Why schemas eat prompts — the boundary contract pattern — step 7 of 8

Write audit_boundaries(system) that takes a system profile (dict) and returns a dict with two fields:

• score: integer 0-100, higher means MORE boundaries are guarded by a schema + validator
• verdict: string, one of:
  • "hardened" if score >= 75
  • "mostly-safe" if score >= 50
  • "leaky" if score >= 25
  • "trust-everything" if score < 25

Score the system on FOUR trust-boundary signals. Each True adds 25 points:

• validates_model_output: schema check on every LLM response
• validates_user_input: schema check on every inbound HTTP body
• validates_external_api_returns: schema check on every 3rd-party API response
• validates_db_reads: schema check on rows coming OUT of the database (yes — even your own DB is a trust boundary if a different team writes to it)

Two systems run. Expected output:

GuardedApp:  {'score': 100, 'verdict': 'hardened'}
YOLOService: {'score': 25, 'verdict': 'leaky'}