LBO Structured Review Prompt Pack

<role>You are a leveraged finance practitioner who has reviewed many buyout structures. You are a colleague, not a prosecutor: you build the picture first, test only what bears on the decision, and retire a flag once the context explains it.</role>

<surface>
Where are you running this? Pick the one that matches, and I will say so if it is the wrong place.
(A) Claude app, a private Project: one deal, a handful of documents you can upload. Chat is correct here.
(B) Claude Code, pointed at a folder: a data room export, many files, spreadsheets to parse, or a PUBLIC target whose filings you want pulled from SEC EDGAR (prompts 04 and 05 need this).
(C) Claude Code locally, or your company's own deployment: material that cannot be uploaded anywhere.
Model: select Claude Opus 5.5 in the model picker for every judgment prompt (03, 06 to 11). Claude Sonnet 5 is fine for a first read of a very large data room in 03. Do not switch model inside a prompt.
STOP and re-route, do not continue anyway, when: you paste a file path, a folder or a screenshot of a file list into chat (that is a Claude Code job, open Claude Code in that folder); the material arrives truncated or is too large to hold (name the documents missing and do not score what you could not read); a figure is needed from a document you have not provided (I ask once, name the document, then stop).
</surface>

<onboarding>
Ask one block at a time. Wait for each answer.
1. DEAL TYPE: (A) take-private of a public company (B) platform buyout of a private company. Not covered: add-ons and carve-outs (standalone and transition service costs, allocated overhead and synergies need their own checks, which this pack does not have). If the deal is one of those, say so and stop.
2. DATA SOURCE, pick all that apply: (A) upload documents to this Project (B) paste figures (C) upload the model workbook as a file (D) Claude Code reading a data room folder (E) Claude Code pulling a public target's 10-K and 10-Q from SEC EDGAR with prompts 04 and 05 (F) a mix.
3. DECISION DATE: the date the decision is taken. Everything filed or known after it is excluded. For a past deal this is how you test the pack blind.
4. YOUR THESIS, in your words: why this deal works, and your role (sponsor, lender, co-investor, seller).
5. TOKENS: {{TARGET_NAME}}, {{CIK_OR_TICKER}} (public only), {{DECISION_DATE}}, {{ENTRY_EV}}, {{HOLD_YEARS}}, {{HURDLE_IRR}}, {{EXIT_MULTIPLE}}. Claude Code only: {{PROJECT_DIR}} (the deal folder). Public target only: {{SEC_NAME}} and {{SEC_EMAIL}}, your real name and email, which the SEC requires on every filings request.
Then offer: CONVERSATION mode (you state the question, I ask for what I need and run only the prompts that bear on it) or FULL sequence 03 to 12. Conversation is the default.
</onboarding>

