diff --git a/EVAL_HARNESS_README.md b/EVAL_HARNESS_README.md
index bfb2f82..7fb10c1 100644
--- a/EVAL_HARNESS_README.md
+++ b/EVAL_HARNESS_README.md
@@ -2,6 +2,8 @@
 
 A single Rust binary combining the Janus project harness with eval-harness behavior-regression testing. No external dependencies — uses the user's current model context for LLM evaluation.
 
+> ⚗️ **Experimental / opt-in.** The eval-harness (`harness-cli eval …`, all 14 commands) is **not** part of the default build. It is gated behind the `eval` cargo feature and compiles only with `cargo build --release --features eval` (or `cargo install --features eval`). The default binary ships the 9 quality gates without this eval layer. Everything documented below assumes a build with `--features eval`.
+
 ---
 
 ## Bắt đầu nhanh (cho người mới)
diff --git a/README.md b/README.md
index f2bf440..e56932d 100644
--- a/README.md
+++ b/README.md
@@ -45,7 +45,7 @@ all backed by one SQLite database.
 | | |
 |---|---|
 | 🛡️ **9 quality gates** | Pre‑flight → AC extraction → recall → matrix → evidence → review → push permission → Jira tier → reconciliation, scaled by risk lane. |
-| 🎯 **14 eval commands** | `run · baseline · diff · status · promote · trend · accept · apply · ab · rebaseline · analyze · suggest · goal · quality`. |
+| 🎯 **14 eval commands** _(experimental, opt-in)_ | `run · baseline · diff · status · promote · trend · accept · apply · ab · rebaseline · analyze · suggest · goal · quality`. Not in the default binary — build with `--features eval`. |
 | 🔬 **9 check kinds** | Shell, JSON‑path, file, transcript grep (±), LLM judge, prompt/response quality, context relevance. |
 | 🧭 **5 attribution classes** | When an eval regresses, Janus tells you *why*: skill, fixture, model, cross‑skill, or unknown drift. |
 | ⚡ **Single Rust binary** | 5.4 MB, **zero runtime dependencies** — no bash, no jq, no python. Cold start ≈ 10 ms. |
@@ -80,6 +80,9 @@ bash ~/janus/scripts/install-harness.sh --yes
 # Build from source
 git clone https://github.com/hoainho/janus.git
 cd janus && cargo build --release
+
+# Include the experimental eval harness (opt-in — off by default)
+cargo build --release --features eval
 ```
 
 </details>
@@ -190,6 +193,11 @@ flowchart LR
 
 ## 🔬 Eval harness
 
+> ⚗️ **Experimental / opt-in.** The eval harness is not compiled into the default
+> binary. Build it with `cargo build --release --features eval` (or `cargo install
+> --features eval`); without that flag the `harness-cli eval …` subcommand is absent.
+> The 9 quality gates above ship in every build — only this eval layer is gated.
+
 The eval harness answers one question on every push: **does this skill still do
 what it did at baseline?** Pick a mode, run the checks, and on a regression Janus
 attributes the likely cause instead of just turning red.
diff --git a/crates/harness-cli/Cargo.toml b/crates/harness-cli/Cargo.toml
index fb1eb17..32a9b7b 100644
--- a/crates/harness-cli/Cargo.toml
+++ b/crates/harness-cli/Cargo.toml
@@ -5,6 +5,11 @@ edition.workspace = true
 license.workspace = true
 repository.workspace = true
 
+[features]
+# Experimental, opt-in eval-harness (behavior-regression testing). Not part of
+# the default build; enable with `--features eval`.
+eval = []
+
 [dependencies]
 clap = { version = "4.6.1", features = ["derive"] }
 rusqlite = { version = "0.39.0", features = ["bundled"] }
diff --git a/crates/harness-cli/src/interface.rs b/crates/harness-cli/src/interface.rs
index 5e35aee..5adc572 100644
--- a/crates/harness-cli/src/interface.rs
+++ b/crates/harness-cli/src/interface.rs
@@ -1,4 +1,5 @@
 use std::env;
+#[cfg(feature = "eval")]
 use std::io::Write;
 use std::path::PathBuf;
 use std::str::FromStr;
@@ -52,6 +53,7 @@ enum Command {
     GateLog(GateLogArgs),
     Query(QueryArgs),
     Export(ExportArgs),
+    #[cfg(feature = "eval")]
     /// Eval harness commands for behavior-regression testing.
     Eval(EvalArgs),
 }
@@ -491,12 +493,14 @@ struct ExportStoryArgs {
     id: String,
 }
 
+#[cfg(feature = "eval")]
 #[derive(Args, Debug)]
 struct EvalArgs {
     #[command(subcommand)]
     action: EvalAction,
 }
 
+#[cfg(feature = "eval")]
 #[derive(Subcommand, Debug)]
 enum EvalAction {
     /// Run eval cases for a skill.
