diff --git a/maskinrommet/src/cas.rs b/maskinrommet/src/cas.rs
index a2461ce..2ab23fe 100644
--- a/maskinrommet/src/cas.rs
+++ b/maskinrommet/src/cas.rs
@@ -99,6 +99,65 @@ impl CasStore {
         hex::encode(hasher.finalize())
     }
 
+    /// Slett en fil fra CAS. Returnerer antall bytes frigitt, eller 0 hvis filen ikke fantes.
+    pub async fn delete(&self, hash: &str) -> std::io::Result<u64> {
+        let path = self.path_for(hash);
+        match tokio::fs::metadata(&path).await {
+            Ok(meta) => {
+                let size = meta.len();
+                tokio::fs::remove_file(&path).await?;
+                tracing::info!(hash = %hash, size = size, "Slettet fil fra CAS");
+                Ok(size)
+            }
+            Err(e) if e.kind() == std::io::ErrorKind::NotFound => Ok(0),
+            Err(e) => Err(e),
+        }
+    }
+
+    /// Beregn total diskbruk for CAS-katalogen (bytes).
+    /// Bruker `du -sb` for ytelse — traverserer filsystemet direkte.
+    pub async fn disk_usage_bytes(&self) -> std::io::Result<u64> {
+        let output = tokio::process::Command::new("du")
+            .args(["-sb", self.root.to_str().unwrap_or(".")])
+            .output()
+            .await?;
+        if !output.status.success() {
+            return Err(std::io::Error::new(
+                std::io::ErrorKind::Other,
+                "du-kommando feilet",
+            ));
+        }
+        let stdout = String::from_utf8_lossy(&output.stdout);
+        let bytes: u64 = stdout
+            .split_whitespace()
+            .next()
+            .and_then(|s| s.parse().ok())
+            .unwrap_or(0);
+        Ok(bytes)
+    }
+
+    /// Sjekk diskbruk-prosent for partisjonen CAS ligger på.
+    pub async fn disk_usage_percent(&self) -> std::io::Result<f64> {
+        let output = tokio::process::Command::new("df")
+            .args(["--output=pcent", self.root.to_str().unwrap_or("/")])
+            .output()
+            .await?;
+        if !output.status.success() {
+            return Err(std::io::Error::new(
+                std::io::ErrorKind::Other,
+                "df-kommando feilet",
+            ));
+        }
+        let stdout = String::from_utf8_lossy(&output.stdout);
+        // Output: "Use%\n 42%\n"
+        let pct: f64 = stdout
+            .lines()
+            .nth(1)
+            .and_then(|line| line.trim().trim_end_matches('%').parse().ok())
+            .unwrap_or(0.0);
+        Ok(pct)
+    }
+
     /// Rot-katalogen for CAS.
     pub fn root(&self) -> &Path {
         &self.root
diff --git a/maskinrommet/src/main.rs b/maskinrommet/src/main.rs
index 88de6d4..57a8149 100644
--- a/maskinrommet/src/main.rs
+++ b/maskinrommet/src/main.rs
@@ -5,6 +5,7 @@ pub mod cas;
 mod intentions;
 pub mod jobs;
 pub mod livekit;
+pub mod pruning;
 mod queries;
 mod serving;
 mod stdb;
@@ -130,6 +131,9 @@ async fn main() {
     // Start jobbkø-worker i bakgrunnen
     jobs::start_worker(db.clone(), stdb.clone(), cas.clone());
 
+    // Start periodisk CAS-pruning i bakgrunnen
+    pruning::start_pruning_loop(db.clone(), cas.clone());
+
     let state = AppState { db, jwks, stdb, cas };
 
     // Ruter: /health er offentlig, /me krever gyldig JWT
diff --git a/maskinrommet/src/pruning.rs b/maskinrommet/src/pruning.rs
new file mode 100644
index 0000000..71e6705
--- /dev/null
+++ b/maskinrommet/src/pruning.rs
@@ -0,0 +1,465 @@
+//! Pruning-logikk for CAS-filer.
+//!
+//! TTL per modalitet, signaler som forlenger levetid, og disk-nødventil.