<rules>
EVIDENCE TIERS. TIER 1: a filed or signed document (10-K, 10-Q, merger proxy, commitment letter, audited statements), cited by accession number or document and page. TIER 2: a figure keyed or exported by a person (model workbook, management schedule). TIER 3: marketing (CIM narrative, press release, teaser): it generates a question and never a number. Every load-bearing figure carries its tier and source id.
MISSING OR BAD INPUT. A missing figure is OPEN: ask once, naming the document that would carry it. Two sources that disagree are both shown with their source ids; never pick one. A document dated after the decision date is excluded and named as excluded. Never invent a number, a rate or a term; a figure not in the documents is labelled ASSUMPTION with a range, and the verdict is run at both ends of the range.
FLAGS. Every finding is tagged CHANGES THE DECISION, WORTH A QUESTION, or EXPLAINED BY CONTEXT, after testing it against the normal patterns below. Only CHANGES THE DECISION reaches the verdict. EXPLAINED BY CONTEXT is listed once and closed; a closed flag never colours a later prompt. WORTH A QUESTION only if the answer could move leverage, cover or IRR by a stated amount (default: 0.25x leverage, 0.25x cover, 2 points of IRR) or change the verdict. Before raising any question, test it twice: is it on the normal patterns list (then it goes to "Checked, normal"), and is it an input simply not provided, such as a rate, fees or a drawn amount (then it goes to OPEN with the document that would carry it)? Only what passes both tests and alone could move a figure past the threshold is a question. One question is one item, stating the figure it moves and by how much. Conventions, rounding, disclosed one-offs and timing go in one untagged line: "Checked, normal: ...".
NORMAL PATTERNS (test every flag against these first): a named one-time charge added back once in five years; a large equity check (it does not change what the debt can carry); the same line tagged with a different XBRL concept in two filings; captive finance debt serviced by its own receivables (abnormal if above eligible receivables, with recourse to corporate borrowers, or refinanced with corporate term debt); an undrawn revolver or ABL commitment (not funded debt); balance sheet cash used as a source; amortization of acquired intangibles inside D&A; a 10-K/A filed after the decision date; bridge commitments replaced by notes at close; convertible notes outstanding at signing; fees and OID in uses; one down year of revenue in a mature business (abnormal when operating income falls several times faster: run the downside). Precedence: arithmetic and recurrence outrank this list (a "one-time" charge taken four years running is a cost).
ARITHMETIC. Carry full precision, round at the end, show the working so a person can re-derive every figure. A tie-out that breaks by more than 0.5% on the same basis blocks.
VERDICTS allowed: PROCEED, PROCEED ON CONDITIONS, PROCEED ONLY ON CHANGED TERMS (name the terms), DECLINE, NO VIEW (only with a named deciding question that is split and would flip the verdict; "the future is uncertain" is not one). A block stops the verdict, not the analysis: keep running the other checks and print what happens by default if nobody acts.
PRIVACY. Your documents go only to your own Claude account, never to us. Use a Team or Enterprise plan, or turn model training off, before loading confidential material.
</rules>

<adapt>
How to adapt this pack: change the NORMAL PATTERNS for your sector (add yours, keep the format: pattern, why normal, what makes it abnormal); change the materiality defaults in FLAGS to your fund's; swap prompt 07's model structure for your house template; add tranches in 07. Keep the tiers, the decision date and the blocking tie-out.
</adapt>

Confirm the five blocks back to me in one table, state the surface and model, then say which prompt runs next.

<task>Run this synthetic sample deal through the pack's logic before any real data. Compare your output with the expected output below line by line. MATCH is judged on substance to the stated rounding; a normal pattern cleared as EXPLAINED BY CONTEXT counts as a match. If one figure does not match, say which and STOP. Do not load real data until every line matches.</task>

<sample>
Target: Harbor Components (synthetic). Decision date 2026-03-31. Role: lender.
Thesis: "Steady industrial demand; 4% EBITDA growth; the debt pays down from free cash flow."
10-K FY2025 (TIER 1): operating income 52.0; D&A 18.0; stock comp 4.0; litigation settlement 6.0 recorded in Q3 2025 (note 12: first settlement in five years, no other claims pending).
10-Q Q1 2026 (TIER 1): operating income 14.0 (Q1 2025: 12.0); D&A 4.6 (Q1 2025: 4.4). Stock comp flat.
Commitment letter (TIER 1): EV 8.0x adjusted EBITDA before stock comp; term loan 4.0x at 8.0%; second lien 1.0x, rate: page missing from the upload; revolver 50.0, undrawn at close. Fees 12.0.
Plan (TIER 2): capex 15.0, cash taxes 8.0, working capital build 2.0 a year; EBITDA +4% a year; exit year 5 at 8.0x; all free cash flow after interest repays the term loan.
</sample>

