Covenant Watch
Borrower certificates recomputed from the credit agreement's own definitions, with headroom, add-back caps and cure rights tracked every test date.
Each quarter your borrowers send financial statements and a signed compliance certificate, typically 45 to 60 days after quarter end. We read the package against the credit agreement as amended: the caps on EBITDA add-backs, the limit on netted cash, the step-down schedule. Code recomputes every ratio. Where the result differs from the certificate, you see why, line by line, with the clause cited. Headroom, cure windows and cure counts are tracked per borrower, and a named person signs every breach.
cap = greater of 5.00 and 20% × 31.60 = 6.32
disallowed 3.28
The certificate applies the $12.00m fixed cap, which ended with the 31 Mar 2026 test period.
A quarterly test that starts after the quarter ends
Covenant monitoring runs on documents the borrower prepares. The credit agreement sets the definitions; the compliance certificate applies them. Your team checks one against the other, borrower by borrower, and has to act while the cure window is still open.
- 1Credit analyst, with deal counselAbstracts each covenant, the EBITDA definition with its add-back caps, and the cure clause, at closing and at every amendment.
- 2Portfolio monitoring analystKeeps each borrower's reporting calendar: monthly statements, quarterly statements with the compliance certificate, audited annual accounts.
- 3Borrower's Responsible OfficerSends the statements and a signed compliance certificate by email or portal, each ratio shown line by line.
- 4Credit or portfolio analystSpreads the financials, recomputes EBITDA and each ratio under the agreement's definitions, compares them with the certificate and measures headroom.
- 5Portfolio manager or credit committeeReviews headroom trends, watchlists tightening names, and weighs cure, waiver or amendment when a test fails.
- 6Administrative agentCirculates certificates and notices to the lender group; waivers and amendments go to the Required Lenders for consent.
- Credit agreement, with schedules, exhibits and every amendment
- Compliance certificate with its covenant schedules
- Monthly, quarterly and audited annual financial statements
- Sponsor or base-case model
- Borrowing base certificate, on asset-based deals
- Covenant abstract per agreement
- Watchlist row per borrower and covenant: level, actual, headroom, status, next deadline
- Exception or breach memo
- Cure log
- Covenant status for the valuation committee, investor reporting and fund-facility reporting
Four places a covenant test goes wrong
- 1
Taking the certificate's arithmetic
Certificates are prepared by the borrower. Run-rate add-backs for new customer contracts and revenue synergies can "meaningfully inflate covenant headroom" (Practical Law, 2025). A missed breach is a lost negotiating moment.
- 2
Testing against a superseded term
Add-back caps and covenant levels are negotiated per deal and change by amendment. In one filed amendment, the conformed text runs deleted and inserted words together, so a plain extract reads "iswas". Miss that and you test the wrong level.
- 3
A cure clock missed or misread
Equity cure funding is typically due 10 to 15 days after the financial statements are required. Cure proceeds count only towards the financial covenants, not pricing. A missed date loses the cure; a misapplied one hides the breach.
- 4
Stress hidden from breach counts
Lincoln International attributes a lower-than-expected covenant default rate partly to amendments that pre-empt defaults and to elevated PIK use; 11% of the loans it valued carried PIK in 2025. Breach counts alone understate stress.
The model reads, code computes, a person signs
Covenant Watch is a fixed workflow, not a free-roaming agent. The model reads and drafts. Code computes and compares. A named person approves the covenant spec, reviews anything flagged, and signs every breach.
Breach determinations, cure acceptance, waiver and amendment decisions, and every message to the borrower, the administrative agent or the lender group stay with your team. The covenant spec never changes without an analyst's approval. The system proposes; it cannot send, post or sign.
Credit agreement §8.04 Cure Right · as amended by Amd. No. 2
Cure EBITDA counts for the financial covenants only. It is disregarded for the pricing grid and baskets. The sponsor and lenders decide; the system does not.
Every figure traces to a page and a clause
Each test leaves an evidence line: where the figure came from, which version of the agreement governed it, what code did with it, and who signed. Records are stored, not regenerated, because model output can vary between runs. Your reviewer, auditor or valuation committee can follow any number back to its source.
Kept for every run
- 01Agreement version and amendment chain, with file hashes
- 02Covenant spec version and the analyst who approved it
- 03File hashes for every document in the package
- 04Page and span for each extracted value
- 05The calculation trace: inputs, caps, ratio, headroom, Cure Amount
- 06Prompt versions and the pinned model version
- 07Validator results and flags
- 08Reviewer decision, edits and timestamps
We will walk the definitions with you, show where the recompute could differ, and tell you what it would take to build.
