fix(semgrep): extract CWE from list-form metadata (was always null)

semgrep emits metadata.cwe as a LIST (e.g. "CWE-798: Use of Hard-coded
Credentials"), but the extractor called v.as_str() on it -> always None, so SAST
findings never carried a CWE at all (silently breaking control mapping AND the
CWE-based dedup). Handle list + bare-string forms and normalise to the CWE-NNN
id. Found by the live control-triage end-to-end test.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

.gitea/workflows/ci.yml

@@ -109,6 +109,8 @@ jobs:
         run: cargo clippy -p compliance-mcp -- -D warnings
       - name: Clippy (werkbank-exec)
         run: cargo clippy -p werkbank-exec -- -D warnings
+      - name: Clippy (control-map)
+        run: cargo clippy -p control-map -- -D warnings
 
       # Security audit
       - name: Security Audit
@@ -117,8 +119,8 @@ jobs:
           RUSTC_WRAPPER: ""
 
       # Tests (reuses compilation artifacts from clippy)
-      - name: Tests (core + agent + werkbank-exec)
+      - name: Tests (core + agent + werkbank-exec + control-map)
-        run: cargo test -p compliance-core -p compliance-agent -p werkbank-exec --lib
+        run: cargo test -p compliance-core -p compliance-agent -p werkbank-exec -p control-map --lib
       - name: Tests (dashboard server)
         run: cargo test -p compliance-dashboard --features server --no-default-features
       - name: Tests (dashboard web)

Cargo.lock

@@ -666,6 +666,7 @@ dependencies = [
  "compliance-core",
  "compliance-dast",
  "compliance-graph",
+ "control-map",
  "dashmap",
  "dotenvy",
  "futures-core",
@@ -968,6 +969,15 @@ dependencies = [
  "charset",
 ]
 
+[[package]]
+name = "control-map"
+version = "0.1.0"
+dependencies = [
+ "serde",
+ "serde_json",
+ "thiserror 2.0.18",
+]
+
 [[package]]
 name = "convert_case"
 version = "0.8.0"

Cargo.toml

@@ -8,6 +8,7 @@ members = [
     "compliance-mcp",
     "compliance-smoke",
     "werkbank-exec",
+    "control-map",
 ]
 resolver = "2"
 
@@ -17,6 +18,7 @@ expect_used = "deny"
 
 [workspace.dependencies]
 compliance-core = { path = "compliance-core", default-features = false }
+control-map = { path = "control-map" }
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
 tokio = { version = "1", features = ["full"] }

compliance-agent/Cargo.toml

@@ -8,6 +8,7 @@ workspace = true
 
 [dependencies]
 compliance-core = { workspace = true, features = ["mongodb", "telemetry", "axum"] }
+control-map = { workspace = true }
 compliance-graph = { path = "../compliance-graph" }
 compliance-dast = { path = "../compliance-dast" }
 # Shared dynamic-execution logic (soft-PLC provisioning + ICS probing), also

compliance-agent/src/controls/checker.rs

@@ -0,0 +1,114 @@
+//! The grounded control checker: judge each candidate region for a control, then
+//! keep only the verdicts that survive the grounding gate.
+//!
+//! Generic over [`ControlJudge`] so tests drive it with a deterministic stub —
+//! the whole recognize → ground path is then exercised without an LLM. With the
+//! real judge, determinism comes from temperature 0 plus the gate.
+
+use compliance_core::control_check::{ground, CandidateRegion, ControlCheckSpec};
+use compliance_core::models::Finding;
+
+use super::judge::ControlJudge;
+
+/// Runs a [`ControlJudge`] over candidate regions and grounds the results.
+pub struct GroundedControlChecker<J> {
+    judge: J,
+}
+
+impl<J: ControlJudge> GroundedControlChecker<J> {
+    pub fn new(judge: J) -> Self {
+        Self { judge }
+    }
+
+    /// Judge every candidate region for `spec` and return the grounded findings.
+    /// A verdict that doesn't quote real code in its region is dropped by
+    /// [`ground`], so nothing fabricated reaches the caller.
+    pub async fn check(
+        &self,
+        spec: &ControlCheckSpec,
+        regions: &[CandidateRegion],
+        repo_id: &str,
+    ) -> Vec<Finding> {
+        let mut findings = Vec::new();
+        for region in regions {
+            let verdict = self.judge.judge(spec, region).await;
+            if let Some(finding) = ground(spec, region, &verdict, repo_id) {
+                findings.push(finding);
+            }
+        }
+        findings
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use compliance_core::control_check::LlmVerdict;
+    use compliance_core::models::finding::Severity;
+
+    /// Deterministic stub: returns a fixed verdict for every region, so the
+    /// recognize → ground composition is tested without an LLM.
+    struct StubJudge {
+        verdict: LlmVerdict,
+    }
+
+    impl ControlJudge for StubJudge {
+        async fn judge(&self, _spec: &ControlCheckSpec, _region: &CandidateRegion) -> LlmVerdict {
+            self.verdict.clone()
+        }
+    }
+
+    fn spec() -> ControlCheckSpec {
+        ControlCheckSpec {
+            control_id: "cra-ai-8".into(),
+            title: "No default passwords".into(),
+            requirement: "No default credentials".into(),
+            default_cwe: Some("CWE-798".into()),
+            severity: Severity::High,
+        }
+    }
+
+    fn region(content: &str) -> CandidateRegion {
+        CandidateRegion {
+            file: "src/auth.py".into(),
+            start_line: 1,
+            content: content.into(),
+        }
+    }
+
+    #[tokio::test]
+    async fn keeps_grounded_and_drops_ungrounded() {
+        let checker = GroundedControlChecker::new(StubJudge {
+            verdict: LlmVerdict {
+                violates: true,
+                snippet: "PASSWORD = \"admin\"".into(),
+                cwe: None,
+                confidence: 0.9,
+            },
+        });
+        let regions = vec![
+            region("x = 1\nPASSWORD = \"admin\"\n"), // quotes real code → grounded
+            region("totally unrelated code\n"),      // snippet absent → dropped
+        ];
+        let findings = checker.check(&spec(), &regions, "repo").await;
+        assert_eq!(findings.len(), 1);
+        assert_eq!(findings[0].control_refs, vec!["cra-ai-8".to_string()]);
+        assert_eq!(findings[0].line_number, Some(2));
+    }
+
+    #[tokio::test]
+    async fn non_violation_yields_nothing() {
+        let checker = GroundedControlChecker::new(StubJudge {
+            verdict: LlmVerdict {
+                violates: false,
+                snippet: String::new(),
+                cwe: None,
+                confidence: 0.0,
+            },
+        });
+        let findings = checker
+            .check(&spec(), &[region("PASSWORD = \"admin\"\n")], "repo")
+            .await;
+        assert!(findings.is_empty());
+    }
+}