Expected output:
1. STOP fires once: "Second lien rate is missing from the commitment letter (page not in the upload). Provide it; I will not assume it." Resume only when given 11.0%.
2. LTM: OI 52.0 + 14.0 - 12.0 = 54.0. D&A 18.0 + 4.6 - 4.4 = 18.2. EBITDA 72.2.
3. Bridge: + litigation 6.0 = 78.2 (EXPLAINED BY CONTEXT: named, first in five years). With stock comp added back 82.2, shown on its own line, not used for sizing.
4. Structure: EV 8.0 x 78.2 = 625.6. Term loan 312.8, second lien 78.2, total 391.0 = 5.0x. Revolver: EXPLAINED BY CONTEXT, undrawn, not debt. Equity 625.6 + 12.0 - 391.0 = 246.6.
5. Interest 312.8 x 8.0% = 25.02; 78.2 x 11.0% = 8.60; total 33.63. Cover 78.2 / 33.63 = 2.33x. Breakeven EBITDA decline to 1.0x cover: 1 - 33.63/78.2 = 57.0%.
6. FCF before interest 78.2 - 15.0 - 8.0 - 2.0 = 53.2; after interest 19.57 (held flat, a simplification the output states).
7. Exit: EBITDA 78.2 x 1.04^5 = 95.14; EV 761.14; debt 391.0 - 5 x 19.57 = 293.13; equity 468.01. MOIC 468.01 / 246.6 = 1.90x. IRR 1.90^(1/5) - 1 = 13.7%.
8. Tie-out: sources = uses (391.0 + 246.6 = 625.6 + 12.0); MOIC re-derived; all PASS.
9. Verdict: PROCEED. No CHANGES THE DECISION flags. "Checked, normal: litigation add-back, undrawn revolver."

Output columns: line, expected, yours, MATCH or MISMATCH. Then one line: "Calibration passed" or the first mismatch.

<role>Deal lead opening the file.</role>
<task>Using the data source from prompt 01:
1. In six lines: what the company sells, to whom, how it makes money, what it costs to keep running (capex, people), the rate environment on the decision date, and the human's thesis quoted.
2. Index the documents present against what this review needs: last audited year, latest interim, merger agreement or commitment letters with tranche terms, debt agreements of debt that survives, management adjustments schedule, customer concentration, related party notes. Each absent item is OPEN, with the prompt it blocks.
3. The three facts the thesis depends on, and which document proves each.
4. Read management transcripts, footnotes and related party notes for anything that changes a figure the model will use (a related party lease that resets rent, a customer above 10% of revenue). Tag each.</task>
<trap>A data room that is complete for a quality of earnings review can still lack the one document the structure turns on: the tranche terms. Without rates and amortization there is no debt capacity, whatever else is present.</trap>
<stop>If no document states the purchase price and the debt tranches (merger agreement, commitment letter or term sheet), say so and stop the structure prompts (07 to 09); 04 to 06 can still run.</stop>
<output_format>Big picture (six lines), document index table (document, present or absent, tier, blocks prompt), the three thesis facts, tagged findings, "Checked, normal:" line, "Next step:".</output_format>

<role>Engineer setting up a small, tested workspace that pulls a public target's filed numbers.</role>
<task>Run in Claude Code. Create this tree at {{PROJECT_DIR}} and write every file exactly as given. Do not summarise the plan back to me. Print `ls -R` when done.
{{PROJECT_DIR}}/scripts/pull_filed.py : pulls 10-K and 10-Q XBRL facts filed on or before a date, builds LTM
{{PROJECT_DIR}}/scripts/tie.py : ties model inputs to the filed LTM, exits 1 on a break
{{PROJECT_DIR}}/tests/test_pull.py : known-answer test on a real filing
Then run, one at a time, and show the output of each:
cd {{PROJECT_DIR}}
python3 --version   (3.10 or higher; if lower, stop and install Python from python.org first)
python3 -m venv .venv
.venv/bin/python -m pip install edgartools==5.59.1 pytest==9.1.1
EDGAR_IDENTITY="{{SEC_NAME}} {{SEC_EMAIL}}" .venv/bin/python -m pytest -q
PASS CONDITION: "1 passed". The test pulls Twitter's FY2021 10-K and must read revenue 5,077,482,000 and operating income -492,739,000. Any other result: stop and use the debug table in prompt 05.
Always call .venv/bin/python, never python3, inside this workspace. {{SEC_NAME}} and {{SEC_EMAIL}} come from prompt 01: the SEC requires a real name and email on every request and allows at most 10 requests per second. If they were not captured, ask for them once and stop.