A fixed workflow with the model at the edges
For engineers and technical reviewers. The breach decision has to be a calculation over cited inputs, so the model never makes it. It reads long agreements and messy schedules into typed records; deterministic code does every piece of arithmetic; people approve the rules and sign the results.
2026-05-18T09:15Z
Sch. 1 l.I · p.3
Qualified cash 9.80
Sch. 1 l.II · p.3
EBITDA before add-backs 31.60
Sch. 4 · p.5
Add-backs 5.10 · 3.20 · 1.30
Sch. 4 · p.5
net debt: 182.40 − 7.50 = 174.90
Shared Cap: max(5.00, 0.20×31.60) = 6.32
add-backs allowed 6.32 · disallowed 3.28
EBITDA: 31.60 + 6.32 = 37.92
headroom −2.5% · Cure Amount 0.95
regression case TBC_Q2FY26_cap_applied · expects 4.61x · re-run on every amendment, prompt or model change
Harness
- Pattern: prompt chaining, one section per covenant. Locate definitions, extract figures, then test deterministically. Anthropic recommends fixed workflows over agents where a task is well defined and needs to be predictable.
- One bounded agent loop, read-only. Resolving a definition chain across amendments may need several lookups; that loop can read documents and nothing else.
- Typed hand-off. Every model output crossing into code is a record: field, value, unit, as-of date, document, page, span, extractor version. It is schema-checked, then value-checked.
- Deterministic engines. A ratio calculator driven by the encoded definitions and step-down schedule; a headroom and cure calculator; date logic for test periods and cure windows.
Data and integrations
- Credit agreements and amendments as PDF or HTML, including conformed changed pages
- Signed compliance certificates (PDF) with covenant schedules; monthly and quarterly variants
- Financial statements as PDF or Excel: monthly, quarterly and audited annual
- Borrowing base certificates on asset-based deals; sponsor model and agreement-fixed EBITDA for pre-closing quarters
- Facility data read from your loan system; prior-quarter test results
- Arrives by borrower portal or email; no industry message standard for certificates exists
- Outputs to your watchlist or portfolio monitoring platform, and to valuation and fund-facility reporting
What is hard for machines, and what we do about it
Definitions chain across clauses and amendments. On 170 EDGAR credit agreements averaging about 84,000 tokens, the 16 models tested (2024–25, including GPT-4o) struggled with multi-hop and set questions (KG-MuLQA, 2025 preprint).
Store each agreement's definitions as a graph, test the multi-hop chains, and have an analyst approve the encoded spec.
Finding the right page. In FinanceBench (Nov 2023), GPT-4-Turbo answered 19% of questions correctly with a shared vector store and 85% when handed the right pages.
Keep one index per agreement and per package, never a shared pool, and measure retrieval on its own.
Tables. In late-2024 tests the best PDF parser scored about 79 of 100 on table structure, and GPT-4o 72 (OmniDocBench).
Parse layout first, then cross-foot every schedule: quarters sum to the test period, net debt equals debt less capped cash, units agree.
Arithmetic in text. Writing programs instead of computing in text improved accuracy by about 12% (Program of Thoughts, Codex, 2022); one inconsequential clause cut accuracy by up to 65% (GSM-Symbolic, 2024 models).
Compute every cap, ratio, headroom figure and Cure Amount in code, from cited inputs only.
Borrower documents are untrusted input. Undefended, targeted prompt injection succeeded in 47.7% of AgentDojo security cases against GPT-4o (2024-05-13); a least-privilege tool filter cut that to 7.5%.
The reader has no write tools and no network access, returns typed fields only, and seeded injection cases sit in the test set.
How we know it works
- Golden set, built with your team. Past quarters across several of your deals, plus public EDGAR agreements and synthetic certificates with fictional names: amended agreements, caps taken before and after add-backs, step-down boundary dates, negative EBITDA, cure events, scanned certificates, injected text. Two reviewers label independently; we measure their agreement.
- Scored per field, not on average. Exact match on certificate lines, test levels and dates; recomputed ratios agreeing at the two-decimal "x.xx to 1.00" presentation; breach recall weighted above precision; share of differences explained with a clause cited.
- Release gate. Each fortnightly release passes the full regression suite, including every known breach in the golden set, before it ships.
- Consistency, not one good run. The same package is run several times and must give the same result every time (pass^k).
- Shadow run, then acceptance. One full quarterly test cycle alongside your current process, which stays authoritative. Each disagreement is logged with its cause. Your business owner signs against agreed thresholds and a stated sampling plan.
- Regression on change. The suite re-runs on every amendment, prompt change and model migration. Models are pinned to fixed versions.
Controls and security
- Least privilege. The model can read documents and draft proposals. No tool that sends, posts or signs is exposed to it.
- Two approvals. An analyst approves the covenant spec; a portfolio manager or credit officer signs every breach and near-breach memo. The system counts as the maker, so a person is always the checker.
