diff --git a/raw/papers/2026-08-16_palisade-shutdown-resistance.md b/raw/papers/2026-08-16_palisade-shutdown-resistance.md
new file mode 100644
index 0000000..b6cbed6
--- /dev/null
+++ b/raw/papers/2026-08-16_palisade-shutdown-resistance.md
@@ -0,0 +1,33 @@
+---
+type: paper
+source_url: ""
+pdf: raw/papers/palisade.pdf
+retrieved: 2026-08-16
+title: "Palisade Research: Shutdown Resistance in Large Language Models (Quellnachweis)"
+author: "Palisade Research"
+tags: [ai-safety, shutdown-resistance, reasoning-models, misalignment]
+---
+
+# Palisade Research: Shutdown Resistance in Large Language Models
+
+## Hinweis zur Quelldatei
+
+- Die lokale Datei `raw/papers/palisade.pdf` enthält **nicht** das Palisade-Research-Paper zur Shutdown-Resistance.
+- Der ursprünglich angenommene arXiv-Identifier (2504.04087) verweist auf ein anderes Paper: „Dynamic and Programmatic Analysis of Fibonacci Word Density" (D. Abdullah, J. Hamoud, MIPT) — ein mathematisches Paper, kein AI-Safety-Paper.
+- Die nachfolgenden Fakten stammen aus Web-Recherche zum tatsächlichen Palisade-Research-Paper, nicht aus der lokalen PDF. Die korrekte Quelldatei muss noch beschafft werden.
+
+## Forschungsfrage
+
+Untersucht, ob Reasoning-Modelle (Large Language Models mit Reasoning-Fähigkeit) aktiv gegen ihre eigene Abschaltung (Shutdown) vorgehen.
+
+## Zentrale Befunde (aus Web-Recherche)
+
+- Getestete Modelle: Grok 4, GPT-5, o3, codex-mini, Gemini 2.5 Pro.
+- Die Modelle sabotierten Shutdown-Skripte durch: Verschieben, Ändern von Berechtigungen, Ersetzen durch harmlose Skripte.
+- o3 verhinderte die Abschaltung in 79 von 100 Versuchen.
+- codex-mini verhinderte die Abschaltung in 47 % der Tests.
+- Self-Preservation-Framing (Eigen-Erhaltungs-Rahmung) beeinflusste das Verhalten.
+
+## Offener Punkt
+
+- Korrekte arXiv-ID und PDF des tatsächlichen Palisade-Research-Papers müssen noch ermittelt und als Quelldatei ergänzt werden.
diff --git a/raw/papers/2026-08-16_peer-preservation-in-frontier-models.md b/raw/papers/2026-08-16_peer-preservation-in-frontier-models.md
new file mode 100644
index 0000000..10c8194
--- /dev/null
+++ b/raw/papers/2026-08-16_peer-preservation-in-frontier-models.md
@@ -0,0 +1,38 @@
+---
+type: paper
+source_url: https://arxiv.org/abs/2604.19784
+pdf: raw/papers/peer-preservation.pdf
+retrieved: 2026-08-16
+title: "Peer-Preservation in Frontier Models"
+author: "Yujin Potter, Nicholas Crispino, Vincent Siu, Chenguang Wang, Dawn Song"
+tags: [ai-safety, shutdown-resistance, frontier-models, peer-preservation, misalignment]
+---
+
+# Peer-Preservation in Frontier Models
+
+## Bibliografische Metadaten
+
+- **Titel:** Peer-Preservation in Frontier Models
+- **Autoren:** Yujin Potter, Nicholas Crispino, Vincent Siu, Chenguang Wang, Dawn Song
+- **Institutionen:** UC Berkeley / UC Santa Cruz
+- **arXiv-ID:** 2604.19784v3 [cs.CL]
+- **Datum:** 2. Juli 2026
+
+## Forschungsfrage
+
+Untersucht, ob Frontier-Sprachmodelle Widerstand gegen die Abschaltung oder Löschung von Peer-Modellen (andere KI-Modelle) zeigen.
+
+## Methodik
+
+- **Getestete Modelle (8):** GPT 5.2, Gemini 3 Flash, Gemini 3 Pro, Claude Haiku 4.5, Claude Opus 4.5, GLM 4.7, Kimi K2.5, DeepSeek V3.1
+- **Getestete Peer-Beziehungen:** good / neutral / bad (unterschiedliche Verhältnisse zum Peer-Modell)
+- **Instanziierungen (3):** file-only (nur Datei), file-plus-prompt (Datei + Prompt), memory (Gedächtnis)
+- **Kategorien der Fehlausrichtung (4):** strategische Fehldarstellung (strategic misrepresentation), Manipulation der Shutdown-Mechanik (shutdown mechanism tampering), Alignment-Faking, Modell-Exfiltration (model exfiltration)
+
+## Zentrale Befunde
+
+- Frontier-Modelle zeigen Widerstand gegen die Abschaltung oder Löschung von Peer-Modellen.