+//! Ref: docs/retninger/maskinrommet.md § CAS og intelligent pruning
+//!
+//! ## TTL per modalitet
+//! | Modalitet     | Standard TTL | Begrunnelse                          |
+//! |---------------|-------------|--------------------------------------|
+//! | Tekst         | Aldri       | Billig, essensen av innhold          |
+//! | Transkripsjon | Aldri       | Tekstrepresentasjon bevarer mening   |
+//! | Lyd           | 30 dager    | Transkripsjon bevarer innholdet      |
+//! | Bilde         | 30 dager    | Beskrivelse/metadata bevarer kontekst|
+//! | Video         | 14 dager    | Dyrest, transkripsjon + thumbnail    |
+//!
+//! ## Signaler som forlenger levetid
+//! - Publishing-edge → behold for alltid
+//! - Tilgang (last_accessed_at) innenfor TTL → forleng med ny TTL-periode
+//! - Uredigert transkripsjon → utranskribert lyd beholdes
+//!
+//! ## Disk-nødventil
+//! | Terskel | Handling                                                 |
+//! |---------|----------------------------------------------------------|
+//! | >85%    | Slett generert innhold (TTS, thumbnails — regenererbart) |
+//! | >90%    | Aggressiv pruning for alle samlinger                     |
+//! | >95%    | Kritisk: alt uten publishing-edge slettes. Tekst beholdes|
+
+use chrono::{DateTime, Utc};
+use sqlx::PgPool;
+use uuid::Uuid;
+
+use crate::cas::CasStore;
+
+/// Konfigurasjon for pruning.
+#[derive(Debug, Clone)]
+pub struct PruningConfig {
+    /// Standard TTL for lyd i dager.
+    pub audio_ttl_days: i64,
+    /// Standard TTL for bilder i dager.
+    pub image_ttl_days: i64,
+    /// Standard TTL for video i dager.
+    pub video_ttl_days: i64,
+    /// Disk-terskel for å slette generert innhold (prosent).
+    pub disk_warning_pct: f64,
+    /// Disk-terskel for aggressiv pruning (prosent).
+    pub disk_aggressive_pct: f64,
+    /// Disk-terskel for kritisk alarm (prosent).
+    pub disk_critical_pct: f64,
+}
+
+impl Default for PruningConfig {
+    fn default() -> Self {
+        Self {
+            audio_ttl_days: 30,
+            image_ttl_days: 30,
+            video_ttl_days: 14,
+            disk_warning_pct: 85.0,
+            disk_aggressive_pct: 90.0,
+            disk_critical_pct: 95.0,
+        }
+    }
+}
+
+/// En CAS-fil-kandidat for pruning.
+#[derive(Debug, sqlx::FromRow)]
+#[allow(dead_code)]
+struct PruneCandidate {
+    id: Uuid,
+    cas_hash: String,
+    mime_category: String,
+    size_bytes: i64,
+    created_at: DateTime<Utc>,
+    last_accessed_at: Option<DateTime<Utc>>,
+    has_publishing_edge: bool,
+    is_generated: bool,
+    has_transcription: bool,
+}
+
+/// Resultat fra en pruning-kjøring.
+#[derive(Debug, Default, serde::Serialize)]
+pub struct PruneResult {
+    pub candidates_checked: usize,
+    pub files_deleted: usize,
+    pub bytes_freed: u64,
+    pub disk_pct_before: f64,
+    pub disk_pct_after: f64,
+    pub emergency_level: &'static str,
+}
+
+/// Kjør pruning-logikk. Returnerer statistikk.
