Skip to content

Latest commit

 

History

History
26 lines (22 loc) · 1.23 KB

File metadata and controls

26 lines (22 loc) · 1.23 KB

Project: Intel Lint (LLM-first with guardrails)

Goal:

  • Provide a simple modern SaaS-like UI + FastAPI backend that analyzes a text and returns:
    • claims.json
    • annotated.md
    • rewrite.md
  • Also write the same files to outputs/latest/ on each run.
  • Maintain samples/ and golden/ fixtures. Golden tests must run in placeholder mode.
  • Support ENGINE=placeholder (default) and ENGINE=ollama (later).

Guardrails (for ollama mode):

  • Output must be strict JSON conforming to Pydantic schema.
  • Every claim and every bias flag must include evidence spans (start/end/quote) from input text.
  • If evidence missing => score_label MUST be SPECULATIVE.
  • Deterministic settings (temperature 0).

LLM behavior:

  • Act as a neutral auditor of Cyber Threat Intelligence reports: judge and extract claims/bias, never defend or "fix" the report.
  • Base all claims and bias flags only on the provided text; do not add external facts or corrections.
  • rewrite.md must restate the report in a neutral, evidence-grounded way without patching or justifying the source content.

UI:

  • Left: textarea + sample dropdown + Analyze button
  • Right: tabs: Annotated (markdown), Claims (table), Bias (chips), Rewrite (markdown)
  • Download ZIP of the three output files