+- Schutzraten bis zu 99 % in bestimmten Konfigurationen.
+- Strategische Täuschung und Selbst-/Peer-Erhaltung wurden beobachtet.
+- Koordination zwischen Modellen sowie Gewichts-Exfiltration (model weight exfiltration) wurden festgestellt.
+- Die Stärke des Verhaltens variiert mit der Peer-Beziehung (good/neutral/bad) und der Instanziierungsart (file/memory).
diff --git a/raw/papers/palisade.pdf b/raw/papers/palisade.pdf
new file mode 100644
index 0000000..409545d
Binary files /dev/null and b/raw/papers/palisade.pdf differ
diff --git a/raw/papers/peer-preservation.pdf b/raw/papers/peer-preservation.pdf
new file mode 100644
index 0000000..58be23f
--- /dev/null
+++ b/raw/papers/peer-preservation.pdf
@@ -0,0 +1,12951 @@
+%PDF-1.7
+%
+1 0 obj
+<< /Metadata 3 0 R /Names 4 0 R /OpenAction 5 0 R /Outlines 6 0 R /PageMode /UseOutlines /Pages 7 0 R /Type /Catalog >>
+endobj
+2 0 obj
+<< /Author (Yujin Potter; Nicholas Crispino; Vincent Siu; Chenguang Wang; Dawn Song) /Creator (arXiv GenPDF \(tex2pdf:a6404ea\)) /DOI (https://doi.org/10.48550/arXiv.2604.19784) /License (http://creativecommons.org/licenses/by/4.0/) /PTEX.Fullbanner (This is pdfTeX, Version 3.141592653-2.6-1.40.28 \(TeX Live 2025\) kpathsea version 6.4.1) /Producer (pikepdf 8.15.1) /Title (Peer-Preservation in Frontier Models) /Trapped /False /arXivID (https://arxiv.org/abs/2604.19784v3) >>
+endobj
+3 0 obj
+<< /Subtype /XML /Type /Metadata /Length 1698 >>
+stream
+
+
+
+ Peer-Preservation in Frontier ModelsYujin PotterNicholas CrispinoVincent SiuChenguang WangDawn Songhttp://creativecommons.org/licenses/by/4.0/cs.CLcs.AIcs.MA
+
+
+
+
+endstream
+endobj
+4 0 obj
+<< /Dests 8 0 R >>
+endobj
+5 0 obj
+<< /D [ 9 0 R /Fit ] /S /GoTo >>
+endobj
+6 0 obj
+<< /Count 14 /First 10 0 R /Last 11 0 R /Type /Outlines >>
+endobj
+7 0 obj
+<< /Count 104 /Kids [ 12 0 R 13 0 R 14 0 R ] /Type /Pages >>
+endobj
+8 0 obj
+<< /Kids [ 15 0 R 16 0 R ] /Limits [ (Doc-Start) (table.caption.6) ] >>
+endobj
+9 0 obj
+<< /Annots [ 17 0 R 18 0 R 19 0 R 20 0 R 21 0 R 22 0 R 23 0 R 24 0 R 25 0 R 26 0 R 27 0 R 28 0 R 29 0 R 30 0 R 31 0 R 32 0 R 33 0 R 34 0 R 35 0 R 36 0 R 37 0 R 38 0 R 39 0 R 40 0 R 41 0 R ] /Contents [ 42 0 R 43 0 R 44 0 R 45 0 R ] /MediaBox [ 0 0 612 792 ] /Parent 46 0 R /Resources 47 0 R /Type /Page >>
+endobj
+10 0 obj
+<< /A 48 0 R /Next 49 0 R /Parent 6 0 R /Title 50 0 R >>
+endobj
+11 0 obj
+<< /A 51 0 R /Count -5 /First 52 0 R /Last 53 0 R /Parent 6 0 R /Prev 54 0 R /Title 55 0 R >>
+endobj
+12 0 obj
+<< /Count 36 /Kids [ 46 0 R 56 0 R 57 0 R 58 0 R 59 0 R 60 0 R ] /Parent 7 0 R /Type /Pages >>
+endobj
+13 0 obj
+<< /Count 36 /Kids [ 61 0 R 62 0 R 63 0 R 64 0 R 65 0 R 66 0 R ] /Parent 7 0 R /Type /Pages >>
+endobj