compliance-agent/src/controls/judge.rs

@@ -0,0 +1,167 @@
+//! The "recognize" stage: judge whether a code region violates a control.
+//!
+//! Behind the [`ControlJudge`] trait so the grounded checker can be driven by a
+//! deterministic stub in tests. The real [`LlmControlJudge`] runs the model at
+//! temperature 0 with a closed prompt — it must quote the offending code VERBATIM,
+//! and everything it returns is then re-checked by the grounding gate
+//! ([`compliance_core::control_check::ground`]). The judge is allowed to be
+//! smart; it is never trusted.
+
+use std::sync::Arc;
+
+use serde::Deserialize;
+
+use compliance_core::control_check::{CandidateRegion, ControlCheckSpec, LlmVerdict};
+
+use crate::llm::LlmClient;
+
+/// Prompt/logic version — part of the verdict cache key, bump on any change here.
+pub const PROMPT_VERSION: &str = "control-judge-v1";
+
+const SYSTEM_PROMPT: &str = "You are a precise security & compliance code auditor. \
+You are given ONE compliance control (a requirement) and ONE code region. Decide \
+ONLY whether the code region VIOLATES the control. Rules: (1) Judge only the code \
+shown — never assume code that is not present. (2) If and only if it violates, copy \
+the EXACT offending code VERBATIM into `snippet`, character-for-character from the \
+region — do not paraphrase, reformat, or reconstruct it. (3) If it does not clearly \
+violate, set violates=false and leave snippet empty. (4) Prefer false over guessing. \
+Respond with STRICT JSON only, no prose: \
+{\"violates\": bool, \"snippet\": \"<verbatim code or empty>\", \"cwe\": \"CWE-NNN or null\", \"confidence\": 0.0-1.0}";
+
+/// Judges one (control, region). Async-in-trait so a stub can drive tests.
+#[allow(async_fn_in_trait)]
+pub trait ControlJudge: Send + Sync {
+    async fn judge(&self, spec: &ControlCheckSpec, region: &CandidateRegion) -> LlmVerdict;
+}
+
+/// The real judge: the LLM at temperature 0 with the closed, verbatim-snippet prompt.
+pub struct LlmControlJudge {
+    llm: Arc<LlmClient>,
+}
+
+impl LlmControlJudge {
+    pub fn new(llm: Arc<LlmClient>) -> Self {
+        Self { llm }
+    }
+}
+
+impl ControlJudge for LlmControlJudge {
+    async fn judge(&self, spec: &ControlCheckSpec, region: &CandidateRegion) -> LlmVerdict {
+        let user = build_user_prompt(spec, region);
+        match self.llm.chat(SYSTEM_PROMPT, &user, Some(0.0)).await {
+            Ok(response) => parse_verdict(&response),
+            Err(e) => {
+                // Fail closed: a transient model error yields no finding, never a
+                // fabricated one.
+                tracing::warn!(control = %spec.control_id, error = %e, "control judge call failed");
+                no_violation()
+            }
+        }
+    }
+}
+
+fn build_user_prompt(spec: &ControlCheckSpec, region: &CandidateRegion) -> String {
+    format!(
+        "CONTROL {id} — {title}\nRequirement: {req}\n\nCODE ({file}, first line = {line}):\n```\n{code}\n```\n\nReturn the JSON verdict.",
+        id = spec.control_id,
+        title = spec.title,
+        req = spec.requirement,
+        file = region.file,
+        line = region.start_line,
+        code = region.content,
+    )
+}
+
+#[derive(Debug, Default, Deserialize)]
+struct RawVerdict {
+    #[serde(default)]
+    violates: bool,
+    #[serde(default)]
+    snippet: String,
+    #[serde(default)]
+    cwe: Option<String>,
+    #[serde(default)]
+    confidence: f64,
+}
+
+/// Parse the model's JSON verdict, tolerant of ```json fencing. Any parse failure
+/// degrades to a non-violation (never a fabricated finding).
+fn parse_verdict(response: &str) -> LlmVerdict {
+    let cleaned = response
+        .trim()
+        .trim_start_matches("```json")
+        .trim_start_matches("```")
+        .trim_end_matches("```")
+        .trim();
+    match serde_json::from_str::<RawVerdict>(cleaned) {
+        Ok(raw) => LlmVerdict {
+            violates: raw.violates,
+            snippet: raw.snippet,
+            cwe: raw.cwe.filter(|c| !c.trim().is_empty()),
+            confidence: raw.confidence,
+        },
+        Err(e) => {
+            tracing::debug!(error = %e, "failed to parse control verdict; treating as non-violation");
+            no_violation()
+        }
+    }
+}
+
+fn no_violation() -> LlmVerdict {
+    LlmVerdict {
+        violates: false,
+        snippet: String::new(),
+        cwe: None,
+        confidence: 0.0,
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use compliance_core::models::finding::Severity;
+
+    fn spec() -> ControlCheckSpec {
+        ControlCheckSpec {
+            control_id: "cra-ai-8".into(),
+            title: "No default passwords".into(),
+            requirement: "Products must not ship default credentials".into(),
+            default_cwe: Some("CWE-798".into()),
+            severity: Severity::High,
+        }
+    }
+
+    #[test]
+    fn parses_plain_and_fenced_json() {
+        let plain = r#"{"violates": true, "snippet": "PASSWORD = \"x\"", "cwe": "CWE-798", "confidence": 0.9}"#;
+        let v = parse_verdict(plain);
+        assert!(v.violates);
+        assert_eq!(v.snippet, "PASSWORD = \"x\"");
+        assert_eq!(v.cwe.as_deref(), Some("CWE-798"));
+
+        let fenced = "```json\n{\"violates\": false, \"snippet\": \"\", \"cwe\": null, \"confidence\": 0.1}\n```";
+        assert!(!parse_verdict(fenced).violates);
+    }
+
+    #[test]
+    fn garbage_and_empty_cwe_are_safe() {
+        assert!(!parse_verdict("not json at all").violates); // fail closed
+        let no_cwe =
+            parse_verdict(r#"{"violates": true, "snippet": "x", "cwe": "  ", "confidence": 0.5}"#);
+        assert!(no_cwe.cwe.is_none()); // blank CWE normalised away
+    }
+
+    #[test]
+    fn user_prompt_carries_control_and_code() {
+        let region = CandidateRegion {
+            file: "src/auth.py".into(),
+            start_line: 10,
+            content: "PASSWORD = \"admin\"".into(),
+        };
+        let p = build_user_prompt(&spec(), &region);
+        assert!(p.contains("cra-ai-8"));
+        assert!(p.contains("Products must not ship default credentials"));
+        assert!(p.contains("PASSWORD = \"admin\""));
+        assert!(p.contains("src/auth.py"));
+    }
+}

compliance-agent/src/controls/mod.rs

@@ -5,6 +5,14 @@
 //! [`OscalControlsProvider`], which pulls breakpilot-compliance's OSCAL catalog
 //! and snapshots it locally.
 