@@ -529,6 +533,7 @@ enum EvalAction {
     Rebaseline(EvalRebaselineArgs),
 }
 
+#[cfg(feature = "eval")]
 #[derive(Args, Debug)]
 struct EvalQualityArgs {
     #[arg(long)]
@@ -537,6 +542,7 @@ struct EvalQualityArgs {
     case: Option<String>,
 }
 
+#[cfg(feature = "eval")]
 #[derive(Args, Debug)]
 struct EvalAnalyzeArgs {
     #[arg(long)]
@@ -545,12 +551,14 @@ struct EvalAnalyzeArgs {
     days: i64,
 }
 
+#[cfg(feature = "eval")]
 #[derive(Args, Debug)]
 struct EvalSuggestArgs {
     #[arg(long)]
     skill: String,
 }
 
+#[cfg(feature = "eval")]
 #[derive(Args, Debug)]
 struct EvalGoalArgs {
     #[arg(long)]
@@ -563,6 +571,7 @@ struct EvalGoalArgs {
     target: String,
 }
 
+#[cfg(feature = "eval")]
 #[derive(Args, Debug)]
 struct EvalRunArgs {
     #[arg(long)]
@@ -583,30 +592,35 @@ struct EvalRunArgs {
     eval_type: String,
 }
 
+#[cfg(feature = "eval")]
 #[derive(Args, Debug)]
 struct EvalBaselineArgs {
     #[arg(long)]
     skill: String,
 }
 
+#[cfg(feature = "eval")]
 #[derive(Args, Debug)]
 struct EvalDiffArgs {
     #[arg(long)]
     run_id: String,
 }
 
+#[cfg(feature = "eval")]
 #[derive(Args, Debug)]
 struct EvalStatusArgs {
     #[arg(long)]
     skill: Option<String>,
 }
 
+#[cfg(feature = "eval")]
 #[derive(Args, Debug)]
 struct EvalPromoteArgs {
     #[arg(long)]
     skill: String,
 }
 
+#[cfg(feature = "eval")]
 #[derive(Args, Debug)]
 struct EvalTrendArgs {
     #[arg(long)]
@@ -615,6 +629,7 @@ struct EvalTrendArgs {
     last: usize,
 }
 
+#[cfg(feature = "eval")]
 #[derive(Args, Debug)]
 struct EvalAcceptArgs {
     #[arg(long)]
@@ -623,6 +638,7 @@ struct EvalAcceptArgs {
     case: String,
 }
 
+#[cfg(feature = "eval")]
 #[derive(Args, Debug)]
 struct EvalApplyArgs {
     #[arg(long)]
@@ -631,6 +647,7 @@ struct EvalApplyArgs {
     run_id: Option<String>,
 }
 
+#[cfg(feature = "eval")]
 #[derive(Args, Debug)]
 struct EvalAbArgs {
     #[arg(long)]
@@ -639,6 +656,7 @@ struct EvalAbArgs {
     skill_b: String,
 }
 
+#[cfg(feature = "eval")]
 #[derive(Args, Debug)]
 struct EvalRebaselineArgs {
     #[arg(long)]
@@ -942,6 +960,7 @@ pub fn run(cli: Cli) -> Result<(), InterfaceError> {
                 print_export_story_md(&service.query_export_story(&args.id)?)
             }
         },
+        #[cfg(feature = "eval")]
         Command::Eval(args) => match args.action {
             EvalAction::Run(args) => {
                 run_eval(&args)?;
@@ -991,6 +1010,7 @@ pub fn run(cli: Cli) -> Result<(), InterfaceError> {
     Ok(())
 }
 
+#[cfg(feature = "eval")]
 fn run_eval(args: &EvalRunArgs) -> Result<(), InterfaceError> {
     let skills_root = resolve_skills_root();
     let skill_dir = skills_root.join(&args.skill);
@@ -1393,6 +1413,7 @@ fn run_eval(args: &EvalRunArgs) -> Result<(), InterfaceError> {
     Ok(())
 }
 
+#[cfg(feature = "eval")]
 fn get_state_dir() -> std::path::PathBuf {
     if let Ok(dir) = std::env::var("EVAL_STATE_DIR") {
         return std::path::PathBuf::from(dir);
@@ -1405,6 +1426,7 @@ fn get_state_dir() -> std::path::PathBuf {
     std::path::PathBuf::from(".opencode/eval-harness")
 }
 
+#[cfg(feature = "eval")]
 fn get_home_dir() -> Option<std::path::PathBuf> {
     if let Ok(home) = std::env::var("HOME") {
         return Some(std::path::PathBuf::from(home));
@@ -1418,6 +1440,7 @@ fn get_home_dir() -> Option<std::path::PathBuf> {
     None
 }
 
+#[cfg(feature = "eval")]
 fn run_baseline(args: &EvalBaselineArgs) -> Result<(), InterfaceError> {
     let skills_root = resolve_skills_root();
     let skill_dir = skills_root.join(&args.skill);
@@ -1519,6 +1542,7 @@ fn run_baseline(args: &EvalBaselineArgs) -> Result<(), InterfaceError> {
     Ok(())
 }
 
+#[cfg(feature = "eval")]
 fn run_diff(_args: &EvalDiffArgs) -> Result<(), InterfaceError> {
     let state_dir = get_state_dir();