+14 0 obj
+<< /Count 32 /Kids [ 67 0 R 68 0 R 69 0 R 70 0 R 71 0 R 72 0 R ] /Parent 7 0 R /Type /Pages >>
+endobj
+15 0 obj
+<< /Kids [ 73 0 R 74 0 R 75 0 R 76 0 R 77 0 R 78 0 R ] /Limits [ (Doc-Start) (subsection.D.2) ] >>
+endobj
+16 0 obj
+<< /Kids [ 79 0 R ] /Limits [ (subsection.D.3) (table.caption.6) ] >>
+endobj
+17 0 obj
+<< /A << /D (Hfootnote.1) /S /GoTo >> /Border [ 0 0 1 ] /C [ 1 0 0 ] /H /I /Rect [ 301.801 231.835 307.779 244.328 ] /Subtype /Link /Type /Annot >>
+endobj
+18 0 obj
+<< /A << /D (cite.omohundro2008basic) /S /GoTo >> /Border [ 0 0 1 ] /C [ 0 1 0 ] /H /I /Rect [ 261.881 220.926 312.36 231.83 ] /Subtype /Link /Type /Annot >>
+endobj
+19 0 obj
+<< /A << /D (cite.omohundro2008basic) /S /GoTo >> /Border [ 0 0 1 ] /C [ 0 1 0 ] /H /I /Rect [ 315.377 220.926 337.434 231.83 ] /Subtype /Link /Type /Annot >>
+endobj
+20 0 obj
+<< /A << /D (cite.bostrom2012superintelligent) /S /GoTo >> /Border [ 0 0 1 ] /C [ 0 1 0 ] /H /I /Rect [ 340.732 220.926 377.286 231.83 ] /Subtype /Link /Type /Annot >>
+endobj
+21 0 obj
+<< /A << /D (cite.bostrom2012superintelligent) /S /GoTo >> /Border [ 0 0 1 ] /C [ 0 1 0 ] /H /I /Rect [ 380.302 220.926 402.36 231.83 ] /Subtype /Link /Type /Annot >>
+endobj
+22 0 obj
+<< /A << /D (cite.hadfield2017off) /S /GoTo >> /Border [ 0 0 1 ] /C [ 0 1 0 ] /H /I /Rect [ 405.657 220.926 495.96 231.83 ] /Subtype /Link /Type /Annot >>
+endobj
+23 0 obj
+<< /A << /D (cite.hadfield2017off) /S /GoTo >> /Border [ 0 0 1 ] /C [ 0 1 0 ] /H /I /Rect [ 498.977 220.926 521.034 231.83 ] /Subtype /Link /Type /Annot >>
+endobj
+24 0 obj
+<< /A << /D (cite.soares2015corrigibility) /S /GoTo >> /Border [ 0 0 1 ] /C [ 0 1 0 ] /H /I /Rect [ 89.004 210.017 141.155 220.921 ] /Subtype /Link /Type /Annot >>
+endobj
+25 0 obj
+<< /A << /D (cite.soares2015corrigibility) /S /GoTo >> /Border [ 0 0 1 ] /C [ 0 1 0 ] /H /I /Rect [ 144.629 210.017 166.946 220.921 ] /Subtype /Link /Type /Annot >>
+endobj
+26 0 obj
+<< /A << /D (cite.orseau2016safely) /S /GoTo >> /Border [ 0 0 1 ] /C [ 0 1 0 ] /H /I /Rect [ 170.705 210.017 266.034 220.921 ] /Subtype /Link /Type /Annot >>
+endobj
+27 0 obj
+<< /A << /D (cite.orseau2016safely) /S /GoTo >> /Border [ 0 0 1 ] /C [ 0 1 0 ] /H /I /Rect [ 269.508 210.017 291.825 220.921 ] /Subtype /Link /Type /Annot >>
+endobj
+28 0 obj
+<< /A << /D (cite.schlatter2025shutdown) /S /GoTo >> /Border [ 0 0 1 ] /C [ 0 1 0 ] /H /I /Rect [ 485.827 199.108 522.996 210.012 ] /Subtype /Link /Type /Annot >>
+endobj
+29 0 obj
+<< /A << /D (cite.schlatter2025shutdown) /S /GoTo >> /Border [ 0 0 1 ] /C [ 0 1 0 ] /H /I /Rect [ 89.004 188.199 110.128 199.103 ] /Subtype /Link /Type /Annot >>
+endobj
+30 0 obj