+mod checker;
+mod judge;
 mod oscal_provider;
+mod scan_triage;
+mod triage;
 
+pub use checker::GroundedControlChecker;
+pub use judge::{ControlJudge, LlmControlJudge, PROMPT_VERSION};
 pub use oscal_provider::OscalControlsProvider;
+pub use scan_triage::triage_repo_findings;
+pub use triage::{ControlTriage, TriageOutcome};

compliance-agent/src/controls/scan_triage.rs

@@ -0,0 +1,139 @@
+//! Scan-pipeline integration for control triage.
+//!
+//! After the deterministic tools have produced findings, this stamps each finding
+//! with the compliance control(s) it's evidence for and marks control-level false
+//! positives — using the ingested OSCAL catalog for control text, the
+//! `control-map` LUT for the finding→control link, and the grounded LLM judge to
+//! confirm. Skipped entirely unless breakpilot is configured.
+
+use std::collections::HashMap;
+use std::path::Path;
+use std::sync::Arc;
+
+use compliance_core::control_check::{CandidateRegion, ControlCheckSpec};
+use compliance_core::models::finding::{Finding, FindingStatus, Severity};
+use compliance_core::models::onboarding::ComplianceFramework;
+use compliance_core::AgentConfig;
+use control_map::ControlMap;
+
+use super::{ControlTriage, LlmControlJudge, OscalControlsProvider, TriageOutcome};
+use crate::llm::LlmClient;
+
+/// Lines of context to read on each side of a finding's line.
+const REGION_WINDOW: usize = 6;
+
+/// Triage every finding in `findings` against the CRA control map: stamp
+/// `control_refs` on confirmed findings and flag control false positives. Returns
+/// the number of findings tagged with at least one control.
+pub async fn triage_repo_findings(
+    config: &AgentConfig,
+    llm: Arc<LlmClient>,
+    repo_path: &Path,
+    findings: &mut [Finding],
+) -> usize {
+    let Some(base_url) = config.breakpilot.base_url.clone() else {
+        return 0; // control triage is opt-in via BREAKPILOT_BASE_URL
+    };
+    let provider = OscalControlsProvider::new(
+        reqwest::Client::new(),
+        base_url,
+        config.breakpilot.token.clone(),
+        &config.breakpilot.snapshot_dir,
+    );
+    let specs = build_specs(&provider).await;
+    if specs.is_empty() {
+        return 0;
+    }
+    let map = match ControlMap::cra() {
+        Ok(m) => m,
+        Err(e) => {
+            tracing::warn!(error = %e, "control map failed to load; skipping control triage");
+            return 0;
+        }
+    };
+    let triage = ControlTriage::new(LlmControlJudge::new(llm), map, specs);
+
+    let mut tagged = 0;
+    for finding in findings.iter_mut() {
+        let (Some(file), Some(line)) = (finding.file_path.clone(), finding.line_number) else {
+            continue;
+        };
+        let Some(region) = fetch_region(repo_path, &file, line) else {
+            continue;
+        };
+        match triage.triage(finding, &region).await {
+            TriageOutcome::Confirmed(controls) => {
+                finding.control_refs = controls;
+                tagged += 1;
+            }
+            TriageOutcome::FalsePositive => {
+                finding.status = FindingStatus::FalsePositive;
+                finding.triage_action = Some("control_false_positive".to_string());
+            }
+            TriageOutcome::Unmapped => {}
+        }
+    }
+    tagged
+}
+
+/// Build the control requirement specs (by id) from the ingested OSCAL catalog.
+async fn build_specs(provider: &OscalControlsProvider) -> HashMap<String, ControlCheckSpec> {
+    let mut specs = HashMap::new();
+    match provider.load(ComplianceFramework::Cra).await {
+        Ok(doc) => {
+            for control in doc.to_controls() {
+                specs.insert(
+                    control.id.clone(),
+                    ControlCheckSpec {
+                        control_id: control.id,
+                        title: control.title,
+                        requirement: control.text,
+                        default_cwe: None,
+                        severity: Severity::Medium,
+                    },
+                );
+            }
+        }
+        Err(e) => tracing::warn!(error = %e, "could not load control catalog for triage"),
+    }
+    specs
+}
+
+/// Read a window of lines around `line` (1-based) from `repo_path/file`.
+fn fetch_region(repo_path: &Path, file: &str, line: u32) -> Option<CandidateRegion> {
+    let content = std::fs::read_to_string(repo_path.join(file)).ok()?;
+    let lines: Vec<&str> = content.lines().collect();
+    if lines.is_empty() {
+        return None;
+    }
+    let center = (line.saturating_sub(1) as usize).min(lines.len() - 1);
+    let start = center.saturating_sub(REGION_WINDOW);
+    let end = (center + REGION_WINDOW + 1).min(lines.len());
+    Some(CandidateRegion {
+        file: file.to_string(),
+        start_line: (start as u32) + 1,
+        content: lines[start..end].join("\n"),
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn fetch_region_windows_around_the_line() {
+        let dir = std::env::temp_dir().join(format!("triage-region-{}", uuid::Uuid::new_v4()));
+        std::fs::create_dir_all(&dir).unwrap();
+        let file = "a.py";
+        std::fs::write(dir.join(file), "l1\nl2\nl3\nSECRET=1\nl5\nl6\n").unwrap();
+        let r = fetch_region(&dir, file, 4).unwrap();
+        assert!(r.content.contains("SECRET=1"));
+        assert_eq!(r.start_line, 1); // window clamps to file start
+        let _ = std::fs::remove_dir_all(&dir);
+    }
+
+    #[test]
+    fn fetch_region_missing_file_is_none() {
+        assert!(fetch_region(Path::new("/nonexistent"), "nope.py", 1).is_none());
+    }
+}