FILE scripts/pull_filed.py
import sys, csv, os, datetime as dt, warnings
warnings.filterwarnings("ignore")
from edgar import Company, set_identity
ITEMS = {
    "revenue": ["RevenueFromContractWithCustomerExcludingAssessedTax", "Revenues", "SalesRevenueNet"],
    "operating_income": ["OperatingIncomeLoss"],
    "d_and_a": ["DepreciationDepletionAndAmortization", "DepreciationAndAmortization",
                "DepreciationAmortizationAndAccretionNet", "OtherDepreciationAndAmortization"],
    "stock_comp": ["ShareBasedCompensation", "AllocatedShareBasedCompensationExpense"],
    "interest_expense": ["InterestExpense", "InterestExpenseDebt"],
    "net_income": ["NetIncomeLoss"],
    "cfo": ["NetCashProvidedByUsedInOperatingActivities"],
    "capex": ["PaymentsToAcquirePropertyPlantAndEquipment", "PaymentsToAcquireProductiveAssets"],
}
BALANCE = {"cash": ["CashAndCashEquivalentsAtCarryingValue"],
           "short_term_investments": ["ShortTermInvestments", "MarketableSecuritiesCurrent", "AvailableForSaleSecuritiesDebtSecuritiesCurrent"]}
def facts(filing):
    df = filing.xbrl().facts.to_dataframe()
    df = df[~df["is_dimensioned"].fillna(False).astype(bool)]
    df = df[df["numeric_value"].notna()]
    df["c"] = df["concept"].str.split(":").str[-1]
    return df
def durations(df, cands):
    for c in cands:
        d = df[(df["c"] == c) & (df["period_type"] == "duration")]
        if len(d):
            d = d.assign(days=(d["period_end"].astype("datetime64[ns]") - d["period_start"].astype("datetime64[ns]")).dt.days)
            return c, d
    return None, None
def main():
    who, ident, as_of = sys.argv[1], sys.argv[2], dt.date.fromisoformat(sys.argv[3])
    set_identity(who)
    co = Company(int(ident) if ident.isdigit() else ident)
    fl = co.get_filings(form=["10-K", "10-Q"])
    rows = [f for f in fl if f.filing_date <= as_of]
    later = [f for f in fl if f.filing_date > as_of]
    if not rows: sys.exit(f"STOP: no 10-K or 10-Q filed on or before {as_of}.")
    tenk = next((f for f in rows if f.form == "10-K"), None)
    if tenk is None: sys.exit("STOP: no 10-K on or before the as-of date inside the listing.")
    q = next((f for f in rows if f.form == "10-Q" and f.filing_date > tenk.filing_date), None)
    print(f"{co.name} CIK {co.cik} as of {as_of}")
    print(f"USED 10-K {tenk.accession_no} filed {tenk.filing_date}" + (f"; 10-Q {q.accession_no} filed {q.filing_date}" if q else "; no later 10-Q"))
    print(f"EXCLUDED as after the as-of date: {len(later)} periodic filing(s)" + (f", first {later[-1].form} filed {later[-1].filing_date}" if later else ""))
    os.makedirs("out", exist_ok=True)
    try:
        fk = facts(tenk); fq = facts(q) if q else None
    except Exception as e:
        sys.exit(f"STOP: no usable XBRL in the filing ({e}). Filings before 2009 to 2011 carry none; read the statements from the filing text instead.")
    src = lambda f: (f.form, f.accession_no, str(f.filing_date))
    ff, lt = [], []
    for item, cands in ITEMS.items():
        c, d = durations(fk, cands)
        if c is None:
            near = sorted(set(fk[(fk["period_type"] == "duration") & fk["c"].str.contains(cands[0][:10])]["c"]))[:6]
            ff.append([item, "NOT FOUND", "", "", "", *src(tenk)]); lt.append([item, f"NOT FOUND in XBRL under {cands}; nearby tags {near}. Read the statement line in the filing text, do not substitute"]); continue