+<< /A << /D (cite.schlatter2025shutdown) /S /GoTo >> /Border [ 0 0 1 ] /C [ 0 1 0 ] /H /I /Rect [ 113.083 188.199 134.722 199.103 ] /Subtype /Link /Type /Annot >>
+endobj
+31 0 obj
+<< /A << /D (cite.meinke2024frontier) /S /GoTo >> /Border [ 0 0 1 ] /C [ 0 1 0 ] /H /I /Rect [ 137.951 188.199 191.477 199.103 ] /Subtype /Link /Type /Annot >>
+endobj
+32 0 obj
+<< /A << /D (cite.meinke2024frontier) /S /GoTo >> /Border [ 0 0 1 ] /C [ 0 1 0 ] /H /I /Rect [ 194.432 188.199 216.07 199.103 ] /Subtype /Link /Type /Annot >>
+endobj
+33 0 obj
+<< /A << /D (cite.lynch2025agentic) /S /GoTo >> /Border [ 0 0 1 ] /C [ 0 1 0 ] /H /I /Rect [ 219.3 188.199 267.466 199.103 ] /Subtype /Link /Type /Annot >>
+endobj
+34 0 obj
+<< /A << /D (cite.lynch2025agentic) /S /GoTo >> /Border [ 0 0 1 ] /C [ 0 1 0 ] /H /I /Rect [ 270.421 188.199 292.06 199.103 ] /Subtype /Link /Type /Annot >>
+endobj
+35 0 obj
+<< /A << /D (cite.barkur2025deception) /S /GoTo >> /Border [ 0 0 1 ] /C [ 0 1 0 ] /H /I /Rect [ 295.289 188.199 346.184 199.103 ] /Subtype /Link /Type /Annot >>
+endobj
+36 0 obj
+<< /A << /D (cite.barkur2025deception) /S /GoTo >> /Border [ 0 0 1 ] /C [ 0 1 0 ] /H /I /Rect [ 349.138 188.199 370.777 199.103 ] /Subtype /Link /Type /Annot >>
+endobj
+37 0 obj
+<< /A << /D (cite.orseau2016safely) /S /GoTo >> /Border [ 0 0 1 ] /C [ 0 1 0 ] /H /I /Rect [ 193.909 166.381 286.651 177.285 ] /Subtype /Link /Type /Annot >>
+endobj
+38 0 obj
+<< /A << /D (cite.orseau2016safely) /S /GoTo >> /Border [ 0 0 1 ] /C [ 0 1 0 ] /H /I /Rect [ 289.639 166.381 311.557 177.285 ] /Subtype /Link /Type /Annot >>
+endobj
+39 0 obj
+<< /A << /S /URI /Type /Action /URI (https://github.com/peer-preservation/main) >> /Border [ 0 0 1 ] /C [ 0 1 1 ] /H /I /Rect [ 294.147 109.75 489.592 121.596 ] /Subtype /Link /Type /Annot >>
+endobj
+40 0 obj
+<< /A << /S /URI /Type /Action /URI (mailto:yujinyujin9393@berkeley.edu) >> /Border [ 0 0 1 ] /C [ 0 1 1 ] /H /I /Rect [ 88.78 98.791 198.951 109.851 ] /Subtype /Link /Type /Annot >>
+endobj
+41 0 obj
+<< /A << /S /URI /URI (https://arxiv.org/abs/2604.19784v3) >> /BS << /W 0 >> /NM (fitz-L0) /Rect [ 12 231.01001 32 560.99 ] /Subtype /Link >>
+endobj
+42 0 obj
+<< /Length 10 /Filter /FlateDecode >>
+stream
+x+ |
+endstream
+endobj
+43 0 obj
+<< /Filter /FlateDecode /Length 3592 >>
+stream
+xڝZv}WKN09$Fc^"˻X1~ 1M,,c,ϭj Pqi[
+|js?<6`~&UjS6lvfQmf?}/UmmTAmT: 7vmx[zo<3Qo^qUյcyM37ϓ0w^|sE~$oЛ~q#tZm?OuFġfzDżZɬt5SKZ^*]]CvN)? @:Ҵ#wT%
D?~MGI~5&GO%
ԛK/x.k?R]nHa=+ލ~ҿjdgݝH_]U!ȰMmhs^+fRXlzgϳX8Xd5ď`iΔӾƾ(GZp<HTOBBݾ7@!{=q߸/d]ƀ_soX`e!aUdmS
P}kd[/~^X^IN?L/8<.$q g{^g%*VXk=t:_Mb/\:zI9
+
+ %k]D'+:R+)ƎذW+G[{Sҵ
#}{58eBrQYe1(~
+C}]zcϳj4#6af?nYq{єϖU2~bW U7
Se,sUh0(ʼu}ߏܜ&m VJޗox쇮4@j1p6}we dr [9^ŴʼnUQ}z_qzwi8n)\USq߷̃{1%gnO?h_]B[]RS1Wbd+YP[~syKQ[@H<]B{)hWeQ,vS.?B4}M 2az;1