How the named-credit review was built
- Universe: every named credit in AG's 2Q26 Financial Supplement (top-50 US exposures, top-50 non-US exposures, top-25 US structured, full BIG list, 300 structured transactions — deduplicated) plus sector aggregates from the AGL Q2 2026 10-Q (filed 8/7/2026).
- Method: stratified sampling, rating-agency style. A-rated-and-above performing names get a one-line verdict from filing rating + sector. Deep research (current news, rating actions, refinancing walls) only for BBB-and-below, all BIG names, and the known-stressed list.
- Traffic lights are reviewer judgment, not agency ratings.
- The CUSIP boundary — stated upfront: AGO discloses named large exposures, not literally every CUSIP. No public filing gives the full CUSIP list. "Line by line" here means every disclosed credit plus stratified sector coverage of the undisclosed tail.
How the CUSIP tape was built
- Source: primary offering documents (official statements) filed on EMMA (emma.msrb.org), found via web-index discovery across all 50 states, DC, and territories, targeting insurer-language variants ("Assured Guaranty Municipal Corp", "Assured Guaranty Corp", "Assured Guaranty Inc.", "Financial Security Assurance", "municipal bond insurance policy").
- Acceptance rule: a document counts only if it contains affirmative bond-insurance language naming an Assured entity as insurer of the bonds being offered. Specimen policies in unrelated documents, merger notices, and continuing-disclosure filings that merely mention AGM are rejected or flagged, not counted.
- CUSIP extraction: full 9-character CUSIPs from CUSIP blocks, or base CUSIP + suffixes from the inside cover. Every candidate must pass the CUSIP check-digit algorithm. Extraction is limited to the first 30 pages (inside cover / maturity schedule) so escrow holdings, investment schedules, and refunded-bond tables in appendices are excluded. Base+suffix CUSIPs must sit within 200 characters of a maturity year.
- Corporate succession: AGM merged into Assured Guaranty Inc. effective August 1, 2024; legacy FSA policies were assumed by AGM in 2009. The tape preserves original insurer entity as named in the document and maps to the current entity.
- Partial-maturity insurance is detected and recorded (
insurance_scope = full / partial-range / partial-listed / partial-series); per-CUSIP insured_likely (yes / no / uncertain) is an analytical aid, not a verification. No partial-scope CUSIP is published as verified wrapped without manual maturity-schedule review.
- Deduplication: the same CUSIP in multiple documents → one row, all sources cited. Refundings are flagged; refunded (defeased) CUSIPs are not current exposure.
- Validation sample: Cedar Rapids CSD 2022 → 10/10 expected CUSIPs; Murrieta Valley USD 2020 → 19/19 with zero leakage from a 244-page appendix (261 unrelated CUSIPs correctly excluded); CommonSpirit 2025A, PREPA Series DDD, Yankee Stadium 2020, Citi Field 2021, Cedar Rapids 2022 all validated case-by-case.
Reinsurance — why book totals are already net
- AGO's "Net Par Outstanding" is, by definition, net of reinsurance cessions (AGL fourth-quarter 2025 Financial Supplement glossary). The $281.4B book total, the $38.66B tape cross-reference, and every figure in the per-share scenario table and Stress Lab are therefore post-cession.
- Most ceded par never leaves the group. Assured Guaranty Inc. (statutory, 2025): gross par $272.3B → net par $211.4B, with $60.8B ceded to affiliates (AG Re / AGRO Bermuda — intra-group capital management) and only $0.1B to unaffiliated third parties (AG fourth-quarter 2025 Financial Supplement).
- Genuine third-party risk transfer is immaterial. AG Re cedes roughly $0.5B of financial guaranty par and $0.4B of specialty exposure to non-affiliates (2025 GAAP financials) — a rounding error against the $281.4B book.
- The history explains the size of the affiliated number. In 2011 AGO had ~$56B ceded to third parties (Radian, Tokio Marine, Syncora, ACA, Ambac and others); most was commuted after the financial crisis or absorbed when AGO acquired the counterparties. The cessions that remain are treaty/quota-share structures disclosed only in aggregate by counterparty type — never per policy — so no public source lets us attribute a ceded share to any individual CUSIP.