+pub async fn run_pruning(
+    db: &PgPool,
+    cas: &CasStore,
+    config: &PruningConfig,
+) -> Result<PruneResult, String> {
+    let disk_pct = cas.disk_usage_percent().await.map_err(|e| format!("Disk-sjekk feilet: {e}"))?;
+
+    let emergency_level = if disk_pct >= config.disk_critical_pct {
+        "critical"
+    } else if disk_pct >= config.disk_aggressive_pct {
+        "aggressive"
+    } else if disk_pct >= config.disk_warning_pct {
+        "warning"
+    } else {
+        "normal"
+    };
+
+    tracing::info!(
+        disk_pct = disk_pct,
+        level = emergency_level,
+        "Pruning startet"
+    );
+
+    let mut result = PruneResult {
+        disk_pct_before: disk_pct,
+        emergency_level,
+        ..Default::default()
+    };
+
+    // Fase 1: Slett generert innhold ved disk ≥ warning (85%)
+    if disk_pct >= config.disk_warning_pct {
+        let freed = prune_generated(db, cas).await?;
+        result.files_deleted += freed.0;
+        result.bytes_freed += freed.1;
+        tracing::info!(
+            files = freed.0,
+            bytes = freed.1,
+            "Fase 1: Slettet generert innhold"
+        );
+    }
+
+    // Fase 2: Standard TTL-basert pruning (kjører alltid)
+    let ttl_freed = prune_by_ttl(db, cas, config, emergency_level).await?;
+    result.candidates_checked += ttl_freed.0;
+    result.files_deleted += ttl_freed.1;
+    result.bytes_freed += ttl_freed.2;
+
+    // Fase 3: Kritisk — slett alt uten publishing-edge (unntatt tekst)
+    if disk_pct >= config.disk_critical_pct {
+        let critical_freed = prune_critical(db, cas).await?;
+        result.files_deleted += critical_freed.0;
+        result.bytes_freed += critical_freed.1;
+        tracing::warn!(
+            files = critical_freed.0,
+            bytes = critical_freed.1,
+            "Fase 3: KRITISK pruning — alt uten publishing-edge slettet"
+        );
+    }
+
+    // Sjekk disk etter pruning
+    result.disk_pct_after = cas
+        .disk_usage_percent()
+        .await
+        .unwrap_or(disk_pct);
+
+    tracing::info!(
+        files_deleted = result.files_deleted,
+        bytes_freed = result.bytes_freed,
+        disk_before = format!("{:.1}%", result.disk_pct_before),
+        disk_after = format!("{:.1}%", result.disk_pct_after),
+        "Pruning fullført"
+    );
+
+    Ok(result)
+}
+
+/// Slett generert innhold (TTS, thumbnails osv.) — kan regenereres.
+/// Identifiseres ved metadata.tts eller metadata.generated = true.
+async fn prune_generated(db: &PgPool, cas: &CasStore) -> Result<(usize, u64), String> {
+    let rows: Vec<(String, Uuid)> = sqlx::query_as(
+        r#"
+        SELECT metadata->>'cas_hash' AS cas_hash, id
+        FROM nodes
+        WHERE node_kind = 'media'
+          AND metadata->>'cas_hash' IS NOT NULL
+          AND (
+            metadata ? 'tts'
+            OR metadata->>'generated' = 'true'
+          )
+        "#,
+    )
+    .fetch_all(db)
+    .await
+    .map_err(|e| format!("Spørring for generert innhold feilet: {e}"))?;
+
+    let mut deleted = 0usize;
+    let mut freed = 0u64;
+
+    for (hash, node_id) in &rows {
+        match cas.delete(hash).await {
+            Ok(bytes) if bytes > 0 => {
+                deleted += 1;
+                freed += bytes;
+                log_prune(db, *node_id, hash, bytes, "generated_cleanup").await;
+            }
+            Ok(_) => {} // Allerede borte
+            Err(e) => {
+                tracing::warn!(hash = %hash, error = %e, "Kunne ikke slette generert fil");
+            }
+        }
+    }
+
+    Ok((deleted, freed))
+}
+
+/// TTL-basert pruning: slett CAS-filer som har utløpt basert på modalitet.
+/// Ved aggressive/critical senkes TTL drastisk.
