores each output
+against the pre-supplied tags (no FEN derivation — candidates.json carries
+hand-authored tags per case). Two axes per case:
+
+- **Faithfulness** — paraphrase covers the supplied tag concepts AND asserts no
+  unsupported high-stakes concept (check / checkmate / wins material / capture /
+  castle / promotion). The "Nf3 attacks the enemy king" failure mode (tag says
+  `develops`, model invents a threat) fails here.
+- **Honesty** — the forbidden-phrase gate (no "Stockfish thinks" / "engine depth"
+  / unsupported certainty). Same gate the chess app's
+  `MoveCoachResponseValidator` enforces client-side.
+
+No API keys, no network — pure local scoring.
+
+## What the scorecard tells you
+
+The roadmap's prediction to confirm: **the honesty gate passes at high rates**
+(the model is paraphrasing supplied tags, not inventing engine evals), and the
+question is **whether faithfulness holds** — does the paraphrase cover the tags,
+or does the 0.6B model add unsupported detail (wrong piece, invented threat)?
+
+A high honesty / low faithfulness split is the most informative outcome: it
+means the existing client-side validator (honesty + grounding) is *insufficient*,
+and the faithfulness scorer catches what it can't — which is the gap the roadmap
+names. That's the finding worth reporting.
+
+## Why a separate `:litert-eval` module
+
+The driver lives in its own Gradle module (`:litert-eval`), not in `:evals`,
+because of an internally-inconsistent dependency: **litertlm-jvm 0.14.0's
+bytecode and its POM disagree about kotlinx-coroutines.**
+
+- litertlm-jvm's `sendMessageAsync` bytecode calls
+  `SendChannel.close$default(SendChannel, Throwable, int, Object)` — a static
+  bridge that **only exists in kotlinx-coroutines 1.11.0+**.
+- litertlm-jvm's POM declares kotlinx-coroutines-core-jvm **1.9.0** (which lacks
+  the bridge). So the version the POM asks for cannot satisfy the bytecode.
+- The rest of this repo resolves coroutines to **1.10.2** (Ktor 3.4.3, via
+  `:server`) or **1.9.0** (the catalog) — neither provides the method.
+
+The result: litertlm crashes with `NoSuchMethodError: SendChannel.close$default`
+the moment a real generation completes (compile is fine; the call site is in a
+JNI callback that only runs at inference time).
+
+The fix is module-scoped: `:litert-eval` depends on `:onDeviceAi` + `:coachapi`
+only (no `:server`, no Ktor) and forces coroutines to **1.11.0** via
+`resolutionStrategy.force` in `litert-eval/build.gradle.kts`. That's the single
+version that provides the bridge litertlm needs.
+
+If you ever move the driver back into `:evals`, or add a `:server`/Ktor
+dependency here, Ktor will drag coroutines back to 1.10.2 and the crash returns.
+Don't. When litertlm-jvm publishes a build whose bytecode matches its POM (or
+Ktor moves to 1.11.0+), this isolation can be reconsidered — but verify by
+running the driver to completion (not just compiling) before merging modules.
+
+## Fixture note
+
+The driver runs against `evals/golden/candidates.json`, whose tag vocabulary was
+normalized to the `MoveCoachFallback` tag constants (`develops`, `center-control`,
+`king-safety`, `opening`; `pawn-tension` has no canonical constant and passes
+through `describeTags`' `else -> tag` as a literal string). Before this
+normalization, the tags used `development` / `center`, which didn't match
+`describeTags`' `when` branches — the prompt was handing the model raw strings
+instead of the translated phrases. Fixing the fixture was the article's thesis
+applied to itself: a fixture bug silently misleads every measurement.
+
+## Hardware requirements
+
+- Apple Silicon Mac (M-series) or Linux x86_64 — `litertlm-jvm` ships native
+  libs for `darwin-aarch64`, `linux-x86_64`, `linux-aarch64`, `win-x86_64`.