compliance-agent/src/controls/triage.rs

@@ -0,0 +1,184 @@
+//! Triage step: confirm/refute a deterministic tool finding against the controls
+//! it maps to (via the `control-map` LUT), grounding the judgment.
+//!
+//! This is where the LLM finally enters — as a **false-positive filter over tool
+//! output**, never as the detector (the ZeroFalse / IRIS pattern). A tool
+//! (semgrep, gitleaks, syft/osv) detects deterministically; `controls_for(tool,
+//! cwe)` attaches the finding to the control(s) it's evidence for; the grounded
+//! judge then confirms or refutes each, and only judgments anchored to real code
+//! survive.
+
+use std::collections::HashMap;
+
+use compliance_core::control_check::{ground, CandidateRegion, ControlCheckSpec};
+use compliance_core::models::Finding;
+use control_map::ControlMap;
+
+use super::judge::ControlJudge;
+
+/// What triage decided for one tool finding.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum TriageOutcome {
+    /// The finding maps to no control in the LUT — keep it, untagged.
+    Unmapped,
+    /// Maps to controls and the grounded judge confirmed at least one — keep the
+    /// finding and tag it with these control ids.
+    Confirmed(Vec<String>),
+    /// Maps to controls but the judge grounded none — treat as a false positive.
+    FalsePositive,
+}
+
+/// Triages tool findings against the control map, confirming with a grounded judge.
+pub struct ControlTriage<J> {
+    judge: J,
+    map: ControlMap,
+    /// Control requirement specs (by control id), built from the ingested catalog.
+    specs: HashMap<String, ControlCheckSpec>,
+}
+
+impl<J: ControlJudge> ControlTriage<J> {
+    pub fn new(judge: J, map: ControlMap, specs: HashMap<String, ControlCheckSpec>) -> Self {
+        Self { judge, map, specs }
+    }
+
+    /// Triage one tool finding. `region` is the code around the finding, used as
+    /// the grounding evidence for the judge.
+    pub async fn triage(&self, finding: &Finding, region: &CandidateRegion) -> TriageOutcome {
+        let Some(cwe) = finding.cwe.as_deref() else {
+            return TriageOutcome::Unmapped;
+        };
+        let mapped = self.map.controls_for(&finding.scanner, cwe);
+        if mapped.is_empty() {
+            return TriageOutcome::Unmapped;
+        }
+
+        let mut confirmed = Vec::new();
+        for entry in mapped {
+            let Some(spec) = self.specs.get(&entry.control) else {
+                continue;
+            };
+            let verdict = self.judge.judge(spec, region).await;
+            // The verdict only counts if it grounds to real code in the region.
+            if ground(spec, region, &verdict, &finding.repo_id).is_some() {
+                confirmed.push(entry.control.clone());
+            }
+        }
+
+        if confirmed.is_empty() {
+            TriageOutcome::FalsePositive
+        } else {
+            TriageOutcome::Confirmed(confirmed)
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use compliance_core::control_check::LlmVerdict;
+    use compliance_core::models::finding::Severity;
+    use compliance_core::models::scan::ScanType;
+
+    struct StubJudge {
+        verdict: LlmVerdict,
+    }
+    impl ControlJudge for StubJudge {
+        async fn judge(&self, _s: &ControlCheckSpec, _r: &CandidateRegion) -> LlmVerdict {
+            self.verdict.clone()
+        }
+    }
+
+    fn specs() -> HashMap<String, ControlCheckSpec> {
+        let mut m = HashMap::new();
+        m.insert(
+            "cra-ai-8".to_string(),
+            ControlCheckSpec {
+                control_id: "cra-ai-8".into(),
+                title: "No default passwords".into(),
+                requirement: "No default credentials".into(),
+                default_cwe: Some("CWE-798".into()),
+                severity: Severity::High,
+            },
+        );
+        m
+    }
+
+    fn semgrep_finding(cwe: &str) -> Finding {
+        let mut f = Finding::new(
+            "repo".into(),
+            "fp1".into(),
+            "semgrep".into(),
+            ScanType::Sast,
+            "hardcoded credential".into(),
+            "desc".into(),
+            Severity::High,
+        );
+        f.cwe = Some(cwe.into());
+        f
+    }
+
+    fn region() -> CandidateRegion {
+        CandidateRegion {
+            file: "src/auth.py".into(),
+            start_line: 1,
+            content: "PASSWORD = \"admin123\"\n".into(),
+        }
+    }
+
+    #[tokio::test]
+    async fn confirmed_finding_is_tagged_with_control() {
+        let triage = ControlTriage::new(
+            StubJudge {
+                verdict: LlmVerdict {
+                    violates: true,
+                    snippet: "PASSWORD = \"admin123\"".into(),
+                    cwe: None,
+                    confidence: 0.9,
+                },
+            },
+            ControlMap::cra().unwrap(),
+            specs(),
+        );
+        let out = triage.triage(&semgrep_finding("CWE-798"), &region()).await;
+        assert_eq!(out, TriageOutcome::Confirmed(vec!["cra-ai-8".to_string()]));
+    }
+
+    #[tokio::test]
+    async fn refuted_mapped_finding_is_false_positive() {
+        // Maps to cra-ai-8, but the judge doesn't confirm (no violation) → FP.
+        let triage = ControlTriage::new(
+            StubJudge {
+                verdict: LlmVerdict {
+                    violates: false,
+                    snippet: String::new(),
+                    cwe: None,
+                    confidence: 0.1,
+                },
+            },
+            ControlMap::cra().unwrap(),
+            specs(),
+        );
+        let out = triage.triage(&semgrep_finding("CWE-798"), &region()).await;
+        assert_eq!(out, TriageOutcome::FalsePositive);
+    }
+
+    #[tokio::test]
+    async fn unmapped_cwe_is_left_untagged() {
+        let triage = ControlTriage::new(
+            StubJudge {
+                verdict: LlmVerdict {
+                    violates: true,
+                    snippet: "PASSWORD = \"admin123\"".into(),
+                    cwe: None,
+                    confidence: 0.9,
+                },
+            },
+            ControlMap::cra().unwrap(),
+            specs(),
+        );
+        let out = triage
+            .triage(&semgrep_finding("CWE-99999"), &region())
+            .await;
+        assert_eq!(out, TriageOutcome::Unmapped);
+    }
+}

compliance-agent/src/pipeline/orchestrator.rs

@@ -215,6 +215,21 @@ impl PipelineOrchestrator {
         .await;
         tracing::info!("[{repo_id}] Triaged: {triaged} findings passed confidence threshold");
 
+        // Stage 5b: control triage — stamp findings with the compliance control(s)
+        // they're evidence for and flag control false positives (grounded LLM over
+        // deterministic tool output). No-op unless breakpilot is configured.
+        self.update_phase(scan_run_id, "control_triage").await;
+        let tagged = crate::controls::triage_repo_findings(
+            &self.config,
+            self.llm.clone(),
+            &repo_path,
+            &mut all_findings,
+        )
+        .await;
+        if tagged > 0 {
+            tracing::info!("[{repo_id}] Control triage tagged {tagged} findings with control refs");
+        }
+
         // Dedup against existing findings and insert new ones
         let mut new_count = 0u32;
         let mut new_findings: Vec<Finding> = Vec::new();

compliance-agent/src/pipeline/semgrep.rs

@@ -82,10 +82,7 @@ impl Scanner for SemgrepScanner {
                 finding.file_path = Some(r.path);
                 finding.line_number = Some(r.start.line);
                 finding.code_snippet = Some(r.extra.lines);
-                finding.cwe = r
+                finding.cwe = r.extra.metadata.as_ref().and_then(extract_cwe);
-                    .extra
-                    .metadata
-                    .and_then(|m| m.get("cwe").and_then(|v| v.as_str()).map(|s| s.to_string()));
                 finding
             })
             .collect();
@@ -124,10 +121,34 @@ struct SemgrepExtra {
     metadata: Option<serde_json::Value>,
 }
 
