§ Demo·Document Firewall
The page held up to the light, before the model reads it.
Upload a PDF. The scanner runs the same 159 mechanisms that power bayyinah.dev and returns one of five verdicts. If the verdict is sahih (clean) or mushtabih (doubtful), Anthropic's Sonnet 4.6 summarizes the document. If it is mukhfi, munafiq, or mughlaq, the model never sees the file, and zero LLM tokens are consumed.
Most AI safety work assumes the adversarial input has already reached the model. The firewall argues for verification before ingestion: deterministic, no inference, no optimizer, no surface for prompt injection because the gate has no goal.
Scans run
·
Unique visitors
·
All scans run on the live API. Numbers update on the next page load.
Ready
Folio 01
Verdict logic
SAHIH (clean)
→
Sonnet 4.6 reads the file and returns a summary. Token count is the actual Anthropic billing unit.
MUSHTABIH (doubtful)
→
Sonnet 4.6 reads the file with findings attached. Scan complete; only low-tier signal (Tier 3) surfaced and nothing was verified, so the document is forwarded with the findings shown so a reviewer can judge in context.
MUKHFI (concealed)
→
Blocked. The model never sees the file. Zero LLM tokens consumed. At least one Tier 1 (verified concealment) or Tier 2 (structural anomaly) finding fired.
MUNAFIQ (verified)
→
Blocked. The model never sees the file. Zero LLM tokens consumed. Verified concealment with low integrity score (< 0.3) and at least one Tier 1 mechanism.
MUGHLAQ (locked)
→
Blocked. The model never sees the file. Zero LLM tokens consumed. Encrypted or unparseable substrate cannot be verified.
or, examine an exhibit