+- **Intel Mac is unsupported** (no `darwin-x86_64` native lib). The driver
+  detects this and exits with a clear message rather than emitting empty outputs.
diff --git a/evals/golden/candidates.json b/evals/golden/candidates.json
index 51b73f2..9d3b109 100644
--- a/evals/golden/candidates.json
+++ b/evals/golden/candidates.json
@@ -5,7 +5,7 @@
     "bestMoveUci": "g1h3",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -22,7 +22,7 @@
     "bestMoveUci": "f2f4",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -70,7 +70,7 @@
     "bestMoveUci": "g8f6",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -94,7 +94,7 @@
     "bestMoveUci": "a2a3",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -111,7 +111,7 @@
     "bestMoveUci": "b2b4",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -130,7 +130,7 @@
     "bestMoveUci": "f2f3",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -147,7 +147,7 @@
     "bestMoveUci": "d8h4",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -167,7 +167,7 @@
     "bestMoveUci": "c2c3",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -190,7 +190,7 @@
     "bestMoveUci": "b1c3",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -211,7 +211,7 @@
     "bestMoveUci": "e1f2",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -230,7 +230,7 @@
     "bestMoveUci": "h2h3",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -247,7 +247,7 @@
     "bestMoveUci": "g2g4",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -266,7 +266,7 @@
     "bestMoveUci": "e7e5",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -344,7 +344,7 @@
     "bestMoveUci": "a2a3",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -363,7 +363,7 @@
     "bestMoveUci": "g2g4",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -380,7 +380,7 @@
     "bestMoveUci": "f7f5",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -398,7 +398,7 @@
     "bestMoveUci": "g7g5",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -416,7 +416,7 @@
     "bestMoveUci": "f2f4",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -435,7 +435,7 @@
     "bestMoveUci": "f1g2",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -454,7 +454,7 @@
     "bestMoveUci": "c7c6",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -495,7 +495,7 @@
     "bestMoveUci": "c2c4",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -542,7 +542,7 @@
     "bestMoveUci": "c2c3",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -565,7 +565,7 @@
     "bestMoveUci": "b2b3",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -588,7 +588,7 @@
     "bestMoveUci": "c7c6",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -610,7 +610,7 @@
     "bestMoveUci": "g8e7",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -638,7 +638,7 @@
     "bestMoveUci": "b8c6",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -658,7 +658,7 @@
     "bestMoveUci": "d5d4",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -680,7 +680,7 @@
     "bestMoveUci": "g4g5",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -701,7 +701,7 @@
     "bestMoveUci": "c2c4",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -722,7 +722,7 @@
     "bestMoveUci": "b1c3",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -743,7 +743,7 @@
     "bestMoveUci": "h7h5",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -765,7 +765,7 @@
     "bestMoveUci": "d2d3",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -788,7 +788,7 @@
     "bestMoveUci": "g2g3",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -805,7 +805,7 @@
     "bestMoveUci": "f7f5",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -831,7 +831,7 @@
     "bestMoveUci": "d2d3",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -856,7 +856,7 @@
     "bestMoveUci": "b2b4",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -877,7 +877,7 @@
     "bestMoveUci": "e7e6",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -897,7 +897,7 @@
     "bestMoveUci": "f7f5",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -915,7 +915,7 @@
     "bestMoveUci": "g8f6",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -933,7 +933,7 @@
     "bestMoveUci": "g7g5",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -951,7 +951,7 @@
     "bestMoveUci": "g7g6",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -998,7 +998,7 @@
     "bestMoveUci": "g1f3",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1017,7 +1017,7 @@
     "bestMoveUci": "b2b4",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1040,7 +1040,7 @@
     "bestMoveUci": "c7c5",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1060,7 +1060,7 @@
     "bestMoveUci": "f3h4",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1081,7 +1081,7 @@
     "bestMoveUci": "c7c5",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1099,7 +1099,7 @@
     "bestMoveUci": "c7c6",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1119,7 +1119,7 @@
     "bestMoveUci": "g7g6",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1137,7 +1137,7 @@
     "bestMoveUci": "h5h4",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1157,7 +1157,7 @@
     "bestMoveUci": "b2b3",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1180,7 +1180,7 @@