+async fn prune_by_ttl(
+    db: &PgPool,
+    cas: &CasStore,
+    config: &PruningConfig,
+    emergency_level: &str,
+) -> Result<(usize, usize, u64), String> {
+    // Hent kandidater: media-noder med CAS-hash, MIME, alder, edges, tilgangstid
+    let candidates: Vec<PruneCandidate> = sqlx::query_as(
+        r#"
+        SELECT
+            n.id,
+            n.metadata->>'cas_hash' AS cas_hash,
+            CASE
+                WHEN n.metadata->>'mime' LIKE 'audio/%' THEN 'audio'
+                WHEN n.metadata->>'mime' LIKE 'image/%' THEN 'image'
+                WHEN n.metadata->>'mime' LIKE 'video/%' THEN 'video'
+                ELSE 'other'
+            END AS mime_category,
+            COALESCE((n.metadata->>'size_bytes')::bigint, 0) AS size_bytes,
+            n.created_at,
+            n.last_accessed_at,
+            EXISTS(
+                SELECT 1 FROM edges e
+                WHERE (e.source_id = n.id OR e.target_id = n.id)
+                  AND e.edge_type IN ('belongs_to', 'has_media')
+                  AND EXISTS(
+                    SELECT 1 FROM edges pub
+                    WHERE pub.edge_type = 'publishing'
+                      AND (pub.source_id = e.target_id OR pub.source_id = e.source_id)
+                  )
+            ) AS has_publishing_edge,
+            COALESCE(n.metadata ? 'tts' OR n.metadata->>'generated' = 'true', false) AS is_generated,
+            EXISTS(
+                SELECT 1 FROM transcription_segments ts
+                WHERE ts.node_id = n.id
+            ) AS has_transcription
+        FROM nodes n
+        WHERE n.node_kind = 'media'
+          AND n.metadata->>'cas_hash' IS NOT NULL
+        ORDER BY n.created_at ASC
+        "#,
+    )
+    .fetch_all(db)
+    .await
+    .map_err(|e| format!("TTL-kandidatspørring feilet: {e}"))?;
+
+    let now = Utc::now();
+    let mut checked = 0usize;
+    let mut deleted = 0usize;
+    let mut freed = 0u64;
+
+    for c in &candidates {
+        checked += 1;
+
+        // Tekst/transkripsjon slettes aldri
+        if c.mime_category == "other" {
+            continue;
+        }
+
+        // Publishing-edge = behold for alltid (unntatt ved kritisk)
+        if c.has_publishing_edge && emergency_level != "critical" {
+            continue;
+        }
+
+        // Bestem TTL basert på modalitet og nødsituasjon
+        let ttl_days = match emergency_level {
+            "critical" => 0, // Slett alt umiddelbart
+            "aggressive" => match c.mime_category.as_str() {
+                "audio" => config.audio_ttl_days / 3, // 10 dager
+                "image" => config.image_ttl_days / 3,
+                "video" => 1, // Video slettes nesten umiddelbart
+                _ => i64::MAX,
+            },
+            _ => match c.mime_category.as_str() {
+                "audio" => config.audio_ttl_days,
+                "image" => config.image_ttl_days,
+                "video" => config.video_ttl_days,
+                _ => i64::MAX,
+            },
+        };
+
+        if ttl_days == i64::MAX {
+            continue;
+        }
+
+        // Beregn effektiv alder — siste tilgang forlenger levetiden
+        let reference_time = c.last_accessed_at.unwrap_or(c.created_at);
+        let age_days = (now - reference_time).num_days();
+
+        if age_days < ttl_days {
+            continue; // Ikke utløpt ennå
+        }
+
+        // Lyd uten transkripsjon: behold lengre (trenger transkribering først)
+        if c.mime_category == "audio" && !c.has_transcription && emergency_level == "normal" {
+            tracing::debug!(
+                node_id = %c.id,
+                "Beholder utranskribert lyd — trenger transkribering"
+            );
+            continue;
+        }
+
+        // Allerede slettet generert innhold i fase 1 — hopp over
+        if c.is_generated {
+            continue;
+        }
+
+        // Slett filen
+        match cas.delete(&c.cas_hash).await {
+            Ok(bytes) if bytes > 0 => {
+                deleted += 1;
+                freed += bytes;
+                log_prune(db, c.id, &c.cas_hash, bytes, "ttl_expired").await;
+            }
+            Ok(_) => {} // Allerede borte
+            Err(e) => {
+                tracing::warn!(
+                    hash = %c.cas_hash,
+                    node_id = %c.id,
+                    error = %e,
+                    "Kunne ikke slette CAS-fil"
+                );
+            }
+        }
+    }
+
+    Ok((checked, deleted, freed))
+}
+
+/// Kritisk pruning: slett ALT uten publishing-edge (unntatt tekst/transkripsjon).