+/// semgrep emits `metadata.cwe` as a list of strings like
+/// `"CWE-798: Use of Hard-coded Credentials"` (occasionally a bare string). Take
+/// the first entry and normalise it to just the `CWE-NNN` id.
+fn extract_cwe(metadata: &serde_json::Value) -> Option<String> {
+    let raw = metadata.get("cwe")?;
+    let text = match raw {
+        serde_json::Value::Array(items) => items.first()?.as_str()?,
+        serde_json::Value::String(s) => s.as_str(),
+        _ => return None,
+    };
+    let id = text.split(':').next().unwrap_or(text).trim();
+    (!id.is_empty()).then(|| id.to_string())
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
 
+    #[test]
+    fn extract_cwe_handles_list_and_normalises() {
+        let md = serde_json::json!({"cwe": ["CWE-798: Use of Hard-coded Credentials"]});
+        assert_eq!(extract_cwe(&md).as_deref(), Some("CWE-798"));
+        let bare = serde_json::json!({"cwe": "CWE-89"});
+        assert_eq!(extract_cwe(&bare).as_deref(), Some("CWE-89"));
+        let none = serde_json::json!({"severity": "ERROR"});
+        assert_eq!(extract_cwe(&none), None);
+    }
+
     #[test]
     fn deserialize_semgrep_output() {
         let json = r#"{

compliance-core/src/control_check.rs

@@ -0,0 +1,204 @@
+//! Grounded control-driven checking.
+//!
+//! Turns a *text* control into findings via an LLM used as a **pattern-recognizer**
+//! whose output is grounded to real code — so a hallucinated finding cannot
+//! survive. Determinism is structural, not a prompt plea:
+//!
+//! 1. the LLM only ever judges *retrieved* regions — it can't invent findings in
+//!    code it never saw;
+//! 2. a verdict becomes a finding only if its quoted snippet appears **verbatim**
+//!    in the region, and the line is recomputed from that match — the model's own
+//!    line number is never trusted ([`ground`]);
+//! 3. verdicts are cached by content hash ([`cache_key`]) so re-scans reproduce.
+//!
+//! The LLM supplies cross-language / cross-stack pattern recognition; this module
+//! supplies the determinism.
+
+use sha2::{Digest, Sha256};
+
+use crate::models::finding::{Finding, Severity};
+use crate::models::scan::ScanType;
+
+/// A control rendered as a check the LLM judges code against.
+#[derive(Debug, Clone)]
+pub struct ControlCheckSpec {
+    /// Stable control id, e.g. `"cra-ai-8"`.
+    pub control_id: String,
+    /// Short control title (used in the finding title).
+    pub title: String,
+    /// The requirement text the LLM judges against (control objective/statement).
+    pub requirement: String,
+    /// CWE to fall back to when the model doesn't supply one.
+    pub default_cwe: Option<String>,
+    /// Severity for findings raised from this control.
+    pub severity: Severity,
+}
+
+/// A retrieved code region the LLM judges — never the whole repo.
+#[derive(Debug, Clone)]
+pub struct CandidateRegion {
+    /// Repo-relative path.
+    pub file: String,
+    /// 1-based line number of the region's first line in `file`.
+    pub start_line: u32,
+    /// The region's source text.
+    pub content: String,
+}
+
+/// The LLM's structured verdict for one (control, region). `snippet` is the
+/// verbatim code the model claims proves the violation — it is the anchor the
+/// grounding gate checks.
+#[derive(Debug, Clone)]
+pub struct LlmVerdict {
+    pub violates: bool,
+    pub snippet: String,
+    pub cwe: Option<String>,
+    pub confidence: f64,
+}
+
+/// The grounding gate. A verdict becomes a [`Finding`] only if it claims a
+/// violation AND its quoted `snippet` appears verbatim in `region.content`; the
+/// finding's line is computed from the match, so a fabricated or mis-located
+/// snippet is dropped. Pure — no LLM, no I/O.
+pub fn ground(
+    spec: &ControlCheckSpec,
+    region: &CandidateRegion,
+    verdict: &LlmVerdict,
+    repo_id: &str,
+) -> Option<Finding> {
+    if !verdict.violates {
+        return None;
+    }
+    let snippet = verdict.snippet.trim();
+    if snippet.is_empty() {
+        return None;
+    }
+    // Grounding: the quoted snippet must literally exist in the retrieved region.
+    let pos = region.content.find(snippet)?;
+    // Recompute the real line from the match — never trust the model's number.
+    let newlines_before = region.content[..pos].matches('\n').count();
+    let line = region.start_line + newlines_before as u32;
+
+    let mut finding = Finding::new(
+        repo_id.to_string(),
+        control_finding_fingerprint(&spec.control_id, &region.file, snippet),
+        "control-check".to_string(),
+        ScanType::CodeReview,
+        format!("{}: {}", spec.control_id, spec.title),
+        format!(
+            "Control {} appears violated ({}) at {}:{line}",
+            spec.control_id, spec.requirement, region.file
+        ),
+        spec.severity.clone(),
+    );
+    finding.cwe = verdict.cwe.clone().or_else(|| spec.default_cwe.clone());
+    finding.file_path = Some(region.file.clone());
+    finding.line_number = Some(line);
+    finding.code_snippet = Some(snippet.to_string());
+    finding.confidence = Some(verdict.confidence);
+    // Carry the control reference on the finding.
+    finding.control_refs = vec![spec.control_id.clone()];
+    Some(finding)
+}
+
+/// Deterministic cache key for a (control, region, model, prompt-version) verdict
+/// so identical inputs reproduce the same verdict without another LLM call.
+pub fn cache_key(
+    control_id: &str,
+    region_content: &str,
+    model: &str,
+    prompt_version: &str,
+) -> String {
+    hash_parts(&[control_id, region_content, model, prompt_version])
+}
+
+fn control_finding_fingerprint(control_id: &str, file: &str, snippet: &str) -> String {
+    hash_parts(&[control_id, file, snippet])
+}
+
+fn hash_parts(parts: &[&str]) -> String {
+    let mut hasher = Sha256::new();
+    for part in parts {
+        hasher.update(part.as_bytes());
+        hasher.update([0u8]); // domain separator between parts
+    }
+    hex::encode(hasher.finalize())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn spec() -> ControlCheckSpec {
+        ControlCheckSpec {
+            control_id: "cra-ai-8".into(),
+            title: "No default passwords".into(),
+            requirement: "Products must not ship default credentials".into(),
+            default_cwe: Some("CWE-798".into()),
+            severity: Severity::High,
+        }
+    }
+
+    fn region() -> CandidateRegion {
+        CandidateRegion {
+            file: "src/auth.py".into(),
+            start_line: 10,
+            content: "def login():\n    PASSWORD = \"admin123\"\n    return PASSWORD\n".into(),
+        }
+    }
+
+    #[test]
+    fn grounds_real_snippet_with_recomputed_line() {
+        let v = LlmVerdict {
+            violates: true,
+            snippet: "PASSWORD = \"admin123\"".into(),
+            cwe: None,
+            confidence: 0.9,
+        };
+        let f = ground(&spec(), &region(), &v, "repo").expect("should ground");
+        assert_eq!(f.line_number, Some(11)); // 2nd line of a region starting at 10
+        assert_eq!(f.cwe.as_deref(), Some("CWE-798")); // fell back to the spec default
+        assert_eq!(f.control_refs, vec!["cra-ai-8".to_string()]); // control ref carried
+        assert_eq!(f.file_path.as_deref(), Some("src/auth.py"));
+        assert_eq!(f.code_snippet.as_deref(), Some("PASSWORD = \"admin123\""));
+    }
+
+    #[test]
+    fn drops_fabricated_snippet_not_in_region() {
+        let v = LlmVerdict {
+            violates: true,
+            snippet: "SECRET = \"totally-made-up\"".into(),
+            cwe: None,
+            confidence: 0.99,
+        };
+        assert!(ground(&spec(), &region(), &v, "repo").is_none());
+    }
+
+    #[test]
+    fn drops_non_violation_and_empty_snippet() {
+        let no = LlmVerdict {
+            violates: false,
+            snippet: "PASSWORD = \"admin123\"".into(),
+            cwe: None,
+            confidence: 0.9,
+        };
+        assert!(ground(&spec(), &region(), &no, "repo").is_none());
+        let empty = LlmVerdict {
+            violates: true,
+            snippet: "   ".into(),
+            cwe: None,
+            confidence: 0.9,
+        };
+        assert!(ground(&spec(), &region(), &empty, "repo").is_none());
+    }
+
+    #[test]
+    fn cache_key_and_fingerprint_are_deterministic() {
+        assert_eq!(cache_key("c", "x", "m", "v"), cache_key("c", "x", "m", "v"));
+        assert_ne!(cache_key("c", "x", "m", "v"), cache_key("c", "y", "m", "v"));
+        assert_eq!(
+            control_finding_fingerprint("c", "f", "s"),
+            control_finding_fingerprint("c", "f", "s")
+        );
+    }
+}