- Consequence for this site: tape CUSIPs record wrapped (gross) amounts from offering documents; book-level figures are net. The difference is de minimis at the group level because third-party cessions are ~$0.5B, but never read a single issue's original par as AGO's current retained exposure.
Hard limitations
- EMMA covers US municipal securities only. It cannot see UK water (Thames, South East, Yorkshire), the Channel Tunnel, Sydney Airport, US structured products (CLOs, TruPS CDOs), private/144A transactions, or secondary-market wraps without a filing. Those credits are marked "no public indenture — private/144A or non-US" in the explorer.
- OCR-poor and scanned documents cause false negatives — the tape undercounts by construction.
- Original issue par ≠ current AGO net par. Bonds mature, are called, or are refunded; AGO cedes reinsurance. Harvested par is labeled original par at issuance and never equated to current net par.
- Pre-2009 legacy documents are thinly indexed — FSA-era coverage is partial.
- EMMA shows issuance, not AGO's retained net par after reinsurance, refundings, or amortization.
- Verification statuses (
extracted-unreviewed and others) are shown on every row — treat unreviewed rows as machine-extracted, not manually confirmed.
- Core filings: AG 2Q26 Financial Supplement (par, internal ratings, geography/sector, BIG list, £1.8bn Thames disclosure); AGL 10-Q Q2 2026 (sector table, Note 5 expected loss, Note 6 loss & LAE, BIG tables, Thames £1.8bn disclosure); AGL Q2 2026 8-K (PVP, buybacks, ABV); Q2 2026 Fixed Income Presentation (CPR $9,978M); Q1 2026 Equity Presentation (BIG >$250M detail).
The disclosure layer and the tape snapshot
- Disclosure layer: the site's Disclosure page renders 125 named table exposures and 300 structured-finance transactions transcribed from the 2Q26 Financial Supplement, with 47 logged discrepancies against the Q2 10-Q and equity/fixed-income presentations. Combined named totals ($92.269B) deduplicate to a $86.994B lower-bound — 30.9% of the $281.4B net-par book; the remaining $194.406B is the undisclosed residual.
- Tape snapshot: 167 issues, 2,405 CUSIPs, $38.349B cumulative original-issue par. Every issue links to its primary source — the official statement bundled in
docs/ where available; where no durable PDF survives, the issue cites the official EMMA issue record instead. Missing PDFs are downloaded from EMMA at build time and %PDF-validated; failures are explicit, never silent.
- History: 9,888 rows from 70 quarterly supplements (AGL 2020Q1–2026Q2, AGC 2020Q1–2026Q2, AGM 2020Q1–2024Q2), zero parse failures. AGM's series ends honestly at 2024Q2 — the supplement changed format, and the gap is shown, not hidden.
- Build:
build_site_data.py reads the canonical CSVs and the validated named registry at build time; all site JSON is regenerated from source on every deploy. The site never edits raw data by hand.
- What's not on the tape yet: Pittsburgh Airport 2021A/B, Anaheim PFA 2021A, MWRA 2007A/B — confirmed official statements held for the next wave, not admitted to the canonical tape this round.
Still-open verification items
- Thames exposure: 2Q26 supplement reports $2,365M at internal rating B (the Q2 10-Q’s £1.8bn figure is the same exposure at a different FX cut) — the site uses the supplement’s $2,365M/B.
- Palomar forbearance status after January 2026; identification of the reported $290M CCC tranches.
- Brightline RSA effectiveness and legal status; primary confirmation of the 44¢ recovery source and its assumptions.
- Jackson/Harrisburg claim-payment history; Dartford listing-suspension reason.
- Middle-market CLO attachment points; composition of ~$2.5B structured "Other."
- Whether Fitch's September 2026 Alameda action was a downgrade or watch-negative.
- Reconcile September 30, 2026 AGO loss measures: $113M net expected loss to be paid vs $294M net expected loss to be expensed vs $286M future loss and LAE — not interchangeable.
- The assertion that S&P's capital model explicitly credits non-acceleration needs sourcing before public reuse.
About this site
Independent analytical exercise prepared September 2026 from public filings and press. Not a rating action, not investment advice. Scenario math is illustrative, not a forecast. Data files (data/credits.json, data/cusips.json) are rebuilt from the deck sources and the EMMA tape by regenerate_data.py — see REDEPLOY.md for the refresh runbook.