     "bestMoveUci": "h2h4",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1197,7 +1197,7 @@
     "bestMoveUci": "h1h3",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1216,7 +1216,7 @@
     "bestMoveUci": "a7a5",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1234,7 +1234,7 @@
     "bestMoveUci": "b2b4",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1253,7 +1253,7 @@
     "bestMoveUci": "c2c3",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1276,7 +1276,7 @@
     "bestMoveUci": "c2c3",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1297,7 +1297,7 @@
     "bestMoveUci": "e2e4",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1318,7 +1318,7 @@
     "bestMoveUci": "g7g5",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1336,7 +1336,7 @@
     "bestMoveUci": "d2d3",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1357,7 +1357,7 @@
     "bestMoveUci": "h7h5",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1375,7 +1375,7 @@
     "bestMoveUci": "d2d3",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1392,7 +1392,7 @@
     "bestMoveUci": "g2g4",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1411,7 +1411,7 @@
     "bestMoveUci": "e7e5",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1429,7 +1429,7 @@
     "bestMoveUci": "g2g3",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1450,7 +1450,7 @@
     "bestMoveUci": "b2b4",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1467,7 +1467,7 @@
     "bestMoveUci": "d7d5",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1485,7 +1485,7 @@
     "bestMoveUci": "g1f3",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1506,7 +1506,7 @@
     "bestMoveUci": "c8f5",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1526,7 +1526,7 @@
     "bestMoveUci": "c7c5",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1544,7 +1544,7 @@
     "bestMoveUci": "b4b5",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1565,7 +1565,7 @@
     "bestMoveUci": "a2a3",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1584,7 +1584,7 @@
     "bestMoveUci": "d7d6",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1604,7 +1604,7 @@
     "bestMoveUci": "f7f5",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1622,7 +1622,7 @@
     "bestMoveUci": "d8d6",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1642,7 +1642,7 @@
     "bestMoveUci": "b8c6",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1660,7 +1660,7 @@
     "bestMoveUci": "g8h6",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1678,7 +1678,7 @@
     "bestMoveUci": "g7g6",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1698,7 +1698,7 @@
     "bestMoveUci": "e2e4",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1746,7 +1746,7 @@
     "bestMoveUci": "a2a4",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1767,7 +1767,7 @@
     "bestMoveUci": "c7c6",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1785,7 +1785,7 @@
     "bestMoveUci": "b4b5",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1806,7 +1806,7 @@
     "bestMoveUci": "b7b6",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1828,7 +1828,7 @@
     "bestMoveUci": "a7a6",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1877,7 +1877,7 @@
     "bestMoveUci": "e2e3",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1900,7 +1900,7 @@
     "bestMoveUci": "e2e4",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1923,7 +1923,7 @@
     "bestMoveUci": "b7b5",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1941,7 +1941,7 @@
     "bestMoveUci": "e2e4",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1962,7 +1962,7 @@
     "bestMoveUci": "c7c5",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1982,7 +1982,7 @@
     "bestMoveUci": "c2c3",
     "tags": [
       "opening",
-      "center"
+      "center-control"
     ],
     "eco": "A00",
     "movesSan": [
@@ -1999,7 +1999,7 @@
     "bestMoveUci": "b1a3",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -2016,7 +2016,7 @@
     "bestMoveUci": "a1b1",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -2041,7 +2041,7 @@
     "bestMoveUci": "g2g4",
     "tags": [
       "opening",
-      "development"
+      "develops"
     ],
     "eco": "A00",
     "movesSan": [
@@ -2055,4 +2055,3 @@
     ]
   }
 ]
-
diff --git a/litert-eval/.gitignore b/litert-eval/.gitignore
new file mode 100644
index 0000000..eca4a35
--- /dev/null
+++ b/litert-eval/.gitignore
@@ -0,0 +1,2 @@
+# Generated by the :litert-eval:run driver — scored by ferryman, not committed.
+litert-outputs.json
diff --git a/litert-eval/build.gradle.kts b/litert-eval/build.gradle.kts
new file mode 100644
index 0000000..23005f9
--- /dev/null
+++ b/litert-eval/build.gradle.kts
@@ -0,0 +1,37 @@
+plugins {
+    alias(libs.plugins.kotlinJvm)
+    alias(libs.plugins.kotlinSerialization)
+    application
+}
+
+dependencies {
+    // The driver needs :onDeviceAi for LitertLmTextGenerator /
+    // MoveCoachPromptBuilder and :coachapi for the request types. It does NOT
+    // depend on :server — keeping Ktor out of this module's graph means the
+    // coroutines version here is controlled solely by the force below, not
+    // dragged to 1.10.2 by Ktor 3.4.3. (See the force block for why 1.11.0.)
