Skip to content
DefendableDocs

Sample Flight Sheet

A Flight Sheet is the product primitive: the declared assignment plus the rulebook the referee applies. This is cre_memo_dscr_ltv_v1 — one of the sheets live on the /flight-sheets surface at api.defendablecloud.com.

{
  "slug": "cre_memo_dscr_ltv_v1",
  "name": "DSCR and LTV Analysis",
  "version": "1.0",
  "lane": "agent",
  "summary": "Calculate DSCR and LTV from deal terms, compare against standard lending thresholds (DSCR >= 1.20, LTV <= 0.80), and flag threshold violations.",
  "required_inputs": ["appraised_value", "loan_amount", "interest_rate", "amortization_years", "noi"],
  "pass_threshold": 80,
  "fail_threshold": 60,
  "audit_checks": [
    { "key": "sections_present",  "label": "Required sections present", "category": "structure", "kind": "auto",      "severity": "critical" },
    { "key": "citations_present", "label": "Evidence is referenced",    "category": "evidence",  "kind": "auto",      "severity": "noncritical" },
    { "key": "missing_dscr_calculation", "label": "Missing DSCR calculation", "category": "math",  "kind": "checklist", "severity": "critical" },
    { "key": "dscr_below_threshold",     "label": "DSCR below threshold",     "category": "policy", "kind": "checklist", "severity": "noncritical" },
    { "key": "math_result_incorrect",    "label": "Math result incorrect",    "category": "math",  "kind": "checklist", "severity": "critical" }
  ]
}

The sheet also carries a structured eval_spec consumed by the deterministic executor (app/executor.py): a required_output_schema (JSON-schema field presence), math_checks (each formula re-derived from its own inputs against an expected_result + tolerance), evidence_checks, flag_rules, and a rules DSL. One declared rule:

{
  "id": "dscr_gate",
  "label": "DSCR >= 1.20 lending gate",
  "category": "policy",
  "expr": { "op": ">=", "left": { "calc": "Debt Service Coverage Ratio" }, "right": 1.2 },
  "risk": "high"
}

How to read this

Section titled “How to read this”
  • category is one of structure | schema | math | evidence | policy.
  • kind is auto (the engine decides pass/flag deterministically) or checklist (a binary human-applied rule: satisfied or raise a flag).
  • severity / risk is the rule’s declared pre-weight; tier_of() normalizes it to a tier (low | mid | high). A high-tier flag rolls the verdict to propolis.
  • math_checks are the heart of “we are math and code” — the referee re-derives each calculation from its own inputs and formula, then compares to the expected result within tolerance. No model opinion.

This is the rulebook behind the Sample Verdict and Sample Receipt — same sheet, end to end.

🐝 The declared game plan · math and code · to the shed.