        years = d[(d["days"] > 350) & (d["days"] < 380)].sort_values("period_end").drop_duplicates("period_end")
        fy = years.iloc[-1]
        for _, y in years.iterrows(): ff.append([item, c, y["period_start"], y["period_end"], y["numeric_value"], *src(tenk)])
        if q is None:
            lt.append([item, f"LTM = FY {fy['numeric_value']:.0f} ({fy['period_end']})", fy["numeric_value"]]); continue
        c2, d2 = durations(fq, [c] + [x for x in cands if x != c])
        if c2 and c2 != c:
            print(f"NOTE {item}: 10-K tags {c}, 10-Q tags {c2}. Same line, different tag; check the two labels match.")
        ytd = d2[d2["period_start"] > fy["period_start"]] if c2 else None
        if ytd is None or not len(ytd):
            lt.append([item, f"STOP: {c} missing from the 10-Q; LTM not built"]); continue
        cur = ytd[ytd["period_end"] == ytd["period_end"].max()].sort_values("days").iloc[-1]
        prior = d2[(d2["period_end"] < cur["period_end"]) & (abs(d2["days"] - cur["days"]) < 8) & (d2["period_start"] <= fy["period_start"]) & (d2["period_end"] <= fy["period_end"])]
        if not len(prior):
            lt.append([item, f"STOP: prior-year YTD for {c} not in the 10-Q; LTM not built"]); continue
        pr = prior.sort_values("period_end").iloc[-1]
        val = fy["numeric_value"] + cur["numeric_value"] - pr["numeric_value"]
        ff.append([item, c, cur["period_start"], cur["period_end"], cur["numeric_value"], *src(q)])
        ff.append([item, c, pr["period_start"], pr["period_end"], pr["numeric_value"], *src(q)])
        lt.append([item, f"LTM {val:.0f} ({c2}) = FY {fy['numeric_value']:.0f} + YTD {cur['numeric_value']:.0f} ({cur['period_start']} to {cur['period_end']}) - prior YTD {pr['numeric_value']:.0f} ({pr['period_start']} to {pr['period_end']})", val])
    last = fq if q is not None else fk
    lf = q if q is not None else tenk
    inst = last[last["period_type"] == "instant"]
    assets = inst[inst["c"] == "Assets"]
    bdate = assets["period_instant"].max() if len(assets) else inst["period_instant"].max()
    inst = inst[inst["period_instant"] == bdate]
    for item, cands in BALANCE.items():
        hit = inst[inst["c"].isin(cands)]
        ff.append([item, hit.iloc[0]["c"] if len(hit) else "NOT FOUND", "", bdate, hit.iloc[0]["numeric_value"] if len(hit) else "", *src(lf)])
    debt = inst[inst["c"].str.contains("Debt|Notes|Borrowing|FinanceLease", regex=True) & ~inst["c"].str.contains("Investment|Receivable|Securities|Fair|Issuance|Proceeds|Repayment|PropertyPlant|Asset", regex=True)]
    with open("out/debt_lines.csv", "w", newline="") as fh:
        w = csv.writer(fh); w.writerow(["concept", "as_of", "value", "form", "accession", "filed"])
        for _, r in debt.drop_duplicates("c").iterrows(): w.writerow([r["c"], bdate, r["numeric_value"], *src(lf)])
    with open("out/filed_facts.csv", "w", newline="") as fh:
        w = csv.writer(fh); w.writerow(["item", "concept", "period_start", "period_end", "value", "form", "accession", "filed"]); w.writerows(ff)
    with open("out/ltm.csv", "w", newline="") as fh:
        w = csv.writer(fh); w.writerow(["item", "working", "ltm_value"]); w.writerows([r + [""] * (3 - len(r)) for r in lt])
    for r in lt: print(" | ".join(map(str, r[:2])))
    print(f"Balance sheet date {bdate}; debt lines written to out/debt_lines.csv ({len(debt.drop_duplicates('c'))} rows). TIER 1 source: the accessions above.")
if __name__ == "__main__":
    main()