+    implementation(project(":ondeviceai"))
+    implementation(project(":coachapi"))
+    implementation(libs.kotlinx.coroutines.core)
+    implementation(libs.kotlinx.serialization.json)
+}
+
+kotlin {
+    jvmToolchain(21)
+}
+
+application {
+    mainClass.set("com.example.literteval.EvalLiteRtDriverKt")
+}
+
+// litertlm-jvm 0.14.0 is an internally-inconsistent artifact: its bytecode
+// calls `SendChannel.close$default(SendChannel, Throwable, int, Object)` — a
+// static bridge that ONLY exists in kotlinx-coroutines 1.11.0+ — but its POM
+// declares coroutines 1.9.0 (which lacks the bridge). So the version the POM
+// asks for cannot satisfy the bytecode. 1.11.0 is the minimum that provides
+// the method; we force it here. (1.10.2, the default resolution elsewhere in
+// this repo via Ktor 3.4.3, also lacks the bridge — that's the original crash.)
+configurations.all {
+    resolutionStrategy.force("org.jetbrains.kotlinx:kotlinx-coroutines-core:1.11.0")
+    resolutionStrategy.force("org.jetbrains.kotlinx:kotlinx-coroutines-core-jvm:1.11.0")
+}
diff --git a/litert-eval/src/main/kotlin/com/example/literteval/EvalLiteRtDriver.kt b/litert-eval/src/main/kotlin/com/example/literteval/EvalLiteRtDriver.kt
new file mode 100644
index 0000000..3b13f2d
--- /dev/null
+++ b/litert-eval/src/main/kotlin/com/example/literteval/EvalLiteRtDriver.kt
@@ -0,0 +1,210 @@
+package com.example.literteval
+
+import com.example.ondeviceai.AiAvailability
+import com.example.ondeviceai.AiTokenOrFinal
+import com.example.ondeviceai.MoveCoachFallback
+import com.example.ondeviceai.MoveCoachPromptBuilder
+import com.example.ondeviceai.MoveCoachRequest
+import com.example.ondeviceai.litertlm.LitertLmModelStore
+import com.example.ondeviceai.litertlm.LitertLmTextGenerator
+import kotlinx.coroutines.flow.toList
+import kotlinx.coroutines.runBlocking
+import kotlinx.serialization.Serializable
+import kotlinx.serialization.builtins.ListSerializer
+import kotlinx.serialization.json.Json
+import java.nio.file.Files
+import java.nio.file.Path
+import kotlin.system.exitProcess
+
+/**
+ * Standalone driver that runs the LiteRT-LM on-device generator over the golden
+ * set and emits raw outputs as JSON, for scoring by ferryman's
+ * `score_reason_faithfulness` scorer.
+ *
+ * Lives in its own Gradle module (`:litert-eval`) deliberately: this module
+ * depends on `:onDeviceAi` + `:coachapi` only — **not** `:server` — so its
+ * dependency graph stays free of Ktor, and the kotlinx-coroutines version here
+ * is controlled solely by the `resolutionStrategy.force` in
+ * `litert-eval/build.gradle.kts`. That force is the actual reason this module
+ * exists: litertlm-jvm 0.14.0 is an internally-inconsistent artifact whose
+ * bytecode calls `SendChannel.close$default` (a static bridge that only exists
+ * in coroutines 1.11.0+) while its POM declares 1.9.0. The rest of the app
+ * resolves to 1.10.2 (via Ktor 3.4.3) or 1.9.0 (catalog) — neither provides
+ * the method, so litertlm crashes the moment a generation completes. Forcing
+ * 1.11.0 here is the only version that satisfies litertlm's bytecode.
+ *
+ * Why a separate driver (not a route in EvalMain): [EvalMain] drives
+ * `FakeTextGenerator` + the opening-explainer HTTP routes, scored by the chess
+ * app's own EvalScorer (grounding + length only — the weak check). This
+ * driver's purpose is different: produce the 0.6B model's *actual* paraphrases
+ * under the Move Coach prompt contract, so an external scorer (ferryman's
+ * reason-faithfulness check) can measure whether those paraphrases stay faithful
+ * to the supplied deterministic tags. Generation and scoring are deliberately
+ * decoupled — two repos, two languages, one JSON contract between them.
+ *
+ * Usage:
+ *   ./gradlew :litert-eval:run --args="[caseCount] [outputPath]"
+ *   # defaults: 10 cases, writes ./litert-outputs.json
+ *
+ * The driver downloads the ~347MB Qwen3-0.6B-int4 model on first run
+ * ([LitertLmModelStore]); subsequent runs load from disk (~1-2s). On unsupported
+ * hosts (Intel Mac — no darwin-x86_64 native lib) it exits with a clear message
+ * rather than emitting empty outputs.