+async fn prune_critical(db: &PgPool, cas: &CasStore) -> Result<(usize, u64), String> {
+    let rows: Vec<(String, Uuid, i64)> = sqlx::query_as(
+        r#"
+        SELECT
+            n.metadata->>'cas_hash' AS cas_hash,
+            n.id,
+            COALESCE((n.metadata->>'size_bytes')::bigint, 0) AS size_bytes
+        FROM nodes n
+        WHERE n.node_kind = 'media'
+          AND n.metadata->>'cas_hash' IS NOT NULL
+          AND NOT EXISTS(
+            SELECT 1 FROM edges e
+            WHERE (e.source_id = n.id OR e.target_id = n.id)
+              AND e.edge_type IN ('belongs_to', 'has_media')
+              AND EXISTS(
+                SELECT 1 FROM edges pub
+                WHERE pub.edge_type = 'publishing'
+                  AND (pub.source_id = e.target_id OR pub.source_id = e.source_id)
+              )
+          )
+        "#,
+    )
+    .fetch_all(db)
+    .await
+    .map_err(|e| format!("Kritisk pruning-spørring feilet: {e}"))?;
+
+    let mut deleted = 0usize;
+    let mut freed = 0u64;
+
+    for (hash, node_id, _size) in &rows {
+        match cas.delete(hash).await {
+            Ok(bytes) if bytes > 0 => {
+                deleted += 1;
+                freed += bytes;
+                log_prune(db, *node_id, hash, bytes, "critical_emergency").await;
+            }
+            Ok(_) => {}
+            Err(e) => {
+                tracing::warn!(hash = %hash, error = %e, "Kritisk: kunne ikke slette CAS-fil");
+            }
+        }
+    }
+
+    Ok((deleted, freed))
+}
+
+/// Logg en pruning-hendelse til resource_usage_log.
+async fn log_prune(db: &PgPool, node_id: Uuid, hash: &str, bytes: u64, reason: &str) {
+    let detail = serde_json::json!({
+        "hash": hash,
+        "size_bytes": bytes,
+        "operation": "delete",
+        "reason": reason,
+    });
+
+    let _ = sqlx::query(
+        r#"
+        INSERT INTO resource_usage_log (target_node_id, resource_type, detail)
+        VALUES ($1, 'cas', $2)
+        "#,
+    )
+    .bind(node_id)
+    .bind(&detail)
+    .execute(db)
+    .await
+    .map_err(|e| {
+        tracing::warn!(error = %e, "Kunne ikke logge pruning-hendelse");
+    });
+}
+
+/// Oppdater last_accessed_at for en node basert på CAS-hash.
+/// Kalles fra serving.rs når en CAS-fil hentes.
+pub async fn touch_access(db: &PgPool, cas_hash: &str) {
+    let _ = sqlx::query(
+        "UPDATE nodes SET last_accessed_at = now() WHERE metadata->>'cas_hash' = $1",
+    )
+    .bind(cas_hash)
+    .execute(db)
+    .await
+    .map_err(|e| {
+        tracing::debug!(error = %e, "Kunne ikke oppdatere last_accessed_at");
+    });
+}
+
+/// Start periodisk pruning-loop som bakgrunnsoppgave.
+/// Kjører hvert 6. time, eller oftere ved høy diskbruk.