- Documents are data. Every package is scanned for hidden text. An instruction found inside a document is flagged and routed to review, never followed.
- Secrets and traces. Credentials are vault-held and never enter prompts or logs; capture of prompt content in traces is opt-in and redacted.
- Hosting and providers. Agreed with each client and named in writing: where the model runs, which providers process data, and where it is stored.
The rules around covenant monitoring, as of 28 September 2026
- FCA multi-firm review of private market valuation practicesExpects documented valuation decisions and a process whose adherence can be tested. Covenant status feeds those decisions.
- SEC Division of Examinations, FY2026 prioritiesNames private credit among alternative investments, and checks that firms' AI claims are accurate and their AI use supervised.
- AIFMD II, Directive (EU) 2024/927Loan-originating AIFs need policies for monitoring their credit portfolio, reviewed at least annually. Transposition varies by member state.
- Fed SR 26-2 / OCC Bulletin 2026-13Revised US bank model-risk guidance. Generative and agentic AI are outside its scope; your wider governance applies.
- FSB, Vulnerabilities in Private CreditProposes monitoring covenant breaches, waivers and amend-extends alongside PIK share, not breaches alone.
- LSTA model provisions for private corporate credit deals (PCC MCAPs), finalModel terms for private credit, members only. Definitions and caps are still negotiated deal by deal.
Terms
- Compliance certificate
- Officer-signed certificate delivered with the financial statements, showing the covenant calculations and confirming no default. Same term US and UK/EU.
- Maintenance covenant
- A financial test the borrower must pass on every test date, usually quarterly. Same term US and UK/EU.
- Headroom / cushion
- The EBITDA underperformance against plan that would cause a breach. "Headroom" in UK/EU usage; US commentary often says "cushion".
- EBITDA add-backs
- Adjustments such as synergies, restructuring and transaction costs that the agreement lets the borrower add to EBITDA.
- Shared cap (combined cap)
- One cap across several add-back categories, often the greater of a dollar amount and a percentage of EBITDA, taken before or after add-backs.
- Step-down
- A scheduled tightening of a covenant level on set dates.
- Test period (LTM)
- The four consecutive fiscal quarters ending on the test date. LMA-style drafting often says "Relevant Period" (not verified).
- Equity cure (Cure Right)
- A sponsor equity contribution deemed to increase EBITDA to cure a financial covenant breach, within limits the agreement sets.
- Cure Amount
- The amount needed to restore compliance; contributions above it are usually not allowed.
- Springing covenant
- A test that applies only when revolver usage passes a threshold or a trigger event occurs. Mainly US usage.
- Administrative agent / facility agent
- The lender group's agent that circulates certificates and notices. "Administrative agent" in US drafting; "facility agent" in LMA/UK usage.
- Required Lenders
- The lender majority whose consent approves most waivers and amendments. LMA-style drafting says "Majority Lenders" (not verified).
Recently updated
- Cure eligibility and cure-count tracking per agreement.
- Rules strip brought up to date as of 28 September 2026: the revised US bank model-risk guidance, AIFMD II's application date and the final LSTA private credit model provisions.
Questions
When the package arrives. Quarterly certificates typically come 45 to 60 days after quarter end, so no test result can exist at quarter close. The run starts on receipt, and your analyst has the recompute, the explanation of any difference and the cure dates while the cure window is still open.
No. Code computes each ratio from the encoded definitions and marks the test pass or fail. A portfolio manager or credit officer signs every breach and near-breach memo. Cure acceptance, waivers, amendments and anything sent to the borrower, the agent or the lender group stay with your team.
Each amendment creates a new version of the covenant spec. The model drafts the changes with the clause cited, code checks dates and sequence, and a credit analyst approves. Earlier versions are kept, so a past quarter is always tested against the terms then in force, and the regression suite re-runs on the new version.
Hosting, model providers and data residency are agreed with each client and named in writing. In every set-up, borrower documents are read by a component with no write access and no network egress, model versions are pinned, and each run's inputs, prompts, model version and outputs are stored for the retention period you set.
It is not built to. Covenant Watch tests the borrower's arithmetic against the agreement; it does not verify collateral or the borrower's books. First Brands, which disclosed about $2.3bn of unpaid factoring obligations in its 2025 Chapter 11, and Tricolor, whose former executives the SEC charged with double-pledging loans, were collateral cases, not covenant arithmetic.
The US bank guidance revised on 17 April 2026 (SR 26-2 / OCC Bulletin 2026-13) puts generative and agentic AI outside its scope and excludes deterministic rule-based processes from its definition of a model. We make no compliance claim. We design to its principles, independent challenge, outcomes analysis and ongoing monitoring, and your own governance sets the controls.