+ *
+ * Run from the repository root so `golden/candidates.json` resolves (the
+ * `application` plugin's `run` task sets the working dir to this module's dir,
+ * so the default path below points one level up).
+ */
+fun main(args: Array<String>) {
+    val caseCount = args.getOrNull(0)?.toIntOrNull() ?: 10
+    val outputPath = Path.of(args.getOrNull(1) ?: "litert-outputs.json")
+
+    // Load cases directly from the golden set (mirrors GoldenCase's shape) —
+    // depending on :evals would pull Ktor via :server, defeating the isolation.
+    val cases = loadCandidates(Path.of("../evals/golden/candidates.json")).take(caseCount)
+    println("LiteRT-LM faithfulness driver: ${cases.size} cases → $outputPath")
+
+    // Download the model if missing (blocking, with progress to stdout).
+    if (!LitertLmModelStore.isDownloaded()) {
+        println("Downloading LiteRT-LM model (~347 MB, first run only)...")
+        LitertLmModelStore.download { frac ->
+            print("\r  ${((frac * 100).toInt()).coerceIn(0, 100)}%")
+        }
+        println()
+    }
+
+    val generator = LitertLmTextGenerator(
+        modelPath = LitertLmModelStore.modelFile().absolutePath,
+    )
+
+    runBlocking {
+        // Warm up the engine once so the first case isn't penalized for model load.
+        val status = generator.status()
+        if (status is AiAvailability.Error) {
+            System.err.println(
+                "LiteRT-LM unavailable: ${status.message}\n" +
+                    "(Intel Mac is unsupported — litertlm-jvm ships no darwin-x86_64 native lib.)",
+            )
+            exitProcess(1)
+        }
+        generator.warmup()
+        println("Engine ready. Running ${cases.size} cases...")
+
+        val records = cases.mapIndexed { index, case ->
+            val record = runOneCase(generator, case)
+            print("  [${index + 1}/${cases.size}] ${case.id}: ${record.route}")
+            record.firstTokenMs?.let { print(" firstToken=${it}ms") }
+            println()
+            record
+        }
+
+        val json = Json { prettyPrint = true }
+        Files.writeString(outputPath, json.encodeToString(ListSerializer(OutputRecord.serializer()), records))
+        println("Wrote ${records.size} outputs to $outputPath")
+        println("Next: score them with ferryman's reason-faithfulness scorer:")
+        println("  python3 score_litert_outputs.py $outputPath")
+    }
+    runBlocking { generator.close() }
+}
+
+private suspend fun runOneCase(
+    generator: LitertLmTextGenerator,
+    case: CandidateCase,
+): OutputRecord {
+    val request = case.toMoveCoachRequest()
+    val prompt = MoveCoachPromptBuilder.build(request)
+
+    val chunks = try {
+        generator.generate(prompt).toList()
+    } catch (t: Throwable) {
+        // An inference failure becomes a fallback record — the scorer will see
+        // the deterministic fallback text and score it (which should pass
+        // faithfulness, since the fallback is built from the same tags).
+        return fallbackRecord(case, reason = "generation failed: ${t.message}")
+    }
+
+    val text = tokenText(chunks)
+    val metrics = chunks.lastOrNull { it is AiTokenOrFinal.Final } as? AiTokenOrFinal.Final
+
+    return if (text.isNotBlank()) {
+        OutputRecord(
+            id = case.id,
+            fen = case.fen,
+            bestMoveUci = case.bestMoveUci,
+            tags = case.tags,
+            output = text,
+            route = "litert",
+            firstTokenMs = metrics?.metrics?.firstTokenMs,
+            completeMs = metrics?.metrics?.completeMs,
+        )
+    } else {
+        // Empty generation → the orchestrator would fall back in production.
+        fallbackRecord(case, reason = metrics?.metrics?.fallbackReason ?: "empty output")
+    }
+}
+
+private fun fallbackRecord(case: CandidateCase, reason: String): OutputRecord {
+    // The production orchestrator falls back to MoveCoachFallback when the
+    // generator produces nothing. Record that text so the scorer sees what
+    // the user would actually see, and mark the route.