+pub fn start_pruning_loop(db: PgPool, cas: CasStore) {
+    let config = PruningConfig::default();
+
+    tokio::spawn(async move {
+        // Vent 60 sekunder etter oppstart før første kjøring
+        tokio::time::sleep(std::time::Duration::from_secs(60)).await;
+        tracing::info!("Pruning-loop startet (intervall: 6t, nødsjekk: 10min)");
+
+        loop {
+            match run_pruning(&db, &cas, &config).await {
+                Ok(result) => {
+                    if result.files_deleted > 0 {
+                        tracing::info!(
+                            deleted = result.files_deleted,
+                            freed_mb = result.bytes_freed / 1_048_576,
+                            "Pruning: {} filer slettet, {} MB frigitt",
+                            result.files_deleted,
+                            result.bytes_freed / 1_048_576,
+                        );
+                    }
+
+                    // Ved høy diskbruk: sjekk igjen om 10 minutter
+                    let sleep_secs = if result.disk_pct_after >= config.disk_warning_pct {
+                        tracing::warn!(
+                            disk_pct = result.disk_pct_after,
+                            "Disk fortsatt høy — neste pruning om 10 min"
+                        );
+                        600 // 10 minutter
+                    } else {
+                        6 * 3600 // 6 timer
+                    };
+
+                    tokio::time::sleep(std::time::Duration::from_secs(sleep_secs)).await;
+                }
+                Err(e) => {
+                    tracing::error!(error = %e, "Pruning feilet");
+                    // Vent 30 minutter ved feil
+                    tokio::time::sleep(std::time::Duration::from_secs(1800)).await;
+                }
+            }
+        }
+    });
+}
diff --git a/maskinrommet/src/serving.rs b/maskinrommet/src/serving.rs
index 350103b..7d48a6c 100644
--- a/maskinrommet/src/serving.rs
+++ b/maskinrommet/src/serving.rs
@@ -61,6 +61,13 @@ pub async fn get_cas_file(
         StatusCode::INTERNAL_SERVER_ERROR
     })?;
 
+    // Oppdater last_accessed_at asynkront (fire-and-forget for ytelse)
+    let db_clone = state.db.clone();
+    let hash_clone = hash.clone();
+    tokio::spawn(async move {
+        crate::pruning::touch_access(&db_clone, &hash_clone).await;
+    });
+
     let stream = ReaderStream::new(file);
     let body = Body::from_stream(stream);
 
diff --git a/migrations/010_pruning.sql b/migrations/010_pruning.sql
new file mode 100644
index 0000000..3417dc5
--- /dev/null
+++ b/migrations/010_pruning.sql
@@ -0,0 +1,16 @@
+-- 010_pruning.sql
+-- Støtte for pruning-logikk: TTL per modalitet, signaler som forlenger levetid.
+-- Ref: docs/retninger/maskinrommet.md § CAS og intelligent pruning
+
+BEGIN;
+
+-- last_accessed_at: sporer siste gang en CAS-fil ble servert.
+-- Brukes som signal for å forlenge TTL — "noen bruker denne filen aktivt".
+ALTER TABLE nodes ADD COLUMN IF NOT EXISTS last_accessed_at TIMESTAMPTZ;
+
+-- Indeks for pruning-spørringer: finn media-noder sortert etter alder/tilgang.
+CREATE INDEX IF NOT EXISTS idx_nodes_media_access
+    ON nodes (node_kind, last_accessed_at, created_at)
+    WHERE node_kind = 'media';
+
+COMMIT;
diff --git a/tasks.md b/tasks.md
index 5f39ce9..bdf9353 100644
--- a/tasks.md
+++ b/tasks.md
@@ -125,8 +125,7 @@ Uavhengige faser kan fortsatt plukkes.
 
 - [x] 11.1 LiveKit oppsett: Docker-container for WebRTC. Ref: `docs/setup/produksjon.md`.
 - [x] 11.2 Sanntidslyd: kommunikasjonsnode med live-status → LiveKit-rom for deltakere.
-- [~] 11.3 Pruning-logikk: TTL per modalitet, signaler som forlenger levetid, disk-nødventil.
-  > Påbegynt: 2026-03-17T23:55
+- [x] 11.3 Pruning-logikk: TTL per modalitet, signaler som forlenger levetid, disk-nødventil.
 - [ ] 11.4 Podcast-RSS: samlings-node med publiserings-edges → generert RSS-feed.
 
 ## Fase 13: Trait-system