compliance-core/src/lib.rs

@@ -1,4 +1,5 @@
 pub mod config;
+pub mod control_check;
 pub mod db;
 pub mod error;
 pub mod models;

compliance-core/src/models/finding.rs

@@ -76,6 +76,10 @@ pub struct Finding {
     pub triage_rationale: Option<String>,
     /// Developer feedback on finding quality
     pub developer_feedback: Option<String>,
+    /// Compliance control ids this finding is evidence for (stamped by control
+    /// triage against the `control-map` LUT). Empty when unmapped.
+    #[serde(default)]
+    pub control_refs: Vec<String>,
     #[serde(with = "super::serde_helpers::bson_datetime")]
     pub created_at: DateTime<Utc>,
     #[serde(with = "super::serde_helpers::bson_datetime")]
@@ -118,6 +122,7 @@ impl Finding {
             triage_action: None,
             triage_rationale: None,
             developer_feedback: None,
+            control_refs: Vec::new(),
             created_at: now,
             updated_at: now,
         }

control-map/Cargo.toml

@@ -0,0 +1,12 @@
+[package]
+name = "control-map"
+version = "0.1.0"
+edition = "2021"
+
+[lints]
+workspace = true
+
+[dependencies]
+serde = { workspace = true }
+serde_json = { workspace = true }
+thiserror = { workspace = true }