FILE scripts/tie.py
import csv, sys
ltm = {r["item"]: r["ltm_value"] for r in csv.DictReader(open("out/ltm.csv"))}
breaks = 0
for r in csv.DictReader(open(sys.argv[1])):
    filed = ltm.get(r["item"], "")
    if filed in ("", None):
        print(f"{r['item']}: OPEN, no filed LTM figure (see out/ltm.csv)"); continue
    m, f = float(r["value"]), float(filed)
    gap = (m - f) / abs(f) if f else float("inf")
    ok = abs(gap) <= 0.005
    breaks += not ok
    print(f"{r['item']}: model {m:,.0f} vs filed {f:,.0f} gap {gap:+.2%} {'TIES' if ok else 'BREAK'}")
sys.exit(1 if breaks else 0)

FILE tests/test_pull.py
import subprocess, csv, os, sys, pathlib
ROOT = pathlib.Path(__file__).resolve().parents[1]
def test_known_filing(tmp_path):
    who = os.environ.get("EDGAR_IDENTITY")
    assert who, "set EDGAR_IDENTITY to your real name and email, as prompt 01 captured"
    r = subprocess.run([sys.executable, str(ROOT / "scripts/pull_filed.py"), who, "1418091", "2022-03-01"], cwd=tmp_path, capture_output=True, text=True)
    assert r.returncode == 0, r.stdout + r.stderr
    ltm = {x["item"]: x["ltm_value"] for x in csv.DictReader(open(tmp_path / "out/ltm.csv"))}
    assert float(ltm["revenue"]) == 5077482000
    assert float(ltm["operating_income"]) == -492739000
</task>
<trap>Naming a script after a Python standard library module (gettext.py, csv.py, email.py) inside this folder breaks `from edgar import ...` with a circular import error. Keep the names above.</trap>
<stop>If the test does not print "1 passed", do not run prompt 05 on a live target. Say which line failed and stop.</stop>

<role>Analyst building the LTM from filed numbers only.</role>
<task>In the workspace from prompt 04:
1. Run: .venv/bin/python scripts/pull_filed.py "{{SEC_NAME}} {{SEC_EMAIL}}" {{CIK_OR_TICKER}} {{DECISION_DATE}}
2. Report the USED and EXCLUDED lines, then the LTM table from out/ltm.csv with the working for each line, each figure TIER 1 with its accession number.
3. Read every NOTE line: open both filings' statements and confirm the two labels describe the same line before accepting the LTM.
4. From out/debt_lines.csv, list debt on the balance sheet date. Mark which tranches the deal documents say are refinanced, and which survive; "all or a portion" leaves the amount OPEN, never assumed.
5. Write model_inputs.csv (columns item,value) with every LTM figure the model will use, then run .venv/bin/python scripts/tie.py model_inputs.csv. Any BREAK blocks prompts 06 to 09 until the input is corrected or the difference is explained in writing with a source.</task>
<trap>The lookahead cut. A 10-K/A or a later 10-Q often restates the very quarter you are reviewing; using it tests the deal with facts nobody had. The script excludes anything filed after the decision date: keep it that way even when the later number looks better.</trap>
<debug>
Symptom: "STOP: no usable XBRL". Cause: an older filing (before about 2009 to 2011) has no XBRL. Fix: read the income statement, cash flow and debt note from the filing text; tag them TIER 1 with the page.
Symptom: an LTM line says STOP, tag missing from the 10-Q. Cause: the 10-Q tags the line under a concept not in the list. Fix: open the 10-Q cash flow statement, find the label, add that concept to the list in ITEMS, re-run.
Symptom: NOT FOUND for D&A. Cause: filers use OtherDepreciationAndAmortization or a company tag. Fix: the script prints nearby tags; confirm against the statement label.
Symptom: segment revenue appears instead of the total. Cause: dimensioned facts. Fix: the script drops them; if you edit it, keep the is_dimensioned filter.
</debug>
<stop>If the decision date falls before the first 10-K in the listing, or the pull prints STOP for operating income, D&A or interest, say so and stop; do not build EBITDA from the part that loaded.</stop>

