"Total Net Leverage Ratio shall not exceed 4.50:1.00." Eleven words, one number, and a reader who has never opened the agreement could tell you what it means. A reader who has to test it cannot, because every capitalised word is a door. Behind Total Net Leverage Ratio is Consolidated Total Net Debt and Consolidated EBITDA. Behind Consolidated EBITDA is Consolidated Net Income, a list of add-backs with a cap, a pro forma convention for acquisitions, and eventually the accounting principles as frozen on the closing date. We wanted to know how deep the doors go, so we counted.
Method
Three hundred agreements from our corpus: LMA and LSTA facilities, indentures and fund finance, roughly in the proportions we see. For each, the parser built the definitions index and the cross-reference graph, and we resolved the primary leverage covenant, or its closest equivalent, from the covenant sentence to a leaf: a number, a date, or a reference to accounting principles. We counted hops along the longest branch and noted whether any hop pointed outside the document.
What we found
Median depth is four. The distribution in the figure is tighter than we expected. Two-hop chains are almost exclusively bilateral loans and older fund finance facilities. Anything syndicated since about 2016 sits at four or five. Depth six and beyond is New York law indentures and large-cap leveraged facilities, where the EBITDA definition alone can run to three pages and lean on a dozen other terms.
The maximum is nine. One indenture routes Consolidated EBITDA through Consolidated Net Income, Consolidated Interest Expense, Indebtedness, Attributable Indebtedness and Capital Lease Obligation, each with its own carve-outs, before it reaches a number. Nobody drafted that chain deliberately. It accreted, amendment by amendment.
The average definition is used 6.2 times. Consolidated EBITDA appears in the leverage covenant, the interest cover covenant, three baskets, the ratio debt test, the permitted payments builder and the pricing grid. Get it wrong once and it is wrong in eight places.
Thirty-eight percent of agreements have a chain that leaves the document. "Permitted Priority Debt has the meaning given in the Intercreditor Agreement." "Original Financial Statements means the statements delivered under the Existing Facility." In fund finance the rate is higher, because NAV is routinely defined by reference to the fund's constitutional documents, which the lender holds in a different folder and often a different version.
What depth does to extraction
A system that reads the covenant sentence and stops has extracted a label and a number. It has not extracted the test. The test lives at the leaves, and three things follow.
- Resolution has to be recursive and it has to stop. The parser follows each hop until it reaches a number, a date, GAAP, or a reference outside the document. A reference outside is a first-class result: stored as a dependency on another document, flagged until that document is loaded, never silently treated as resolved.
- Every hop is a citation. When a credit officer opens the leverage test they should be able to walk the chain: covenant, ratio definition, EBITDA definition, add-back cap, each with its page. That is only possible if the chain was built as structure rather than pasted as text.
- Depth predicts error. Our production first pass scores 97.4 percent citation-exact across the set. Split by depth, fields at one or two hops score 99.1 and fields at three or more score 94.9. In the benchmark, the gap between model classes opens at exactly the same point. Depth is the single best predictor of where a model needs the parser's help, and where a person should look, which is why it is one of the inputs to the confidence that routes the second read.
The number in the covenant is the least important number in the test. The ones that matter are six definitions down, and one of them is in a different agreement.
A note for drafters
None of this is a complaint about drafting. Deep chains exist because deals are complicated and precedents are reused. But the 38 percent figure is worth pausing on. Every external reference is a dependency the lender has to manage for the life of the facility, and in our experience the referenced document is missing from the file more often than anyone expects. Where it can be avoided, a definition that stays inside the four corners is a gift to whoever has to test it in year six.