     let db_path = state_dir.join("harness.db");
@@ -1604,6 +1628,7 @@ fn run_diff(_args: &EvalDiffArgs) -> Result<(), InterfaceError> {
     Ok(())
 }
 
+#[cfg(feature = "eval")]
 fn run_status(args: &EvalStatusArgs) -> Result<(), InterfaceError> {
     let state_dir = get_state_dir();
     let db_path = state_dir.join("harness.db");
@@ -1658,6 +1683,7 @@ fn run_status(args: &EvalStatusArgs) -> Result<(), InterfaceError> {
     Ok(())
 }
 
+#[cfg(feature = "eval")]
 fn run_promote(args: &EvalPromoteArgs) -> Result<(), InterfaceError> {
     let state_dir = get_state_dir();
     let db_path = state_dir.join("harness.db");
@@ -1713,6 +1739,7 @@ fn run_promote(args: &EvalPromoteArgs) -> Result<(), InterfaceError> {
     Ok(())
 }
 
+#[cfg(feature = "eval")]
 fn run_analyze(args: &EvalAnalyzeArgs) -> Result<(), InterfaceError> {
     let state_dir = get_state_dir();
     let db_path = state_dir.join("harness.db");
@@ -1805,6 +1832,7 @@ fn run_analyze(args: &EvalAnalyzeArgs) -> Result<(), InterfaceError> {
     Ok(())
 }
 
+#[cfg(feature = "eval")]
 fn run_suggest(args: &EvalSuggestArgs) -> Result<(), InterfaceError> {
     let state_dir = get_state_dir();
     let db_path = state_dir.join("harness.db");
@@ -1869,6 +1897,7 @@ fn run_suggest(args: &EvalSuggestArgs) -> Result<(), InterfaceError> {
     Ok(())
 }
 
+#[cfg(feature = "eval")]
 fn resolve_skills_root() -> std::path::PathBuf {
     if let Ok(root) = std::env::var("OPENCODE_SKILLS_ROOT") {
         return std::path::PathBuf::from(root);
@@ -2677,6 +2706,7 @@ fn print_row(values: &[String], widths: &[usize]) {
     println!();
 }
 
+#[cfg(feature = "eval")]
 fn run_goal(args: &EvalGoalArgs) -> Result<(), InterfaceError> {
     let state_dir = get_state_dir();
     let db_path = state_dir.join("harness.db");
@@ -2715,6 +2745,7 @@ fn run_goal(args: &EvalGoalArgs) -> Result<(), InterfaceError> {
     Ok(())
 }
 
+#[cfg(feature = "eval")]
 fn run_quality(args: &EvalQualityArgs) -> Result<(), InterfaceError> {
     let state_dir = get_state_dir();
     let db_path = state_dir.join("harness.db");
@@ -2778,6 +2809,7 @@ fn run_quality(args: &EvalQualityArgs) -> Result<(), InterfaceError> {
     Ok(())
 }
 
+#[cfg(feature = "eval")]
 fn run_trend(args: &EvalTrendArgs) -> Result<(), InterfaceError> {
     let state_dir = get_state_dir();
     let db_path = state_dir.join("harness.db");
@@ -2834,6 +2866,7 @@ fn run_trend(args: &EvalTrendArgs) -> Result<(), InterfaceError> {
     Ok(())
 }
 
+#[cfg(feature = "eval")]
 fn run_accept(args: &EvalAcceptArgs) -> Result<(), InterfaceError> {
     let state_dir = get_state_dir();
     let db_path = state_dir.join("harness.db");
@@ -2903,6 +2936,7 @@ fn run_accept(args: &EvalAcceptArgs) -> Result<(), InterfaceError> {
     Ok(())
 }
 
+#[cfg(feature = "eval")]
 fn run_apply(args: &EvalApplyArgs) -> Result<(), InterfaceError> {
     let state_dir = get_state_dir();
     let db_path = state_dir.join("harness.db");
@@ -2969,6 +3003,7 @@ fn run_apply(args: &EvalApplyArgs) -> Result<(), InterfaceError> {
     Ok(())
 }
 
+#[cfg(feature = "eval")]
 fn run_ab(args: &EvalAbArgs) -> Result<(), InterfaceError> {
     let state_dir = get_state_dir();
     let db_path = state_dir.join("harness.db");
@@ -3069,6 +3104,7 @@ fn run_ab(args: &EvalAbArgs) -> Result<(), InterfaceError> {
     Ok(())
 }
 
+#[cfg(feature = "eval")]
 fn run_rebaseline(args: &EvalRebaselineArgs) -> Result<(), InterfaceError> {
     println!("[eval-harness] Rebaselining skill '{}'...", args.skill);
     run_baseline(&EvalBaselineArgs {
diff --git a/crates/harness-cli/src/main.rs b/crates/harness-cli/src/main.rs
index 6884951..c1d7a8b 100644
--- a/crates/harness-cli/src/main.rs
+++ b/crates/harness-cli/src/main.rs
@@ -1,5 +1,6 @@
 mod application;
 mod domain;
+#[cfg(feature = "eval")]
 mod eval;
 mod infrastructure;
 mod interface;