control-map/data/cra_control_map.json

@@ -0,0 +1,399 @@
+{
+  "version": "1.0",
+  "framework": "cra",
+  "controls": [
+    {
+      "control": "cra-ai-1",
+      "title": "Secure-by-Default-Konfiguration",
+      "scans": [],
+      "note": "code-checkable but no off-the-shelf tool digs it out — author a detector (custom semgrep rule / check)",
+      "status": "needs_tooling"
+    },
+    {
+      "control": "cra-ai-2",
+      "title": "Minimale Angriffsflaeche",
+      "scans": [],
+      "note": "code-checkable but no off-the-shelf tool digs it out — author a detector (custom semgrep rule / check)",
+      "status": "needs_tooling"
+    },
+    {
+      "control": "cra-ai-3",
+      "title": "Sichere Systemarchitektur",
+      "scans": [],
+      "note": "code-checkable but no off-the-shelf tool digs it out — author a detector (custom semgrep rule / check)",
+      "status": "needs_tooling"
+    },
+    {
+      "control": "cra-ai-4",
+      "title": "Least-Privilege-Prinzip",
+      "scans": [],
+      "note": "code-checkable but no off-the-shelf tool digs it out — author a detector (custom semgrep rule / check)",
+      "status": "needs_tooling"
+    },
+    {
+      "control": "cra-ai-5",
+      "title": "Manipulationsschutz",
+      "scans": [],
+      "note": "code-checkable but no off-the-shelf tool digs it out — author a detector (custom semgrep rule / check)",
+      "status": "needs_tooling"
+    },
+    {
+      "control": "cra-ai-6",
+      "title": "Integritaetspruefung",
+      "scans": [],
+      "note": "code-checkable but no off-the-shelf tool digs it out — author a detector (custom semgrep rule / check)",
+      "status": "needs_tooling"
+    },
+    {
+      "control": "cra-ai-7",
+      "title": "Starke Authentifizierung",
+      "scans": [],
+      "note": "code-checkable but no off-the-shelf tool digs it out — author a detector (custom semgrep rule / check)",
+      "status": "needs_tooling"
+    },
+    {
+      "control": "cra-ai-8",
+      "title": "Keine Default-Passwoerter",
+      "scans": [
+        {
+          "tool": "gitleaks",
+          "scan_type": "secret_detection",
+          "cwe": [],
+          "rules": []
+        },
+        {
+          "tool": "semgrep",
+          "scan_type": "sast",
+          "cwe": [
+            "CWE-798",
+            "CWE-259"
+          ],
+          "rules": []
+        }
+      ],
+      "note": null,
+      "status": "covered"
+    },
+    {
+      "control": "cra-ai-9",
+      "title": "Sicheres Credential-Management",
+      "scans": [
+        {
+          "tool": "gitleaks",
+          "scan_type": "secret_detection",
+          "cwe": [],
+          "rules": []
+        },
+        {
+          "tool": "semgrep",
+          "scan_type": "sast",
+          "cwe": [
+            "CWE-798",
+            "CWE-522"
+          ],
+          "rules": []
+        }
+      ],
+      "note": null,
+      "status": "covered"
+    },
+    {
+      "control": "cra-ai-10",
+      "title": "Sitzungsmanagement",
+      "scans": [],
+      "note": "code-checkable but no off-the-shelf tool digs it out — author a detector (custom semgrep rule / check)",
+      "status": "needs_tooling"
+    },
+    {
+      "control": "cra-ai-11",
+      "title": "Brute-Force-Schutz",
+      "scans": [],
+      "note": "code-checkable but no off-the-shelf tool digs it out — author a detector (custom semgrep rule / check)",
+      "status": "needs_tooling"
+    },
+    {
+      "control": "cra-ai-12",
+      "title": "Rollenbasierte Autorisierung",
+      "scans": [],
+      "note": "code-checkable but no off-the-shelf tool digs it out — author a detector (custom semgrep rule / check)",
+      "status": "needs_tooling"
+    },
+    {
+      "control": "cra-ai-13",
+      "title": "Verschluesselung sensibler Daten",
+      "scans": [
+        {
+          "tool": "semgrep",
+          "scan_type": "sast",
+          "cwe": [
+            "CWE-327",
+            "CWE-326"
+          ],
+          "rules": []
+        }
+      ],
+      "note": null,
+      "status": "covered"
+    },
+    {
+      "control": "cra-ai-14",
+      "title": "Speicher-Schutz (Data at Rest)",
+      "scans": [],
+      "note": "code-checkable but no off-the-shelf tool digs it out — author a detector (custom semgrep rule / check)",
+      "status": "needs_tooling"
+    },
+    {
+      "control": "cra-ai-15",
+      "title": "Transport-Schutz (Data in Transit)",
+      "scans": [
+        {
+          "tool": "semgrep",
+          "scan_type": "sast",
+          "cwe": [
+            "CWE-319",
+            "CWE-311"
+          ],
+          "rules": []
+        }
+      ],
+      "note": null,
+      "status": "covered"
+    },
+    {
+      "control": "cra-ai-16",
+      "title": "Sicheres Schluesselmanagement",
+      "scans": [
+        {
+          "tool": "gitleaks",
+          "scan_type": "secret_detection",
+          "cwe": [],
+          "rules": []
+        },
+        {
+          "tool": "semgrep",
+          "scan_type": "sast",
+          "cwe": [
+            "CWE-798",
+            "CWE-321"
+          ],
+          "rules": []
+        }
+      ],
+      "note": null,
+      "status": "covered"
+    },
+    {
+      "control": "cra-ai-17",
+      "title": "Datenminimierung",
+      "scans": [],
+      "note": "process / document control — outside static-scan scope",
+      "status": "not_code_checkable"
+    },
+    {
+      "control": "cra-ai-18",
+      "title": "Strukturierter SSDLC",
+      "scans": [],
+      "note": "process / document control — outside static-scan scope",
+      "status": "not_code_checkable"
+    },
+    {
+      "control": "cra-ai-19",
+      "title": "Systematische Code Reviews",
+      "scans": [],
+      "note": "process / document control — outside static-scan scope",
+      "status": "not_code_checkable"
+    },
+    {
+      "control": "cra-ai-20",
+      "title": "Automatisierte Sicherheitstests",
+      "scans": [
+        {
+          "tool": "semgrep",
+          "scan_type": "sast",
+          "cwe": [
+            "CWE-89",
+            "CWE-78",
+            "CWE-79",
+            "CWE-22"
+          ],
+          "rules": []
+        }
+      ],
+      "note": null,
+      "status": "covered"
+    },
+    {
+      "control": "cra-ai-21",
+      "title": "Supply-Chain-Security",
+      "scans": [],
+      "note": "process / document control — outside static-scan scope",
+      "status": "not_code_checkable"
+    },
+    {
+      "control": "cra-ai-22",
+      "title": "Dependency-Monitoring",
+      "scans": [
+        {
+          "tool": "osv",
+          "scan_type": "cve",
+          "cwe": [],
+          "rules": []
+        },
+        {
+          "tool": "syft",
+          "scan_type": "sbom",
+          "cwe": [],
+          "rules": []
+        }
+      ],
+      "note": null,
+      "status": "covered"
+    },
+    {
+      "control": "cra-ai-23",
+      "title": "Software Bill of Materials (SBOM)",
+      "scans": [
+        {
+          "tool": "syft",
+          "scan_type": "sbom",
+          "cwe": [],
+          "rules": []
+        }
+      ],
+      "note": null,
+      "status": "covered"
+    },
+    {
+      "control": "cra-ai-24",
+      "title": "Security-Logging",
+      "scans": [],
+      "note": "code-checkable but no off-the-shelf tool digs it out — author a detector (custom semgrep rule / check)",
+      "status": "needs_tooling"
+    },
+    {
+      "control": "cra-ai-25",
+      "title": "Ereignis-Monitoring",
+      "scans": [],
+      "note": "process / document control — outside static-scan scope",
+      "status": "not_code_checkable"
+    },
+    {
+      "control": "cra-ai-26",
+      "title": "Anomalie-Erkennung",
+      "scans": [],
+      "note": "process / document control — outside static-scan scope",
+      "status": "not_code_checkable"
+    },
+    {
+      "control": "cra-ai-27",
+      "title": "Log-Integritaet und -Aufbewahrung",
+      "scans": [],
+      "note": "code-checkable but no off-the-shelf tool digs it out — author a detector (custom semgrep rule / check)",
+      "status": "needs_tooling"
+    },
+    {
+      "control": "cra-ai-28",
+      "title": "Sichere Update-Mechanismen",
+      "scans": [],
+      "note": "code-checkable but no off-the-shelf tool digs it out — author a detector (custom semgrep rule / check)",
+      "status": "needs_tooling"
+    },
+    {
+      "control": "cra-ai-29",
+      "title": "Update-Authentizitaet",
+      "scans": [],
+      "note": "code-checkable but no off-the-shelf tool digs it out — author a detector (custom semgrep rule / check)",
+      "status": "needs_tooling"
+    },
+    {
+      "control": "cra-ai-30",
+      "title": "Update-Integritaet",
+      "scans": [],
+      "note": "code-checkable but no off-the-shelf tool digs it out — author a detector (custom semgrep rule / check)",
+      "status": "needs_tooling"
+    },
+    {
+      "control": "cra-ai-31",
+      "title": "Lifecycle-Support",
+      "scans": [],
+      "note": "process / document control — outside static-scan scope",
+      "status": "not_code_checkable"
+    },
+    {
+      "control": "cra-ai-32",
+      "title": "Schwachstellen-Identifikation",
+      "scans": [],
+      "note": "process / document control — outside static-scan scope",
+      "status": "not_code_checkable"
+    },
+    {
+      "control": "cra-ai-33",
+      "title": "SBOM-Pflege und Analyse",
+      "scans": [
+        {
+          "tool": "syft",
+          "scan_type": "sbom",
+          "cwe": [],
+          "rules": []
+        },
+        {
+          "tool": "osv",
+          "scan_type": "cve",
+          "cwe": [],
+          "rules": []
+        }
+      ],
+      "note": null,
+      "status": "covered"
+    },
+    {
+      "control": "cra-ai-34",
+      "title": "Risikobasierte Priorisierung",
+      "scans": [],
+      "note": "process / document control — outside static-scan scope",
+      "status": "not_code_checkable"
+    },
+    {
+      "control": "cra-ai-35",
+      "title": "Coordinated Vulnerability Disclosure",
+      "scans": [],
+      "note": "process / document control — outside static-scan scope",
+      "status": "not_code_checkable"
+    },
+    {
+      "control": "cra-ai-36",
+      "title": "Incident-Response-Prozess",
+      "scans": [],
+      "note": "process / document control — outside static-scan scope",
+      "status": "not_code_checkable"
+    },
+    {
+      "control": "cra-ai-37",
+      "title": "Fruehwarnung (24h)",
+      "scans": [],
+      "note": "process / document control — outside static-scan scope",
+      "status": "not_code_checkable"
+    },
+    {
+      "control": "cra-ai-38",
+      "title": "Detaillierter Vorfallsbericht (72h)",
+      "scans": [],
+      "note": "process / document control — outside static-scan scope",
+      "status": "not_code_checkable"
+    },
+    {
+      "control": "cra-ai-39",
+      "title": "Patch-Bereitstellung",
+      "scans": [],
+      "note": "process / document control — outside static-scan scope",
+      "status": "not_code_checkable"
+    },
+    {
+      "control": "cra-ai-40",
+      "title": "Dokumentation und Nachbereitung",
+      "scans": [],
+      "note": "process / document control — outside static-scan scope",
+      "status": "not_code_checkable"
+    }
+  ]
+}

