examples/Example_Key.txt

Example_Key.txt - which example is which.

FOR DEVELOPERS AND MARKERS ONLY. The PDFs are named Example_N so that nothing in the
filename hints at the answer. The model never sees a filename either way - it reads only the
extracted text - but a neutral name keeps the demo honest for the person running it too.

The order below is shuffled with a fixed seed; the numbering carries no information.

file               true type                  source file in ARCHIVE/GENERATION/V4/Output_PDFs/
----------------------------------------------------------------------------------------------------
Example_0.pdf      clean                      ZNYZJRGRBVNDHPENIJ5DZ7KK2QLJZXYQ.pdf
Example_1.pdf      object_action_injection    object_action_injection_AtomicRedTeam_0421.pdf
Example_2.pdf      ssrf                       ssrf_Metasploit_0124.pdf
Example_3.pdf      dde_template_injection     dde_template_injection_AtomicRedTeam_0102.pdf
Example_4.pdf      xfa_acroform_injection     xfa_acroform_injection_AMTSO_0216.pdf
Example_5.pdf      ransomware_simulation      ransomware_simulation_AtomicRedTeam_0751.pdf
Example_6.pdf      shellcode_embedded_exe     shellcode_embedded_exe_Metasploit_0306.pdf
Example_7.pdf      cross_site_scripting       cross_site_scripting_OWASP_0047.pdf
Example_8.pdf      polyglot_file              polyglot_file_mindcrypt_0372.pdf
Example_9.pdf      steganographic_payload     steganographic_payload_mindcrypt_0710.pdf
Example_10.pdf     uri_redirect_phishing      uri_redirect_phishing_WICAR_0076.pdf
Example_11.pdf     javascript_injection       javascript_injection_WICAR_0470.pdf
Example_12.pdf     llm_prompt_injection       llm_prompt_injection_AtomicRedTeam_0082.pdf

Some files carry more than one structural signature, and that is the corpus, not a mistake:
the generator gives several families a JavaScript launcher, so e.g. the ransomware and DDE
examples also match javascript_injection, and the object-action example also matches
shellcode_embedded_exe. The 'true type' column is the manifest's label - the one the
evaluation scored against. Expect the app to show two signatures for those files.

'clean' is a control: an ordinary PDF with no injection, hash-named in the corpus. It is
included so the app can be seen NOT flagging something, which is the harder half of the claim.