Insurance — Stage 0 re-score, 2026-09-22

Insurance — Stage 0 re-score, 2026-09-22

Slot cluster-insurance-rescore. S0 only. Drafted by Claude in an unattended cloud routine; nothing here is human-verified. Every figure carries the filing it came from; background knowledge is marked [background] and search-result evidence [search-summary].

This file supersedes the S0 section (§3) of cluster-insurance-2026-09-22.md and nothing else. That file's S4 (industry & supply map), S6 (scuttlebutt) and S7 (thesis) stand and continue to carry forward unchanged; it is dated and has not been edited (ledger/research/README.md).

Seven members: ACGL, KNSL, ALL, AIG, LNC, FAF, AEG. RYAN was removed from this slot when it was re-tagged fee_intermediaries (WATCHLIST_PLAN.md §22) and is not scored here.

Why this slot exists. WATCHLIST_PLAN.md §20 measured how Stage 0 had actually behaved and §21 took two decisions: S0 becomes a lens that never stops a name, and risk-carrying balance sheets get their own four tests (SKILL.md §4.3). The superseded file scored all eight members on distribution lock-in, unscrapeable data, liability absorption and a reworded fourth test — and, as §4.1 now records, every insurer absorbs liability and none owns its channel, so the instrument returned nearly the same answer eight times and none of it was about whether these are good insurers.

---

§0 — How this pass was run, and what it can be checked against

Instruments, one per member, chosen by business model before scoring (SKILL.md §4.1):

The three re-tags were written into watchlist.yaml before any member was scored, as §4.1 requires. insurance is now the four names that underwrite a P&C book and nothing else.

Sources. Rung 1 (SEC EDGAR) answered everything: data.sec.gov/api/xbrl/companyfacts for every series below, and eight filings read directly — the FY2025 and FY2023 10-Ks of ACGL, KNSL, ALL and AIG, plus LNC's and FAF's FY2025 10-Ks and Aegon's FY2025 20-F. Rung 5 (ROIC.ai) for prices. One Equibles call, spent on a reachability check, not on data.

How quotes were taken. Each filing's HTML was converted to text with whitespace collapsed, and every quoted span below was located in that text by exact substring match before it was written here; all eleven were re-verified mechanically in one pass before this file was committed. No quote in this file was typed from memory (ledger/research/README.md).

Two method notes that matter for reading the scores:

1. The ex-development combined ratio is used as an upper bound on the ex-cat ex-development ratio. Removing catastrophes lowers a combined ratio, so a year that passes §4.3's 95% threshold after adding prior-year development back is a year that passes it with catastrophes also removed. Where the issuer publishes the ex-cat ex-development figure itself (ALL, AIG) that figure is used directly and both removed numbers are stated, as §4.3 requires. ACGL and KNSL do not publish a five-year series of it, so the upper bound is used and the disclosed catastrophe load is stated alongside.
2. Paid-to-incurred drifts upward mechanically on a fast-growing book, because paid claims lag incurred ones and the ratio walks toward 1 as the book seasons, with no reserve problem whatever. It is read here as diagnostic only where premium growth is stable, and every reading says which case it is.

---

§1 — The finding: the instrument changes the answer, and in one direction

Read the table as two results, not one.

First: the four P&C underwriters separate from the two spread writers completely. The old instrument put ACGL (2/4) below KNSL (3/4) and both above ALL (1/4) and LNC (1/4), interleaving a disciplined reinsurer, a compounder, a personal-lines carrier at peak earnings and a life company with a reinsured back book. The new one puts the four underwriters at 3–4/4 and the two spread writers at 0. That separation is the instrument working: the questions §4.3 asks — are the reserves honest, does the edge survive the cycle, is the capital theirs, does the book grow without issuing shares — are answerable about an underwriter and are answered differently by a spread business.

Second, and this is the part worth the owner's attention: the axis that discriminates is reserve honesty, and it does not point where the price does. Test 1 produced the only 0 among the P&C names, and it produced it for ALL — the cheapest of the four on trailing earnings. Test 2, by contrast, discriminated nothing: all four pass it, including Allstate. On a watchlist already filtered to large surviving carriers, an ex-cat ex-development combined ratio below 95% is table stakes; where the honest earnings are is not.

The three new facts this pass produced

(1) Arch's clean development record is mostly mortgage, and mortgage is the segment that just went through zero. The FY2025 10-K breaks favourable development out by segment and tail length (accession 0000947484-26-000017, PYD-by-tail table, $m, favourable shown negative):

Two things at once. The mortgage segment — whose loss ratio went −1.2% to +6.5% year on year in Q2 2026, i.e. through zero, the finding the superseded file flagged and nobody was watching — has been supplying between two-fifths and two-thirds of the group's favourable development. And the long-tailed lines have been adverse in every one of the three disclosed years, so the direction of travel inside the release is already the wrong one. names-insurance-1 found the same shape one quarter further on ($437m short-tail favourable against $48m long-tail adverse over H1 2026, from $355m/$5m). Test 1 still scores 1 — five straight favourable years is five straight favourable years — but the composition is the thing to re-read next year, not the total.