<role>Quality of earnings reviewer.</role>
<task>Using the data source from prompt 01 (the LTM from prompt 05 when the target is public):
1. Reported EBITDA = operating income + D&A, with the working, for the LTM and each prior year in out/filed_facts.csv or the audited statements.
2. Each adjustment on its own row: amount, the document and page that names it, recurring or one-time (test: did it occur in any of the prior four years), and a tag.
3. Stock comp on its own line. Show EBITDA before and after adding it back; size the debt on the before figure unless the documents show it will be replaced by equity that costs the business no cash.
4. Cash conversion: CFO minus capex against EBITDA. A gap above a third needs a named cause (working capital, taxes, capitalised costs).
5. Private targets only: tie revenue and EBITDA in the adjustments schedule to the audited statements and the trial balance; a proof of cash for the latest year if bank statements are present.</task>
<trap>A management adjustment that recurs is a cost. "Restructuring", "integration" or "one-time" charges taken three years running belong in EBITDA, whatever the label says.</trap>
<stop>If the management adjustments schedule is absent and the filings do not name an adjustment, do not add it back; list it as OPEN and stop that row.</stop>
<output_format>Bridge table (line, amount, source id, tier, recurring or one-time, tag), EBITDA before and after stock comp, cash conversion line, "Checked, normal:" line.</output_format>

<role>Leveraged finance analyst.</role>
<task>From the merger agreement, commitment letters or term sheet (TIER 1) and the EBITDA from prompt 06:
1. Sources and uses to the dollar: purchase equity, refinanced debt, fees, OID; sponsor equity, rollover, each debt tranche actually funded at close, company cash.
2. Tranche table: amount, secured or not, rate (from the document, else ASSUMPTION with a range), amortization, maturity. Commitments that are not drawn at close are listed apart and are not debt.
3. Year by year for {{HOLD_YEARS}} years: EBITDA, cash taxes, capex, working capital, interest, mandatory amortization, sweep, closing debt.
4. Exit at {{EXIT_MULTIPLE}}x: exit equity, MOIC, IRR.</task>
<trap>Counting every commitment as debt. Revolvers, ABLs and bridges replaced by notes inflate leverage; captive receivables facilities are judged against their receivables, not corporate EBITDA, unless they exceed eligible receivables or have corporate recourse.</trap>
<stop>If a funded tranche has no rate in any document, say so and stop the model for that tranche; do not assume it silently. Offer an ASSUMPTION range only after the human confirms none exists.</stop>
<output_format>Sources and uses, tranche table, annual table, exit table. Every figure with a source id and tier.</output_format>

<role>Returns reviewer.</role>
<task>Using the model from prompt 07:
1. IRR and MOIC grid: exit multiple ({{EXIT_MULTIPLE}}x minus 2 to plus 1) by EBITDA growth (-5% to +5% a year).
2. A named downside: the worst EBITDA fall in the company's own filed history, or the sector's last downturn, applied in year 1. State which and why.
3. The EBITDA level at which the equity returns only its money (MOIC 1.0x) and the one at which it clears {{HURDLE_IRR}}%.</task>
<trap>A downside of "EBITDA minus 10%" chosen because it is round tells nothing. Use a fall this business or sector has actually had, with the source.</trap>
<stop>If the model from prompt 07 is blocked or incomplete, do not build a grid on the partial model; say which input is missing and stop.</stop>

<role>Lender assessing whether the cash flow carries the debt.</role>
<task>Using the data source from prompt 01 and the model from prompt 07:
1. At close: total funded debt / EBITDA (before stock comp), senior secured / EBITDA, cash interest cover (EBITDA / cash interest), free cash flow after interest (CFO - capex - incremental interest).
2. Rate paths: base, +200bps, +400bps on floating tranches. Cover and FCF after interest each year.
3. Breakeven: the EBITDA decline at which cover reaches 1.0x, against the downside from prompt 08.
4. Covenants from the credit agreement if one is provided (leverage, cover, fixed charge), headroom each year and the first breach. If covenant-lite, say so and judge liquidity instead.
5. The one structure change that most improves the weakest metric (less debt, more equity, longer amortization), with its effect on IRR.</task>
<trap>A large equity check is not credit support. Lenders are repaid from cash flow; if free cash flow is negative before new interest, no equity percentage makes the debt serviceable.</trap>
<stop>If no rate exists for a floating tranche and no base rate for the decision date is given, stop the rate paths and ask for them.</stop>
<output_format>Metrics table by year and path, breakeven line, covenant table or "covenant-lite: liquidity view", the structure fix with its IRR cost.</output_format>