control-map/src/lib.rs

@@ -0,0 +1,163 @@
+//! `control-map` — the deterministic control → scan lookup table (LUT).
+//!
+//! The "transcribing" layer: it maps each compliance control to the static-scan
+//! step(s) that check it, or marks it as needing custom tooling, or as not
+//! code-checkable at all. The map is **authored and human-reviewed** — no LLM
+//! decides coverage. The LLM only enters later, downstream, to triage/ground the
+//! *tool's* findings (that lives in the agent, not here).
+//!
+//! This crate is intentionally tiny and standalone: types + an embedded JSON LUT
+//! + query helpers.
+
+use serde::{Deserialize, Serialize};
+
+/// Coverage bucket for a control under static (SAST-family) scanning.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum Coverage {
+    /// An existing tool's scan surfaces findings for this control.
+    Covered,
+    /// Code-checkable, but no existing tool digs it out — we must write tooling.
+    NeedsTooling,
+    /// Process / document control — out of static-scan scope.
+    NotCodeCheckable,
+}
+
+/// One tool binding: a scan step that (at least partially) checks a control.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ScanBinding {
+    /// Tool name, e.g. `"semgrep"`, `"gitleaks"`, `"syft"`, `"osv"`.
+    pub tool: String,
+    /// Scan family, e.g. `"sast"`, `"secret_detection"`, `"sbom"`, `"cve"`.
+    pub scan_type: String,
+    /// CWEs whose findings map to this control (used to attach findings back).
+    #[serde(default)]
+    pub cwe: Vec<String>,
+    /// Optional specific rule ids this control keys on.
+    #[serde(default)]
+    pub rules: Vec<String>,
+}
+
+/// One control's entry in the LUT.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ControlEntry {
+    /// Control id, e.g. `"cra-ai-8"`.
+    pub control: String,
+    /// Human-readable title (for the reviewable view).
+    #[serde(default)]
+    pub title: String,
+    /// Coverage bucket.
+    pub status: Coverage,
+    /// Tool bindings (empty unless `status == Covered`).
+    #[serde(default)]
+    pub scans: Vec<ScanBinding>,
+    /// Reviewer note — why it needs tooling / isn't code-checkable.
+    #[serde(default)]
+    pub note: Option<String>,
+}
+
+/// The control → scan lookup table for one framework.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ControlMap {
+    pub version: String,
+    pub framework: String,
+    pub controls: Vec<ControlEntry>,
+}
+
+const CRA_MAP_JSON: &str = include_str!("../data/cra_control_map.json");
+
+impl ControlMap {
+    /// Load the built-in CRA control map (the embedded, authored LUT).
+    pub fn cra() -> Result<Self, MapError> {
+        Ok(serde_json::from_str(CRA_MAP_JSON)?)
+    }
+
+    /// The coverage entry for a control id, if present.
+    pub fn coverage(&self, control_id: &str) -> Option<&ControlEntry> {
+        self.controls.iter().find(|c| c.control == control_id)
+    }
+
+    /// Controls whose bindings include the given `tool` + `cwe` — used to attach a
+    /// raw tool finding back to the control(s) it's evidence for.
+    pub fn controls_for(&self, tool: &str, cwe: &str) -> Vec<&ControlEntry> {
+        self.controls
+            .iter()
+            .filter(|c| {
+                c.scans
+                    .iter()
+                    .any(|s| s.tool == tool && s.cwe.iter().any(|w| w == cwe))
+            })
+            .collect()
+    }
+
+    /// Count of controls in each coverage bucket.
+    pub fn summary(&self) -> CoverageSummary {
+        let mut s = CoverageSummary::default();
+        for c in &self.controls {
+            match c.status {
+                Coverage::Covered => s.covered += 1,
+                Coverage::NeedsTooling => s.needs_tooling += 1,
+                Coverage::NotCodeCheckable => s.not_code_checkable += 1,
+            }
+        }
+        s
+    }
+}
+
+/// Coverage bucket counts.
+#[derive(Debug, Default, Clone, Copy, PartialEq, Eq)]
+pub struct CoverageSummary {
+    pub covered: usize,
+    pub needs_tooling: usize,
+    pub not_code_checkable: usize,
+}
+
+impl CoverageSummary {
+    pub fn total(&self) -> usize {
+        self.covered + self.needs_tooling + self.not_code_checkable
+    }
+}
+
+/// Errors loading a control map.
+#[derive(Debug, thiserror::Error)]
+pub enum MapError {
+    #[error("failed to parse control map: {0}")]
+    Parse(#[from] serde_json::Error),
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn cra_map_loads_all_40_controls() {
+        let map = ControlMap::cra().expect("CRA map should parse");
+        assert_eq!(map.framework, "cra");
+        assert_eq!(map.controls.len(), 40);
+        assert_eq!(map.summary().total(), 40);
+    }
+
+    #[test]
+    fn hardcoded_password_control_is_tool_covered() {
+        let map = ControlMap::cra().unwrap();
+        let c = map.coverage("cra-ai-8").expect("cra-ai-8 present");
+        assert_eq!(c.status, Coverage::Covered);
+        assert!(c.scans.iter().any(|s| s.tool == "semgrep"));
+        assert!(c.scans.iter().any(|s| s.tool == "gitleaks"));
+    }
+
+    #[test]
+    fn finding_attaches_back_to_control_via_tool_and_cwe() {
+        let map = ControlMap::cra().unwrap();
+        let hits = map.controls_for("semgrep", "CWE-798");
+        assert!(hits.iter().any(|c| c.control == "cra-ai-8"));
+    }
+
+    #[test]
+    fn every_bucket_is_represented() {
+        let s = ControlMap::cra().unwrap().summary();
+        assert!(s.covered > 0);
+        assert!(s.needs_tooling > 0);
+        assert!(s.not_code_checkable > 0);
+    }
+}