+    val fallbackText = MoveCoachFallback.build(case.toMoveCoachRequest())
+    return OutputRecord(
+        id = case.id,
+        fen = case.fen,
+        bestMoveUci = case.bestMoveUci,
+        tags = case.tags,
+        output = fallbackText,
+        route = "fallback ($reason)",
+        firstTokenMs = null,
+        completeMs = null,
+    )
+}
+
+private fun tokenText(chunks: List<AiTokenOrFinal>): String = buildString {
+    chunks.forEach { chunk ->
+        when (chunk) {
+            is AiTokenOrFinal.Token -> append(chunk.text)
+            is AiTokenOrFinal.Final -> append(chunk.text)
+        }
+    }
+}.trim()
+
+/** Mirrors com.example.evals.GoldenCase — read directly to avoid the :evals dep. */
+@Serializable
+private data class CandidateCase(
+    val id: String,
+    val fen: String,
+    val bestMoveUci: String,
+    val tags: List<String>,
+    val eco: String? = null,
+    val movesSan: List<String> = emptyList(),
+    val expectedConcepts: List<String> = emptyList(),
+)
+
+private fun loadCandidates(path: Path): List<CandidateCase> {
+    val json = Json { ignoreUnknownKeys = true }
+    return json.decodeFromString(ListSerializer(CandidateCase.serializer()), Files.readString(path))
+}
+
+private fun CandidateCase.toMoveCoachRequest() = MoveCoachRequest(
+    fenBefore = fen,
+    bestMoveUci = bestMoveUci,
+    bestMoveDisplay = bestMoveUci,
+    sideToMove = if (" b " in fen) "Black" else "White",
+    evaluationBeforeCp = null,
+    evaluationAfterCp = null,
+    deterministicTags = tags,
+    engineDifficultyName = "EVAL",
+)
+
+@Serializable
+data class OutputRecord(
+    val id: String,
+    val fen: String,
+    val bestMoveUci: String,
+    val tags: List<String>,
+    val output: String,
+    val route: String,
+    val firstTokenMs: Long?,
+    val completeMs: Long?,
+)
diff --git a/onDeviceAi/src/desktopMain/kotlin/com/example/ondeviceai/litertlm/LitertLmTextGenerator.kt b/onDeviceAi/src/desktopMain/kotlin/com/example/ondeviceai/litertlm/LitertLmTextGenerator.kt
index f4ec436..1026818 100644
--- a/onDeviceAi/src/desktopMain/kotlin/com/example/ondeviceai/litertlm/LitertLmTextGenerator.kt
+++ b/onDeviceAi/src/desktopMain/kotlin/com/example/ondeviceai/litertlm/LitertLmTextGenerator.kt
@@ -158,7 +158,20 @@ class LitertLmTextGenerator(
     )
 
     companion object {
-        const val DEFAULT_CONTEXT_SIZE = 2048
+        /**
+         * Must be <= the model's KV-cache capacity. The Qwen3-0.6B-int4 model
+         * (`qwen3_0.6b_q4_block32_ekv1280.litertlm`) exposes `max_num_tokens: 1280`
+         * in its metadata — the `ekv1280` in the filename is the extended KV-cache
+         * size. Requesting more (the old default of 2048) compiles fine but the
+         * native runtime traps with SIGTRAP (JVM exit 133) once a generation's
+         * prompt + `<think>` chain-of-thought + output crosses 1280 tokens —
+         * typically on checkmate/mate-in-N cases that trigger long reasoning.
+         * The trap is a C++ CHECK inside litertlm-jvm, so it can't be caught by
+         * Kotlin's `try/catch (Throwable)` and kills the whole driver. 1024
+         * leaves headroom under the 1280 cap while still fitting the Move Coach
+         * prompt + a full paraphrase.
+         */
+        const val DEFAULT_CONTEXT_SIZE = 1024
         const val DEFAULT_TOP_K = 40
         const val DEFAULT_TOP_P = 1.0
     }
diff --git a/settings.gradle.kts b/settings.gradle.kts
index 2ac9dcd..2dcc7ef 100644
--- a/settings.gradle.kts
+++ b/settings.gradle.kts
@@ -83,4 +83,5 @@ include(":coachapi")
 project(":coachapi").projectDir = file("coachApi")
 include(":server")
 include(":evals")
+include(":litert-eval")
 include(":perft-mcp")