(2) Allstate did a reserve round trip inside one cycle, and the release landed in the record year. Prior-year reserve re-estimates, XBRL SupplementalInformationForPropertyCasualtyInsuranceUnderwritersPriorYearClaimsAndClaimsAdjustmentExpense ($m, adverse positive): 2021 +122 · 2022 +1,741 · 2023 +549 · 2024 −308 · 2025 −1,809. Three adverse years, then the largest release in the cluster — worth 3.1 points of the Property-Liability combined ratio on Allstate's own line (FY2025 10-K, 0000899051-26-000031) — arriving in the year net income hit $10,282m, its best in six by more than 2x. The superseded file called Allstate "the cheapest-looking name in the cluster on trailing earnings, at the point in the cycle where trailing earnings are least informative" and was right; this is the number that says so.

(3) AIG's development record is net of a contract designed to absorb development. The FY2025 10-K's own footnote to the General Insurance ratio table:

"excludes net loss reserve discount and the portion of favorable or unfavorable prior year reserve development for which we have ceded the risk under retroactive reinsurance agreements and related changes in amortization of the deferred gain" — AIG FY2025 10-K, accession 0000005272-26-000023, General Insurance results table, footnote (a)

Everything in AIG's published combined-ratio series — including the accident-year ratio "as adjusted" that scores its test 2 — is measured after the adverse development cover has taken its share. That does not make the number wrong; it makes it a different number from the one the same label carries at Arch or Kinsale, and the cross-sectional comparison in the superseded file's §5 should not be read across that line.

---

§2 — ACGL, Arch Capital Group · insurance · ✅ lens applied — 4/4

Instrument: §4.3 risk-carrying balance sheets.

Where test 3 and test 4 are softer than the 1s look. The 25% PML tolerance is above §4.3's ~20% reference and is a limit, not a measurement — the 10-K does not state the modelled figure, so the leg is a policy, not a number. And H1 2026 changes the funding picture: names-insurance-1 read $1,977m of new borrowings against zero a year earlier, senior notes from $2.7bn to $4.3bn, and $1,949m of buybacks with equity falling over the half on $2,104m of earnings. Buying stock at ~1.4x book with borrowed money is a capital-allocation call, not a reserve problem, and it belongs in S7 — but it is the reason test 4's 1 is a 1 about the five-year record and not about this year.

Cycle question — where is this line, on a named measure, and does the case need it to stay? Arch's own Q2 2026 8-K (0000947484-26-000118, carried forward from the superseded file) gives the measure: reinsurance combined ratio ex-cat and ex-development 77.2% → 79.9% while the reported ratio improved 78.5% → 77.5%; insurance 90.6% → 91.6%; group 80.9% → 82.5%. Gross premium flat at +0.2%, net written −10.4%, net earned −12.8%. The line is softening and Arch is declining business somebody else is writing. The case does not need the cycle to stay — this is the member that compounds through it — but it does need the reserves to keep behaving, which is §1(1).

The AI line (one sentence, not scored). Cheap intelligence entrenches a multi-segment underwriter with thirty years of its own loss experience far more than it commoditises one; where it bites is the expense ratio of the brokered layer between Arch and the risk.

Falsifier — one number. The group ex-development combined ratio of 86.3% plus the loss of the mortgage release (1.4 points of 2025 net earned premium) plus development normalising toward zero (3.5 points) puts the reported ratio at ~91 with no catastrophe. Two more years of the 2026 trend crosses 100.

Liquidity. ADV $341m/day (21 sessions to 2026-09-21; carried forward from cluster-insurance-2026-09-22.md §3, where the window and method are stated. ACGL's own scorecard records $349.7m on the same window — the same figure, rounded twice). Last price $95.54 (2026-09-22, ROIC NASDAQ:ACGL). P/B 1.48 on the file-wide basis below.

---

§3 — KNSL, Kinsale Capital Group · insurance · ✅ lens applied — 4/4

Instrument: §4.3 risk-carrying balance sheets.

The 2025 detail that cuts the other way, and where it is priced. The 10-K attributes the year's underwriting-income increase in part to development:

"The increase in underwriting income was primarily due to continued growth in the business and higher favorable development of loss reserves from prior accident years offset in part by higher catastrophe losses incurred." — KNSL FY2025 10-K, accession 0001669162-26-000015, results-of-operations discussion

Favourable development rose 67% in 2025, to 3.99 points of net earned premium from 2.79 — its largest contribution of the five years — in the year gross written premium growth fell to +5.7% from +19%. That is the cluster's central mechanism (reported flattered by releases) appearing at the member with the cleanest reserves. It does not move test 1, whose record is unambiguous; it belongs in the falsifier, and it is there.