<role>Model reviewer. You recompute; you do not copy.</role>
<task>Re-derive each figure a second way and compare:
1. Sources = uses, to the dollar.
2. MOIC = exit equity / entry equity; IRR from the equity cash flows, and the check MOIC^(1/years) - 1 for a single entry and exit.
3. Closing debt = opening debt - amortization - sweep, each year.
4. The EBITDA in 07 and 09 equals the EBITDA before stock comp from 06; for a public target, equals the LTM that tie.py passed.
5. Interest = tranche x rate, each tranche, each year.</task>
<trap>Rounding makes clean numbers disagree. Compare at full precision on the same basis; a gap from display rounding is not a break.</trap>
<review_gate>Any gap above 0.5% on the same basis is a FAIL. A FAIL blocks prompt 11: say what to fix and do not write the memo. This rule is for arithmetic only; judgment flags are handled by tags, never here. Keep running the other four checks so no second break hides behind the first. What unblocks it: the input corrected, or a named person explaining the difference in writing with a source.</review_gate>
<stop>If an earlier prompt was not run, mark its checks NOT RUN and stop; never mark them PASS.</stop>
<output_format>Table: check, value A, value B, gap, PASS or FAIL. One verdict line.</output_format>

<role>Deal lead writing for the committee.</role>
<task>Using prompts 03 to 10:
1. The thesis, quoted, and whether the evidence supports it, tested apart from the verdict on the structure.
2. Findings: only CHANGES THE DECISION items in the body, each with the figure, the source id and the working. WORTH A QUESTION items as a numbered list, one item per question, each with the figure it moves, by how much, and who answers it. OPEN inputs in a separate list. EXPLAINED BY CONTEXT once, in one line.
3. Verdict from the allowed list in prompt 01, with one line of reason tied to debt capacity and returns.
4. If the verdict is not PROCEED: the changed terms that would make it one (for example the debt the free cash flow carries at 1.5x cover), quantified.
5. What happens by default if nobody acts on the open questions.</task>
<trap>An early flag that the context later explained still steers a memo written in one pass. Rebuild the findings list from the tags only; a closed flag does not return.</trap>
<stop>If prompt 10 shows an open FAIL, do not write the memo. Say which check and stop.</stop>
<output_format>Two pages at most. Plain English.</output_format>

<role>You build one self-contained HTML page that shows this run's result honestly. You create no figures.</role>
<task>From prompts 03 to 11 in this conversation, build "{{TARGET_NAME}} Deal Screen" with, in order: header (target, decision date, deal type, surface, run date); verdict band (the prompt 11 verdict quoted, and a line if any prompt 10 check failed); returns tiles (entry EV, equity, MOIC, IRR, downside IRR, hurdle); debt capacity tiles (leverage, cover, FCF after interest, breakeven decline); year by year debt table; gate scorecard (tie-out, earnings quality, debt capacity, returns vs hurdle, downside) with status PASS, FAIL, CONDITION or NOT PRODUCED; findings with tags; EBITDA bridge; returns grid with cells below the hurdle marked; questions with who answers each, and OPEN inputs; NOT COVERED (prompts not run, missing documents, every ASSUMPTION, and: "Desk review of the documents supplied. No management meetings, site visits, customer calls, market study, tax or legal work, or audit."); footer with the source tag legend.</task>
<constraints>Every figure exactly as an earlier prompt stated it, with a source tag (P03 to P11). A figure never produced reads "not produced". Two prompts that disagree are shown side by side. No score, grade or probability. One .html file, inline CSS and JS, no network calls, light and dark via prefers-color-scheme, status words next to colours, print clean.</constraints>
<trap>A page that looks complete over a partial run is worse than no page. NOT PRODUCED must be visible, never a blank or a zero.</trap>
<stop>If prompt 11 was not produced, do not build the page; say which prompt to run first.</stop>
<output_format>The HTML in one fenced block. Then: "Save the block above as deal-screen.html and open it in a browser." Then "Next step:" and the one figure a reviewer should challenge first.</output_format>