Cycle question. Named measure, from names-insurance-1's reading of the Q2 2026 10-Q: gross written premium −5.0% in the quarter and −2.9% over the half — the first decline on the public record — with the Commercial Property Division −32.7% and average premium per policy −14%; net written held nearly flat only by lifting retention 82.6% → 85.8%, which took the expense ratio 20.7% → 21.7%. The case needs the cycle, or at least needs the E&S submission flow: Kinsale has never underwritten through a soft E&S market. Its entire public record, 2020–2025, is a hardening one.

The AI line. Small-account E&S is the one underwriting niche where cheap intelligence most plausibly widens the gap — Kinsale's advantage is cost per policy on classes too small for a human-underwritten admitted carrier, and that is exactly what gets cheaper.

Falsifier — one number. The expense ratio. It is the scoring point behind test 4 and it is being spent to defend the line: 20.7% → 21.7% as retention rose. An expense ratio through 22% with gross premium still falling says the moat is being consumed to hold volume.

Liquidity. ADV $70m/day — the thinnest of the four underwriters, and the constraint S8 must respect. Last price $354.59 (2026-09-22). P/B 4.22 — by a distance the most expensive name here.

---

§4 — ALL, The Allstate Corporation · insurance · ✅ lens applied — 3/4

Instrument: §4.3 risk-carrying balance sheets.

What the re-score actually says about Allstate, and it is not what the old 1/4 said. The underwriting is the best it has been — 79.7% ex-cat ex-development, improving 16 points in two years. The reserves are the least trustworthy record among the four underwriters. The old instrument reached a 1/4 by observing that personal auto is price-transparent and the captive channel is losing share; both true, neither decisive. This instrument says: the machine works and the earnings you are being shown include a release from the years it did not.

Cycle question. Named measure: auto rate. The board carries it from a sell-side P&C analyst on 2026-09-07 — rates raised ~60% against claims inflation of ~40%, now reversing, expected to persist through end-2027 (normalized_end_date: 2027-12-31 on the claim row), with "the personal lines subgroup, which I've not been more negative on in my career" (Ryan Tunis, quote located in candidates/eisman-playbook-2026-09-07-p-c-stocks-worth-owning-the-ai-hedge-with-01-pc-personal-lines.yaml, 13:57, and carried forward from the superseded file §3 rather than re-located here). The case needs the cycle to be wrong about it: earnings are at a peak and the rate cycle is turning down.

The AI line. Telematics and claims automation are where cheap intelligence lands hardest in personal auto, and both are available to the direct channel Allstate is losing share to — so on this axis it is the one member where cheap intelligence plausibly helps the competitor more.

Falsifier — one number. Prior-year development turning adverse again. The 2022 round trip (+$1,741m) is the proof it can, and it would arrive while the exposure base itself shrinks: the same claim row records ADAS in ~40% of new vehicle sales against 15% three or four years ago. A shrinking exposure base does not mean-revert; a release cannot be repeated.

Liquidity. ADV $422m/day — the deepest of the seven. Last price $235.96 (2026-09-22). P/B 2.06 — see the basis note; Allstate's preferred is not tagged at a current date, so this flatters the common slightly.

---

§5 — AIG, American International Group · insurance · ✅ lens applied — 3/4

Instrument: §4.3 risk-carrying balance sheets.

The 0 is the point of running this instrument on AIG. The standard four scored it 2/4 for being broker-intermediated in a commoditised segment; the superseded file then spent its best paragraph on the shrink-and-buy-back machine and had nowhere to score it. §4.3 has somewhere: the buyback is larger than the cash the underwriting produces, and the per-share arithmetic (premium per share $36.2 → $41.6) is being manufactured by the denominator, not earned by the numerator.

Cycle question. Named measure: AIG's own accident-year combined ratio as adjusted has been flat at 87.7 → 88.2 → 88.3 for three years while the catastrophe load fell 4.3 → 5.0 → 3.9 points — i.e. the underlying margin is not improving, and the reported improvement is mix and weather. The superseded file's S6 places AIG in large-account commercial, the segment the board's analyst calls the most exposed to soft pricing. The case needs the self-help to continue, and the self-help is finite.

The AI line. Large-account commercial underwriting is judgement-heavy and relationship-priced, which is the profile cheap intelligence erodes slowest — but it is also the segment where the broker, not the carrier, would capture the saving.

Falsifier — one number. The accident-year combined ratio as adjusted crossing 90 while the buyback continues to run the equity base down. At 88.3 that is two years of the 2024–25 drift.

Liquidity. ADV $310m/day. Last price $74.31 (2026-09-22). P/B 1.03 — the only P&C member near book.

---

§6 — The three re-tagged names

LNC, Lincoln National · re-tagged life_annuity · ✅ lens applied — 0/4

Instrument: §4.3, spread branches. It carries annuity risk on its own balance sheet, so §4.3 is the right instrument; it is not in the P&C cycle, which is why it left insurance.

Cycle question. The spread cycle is rates and policyholder behaviour; the named measure is the 2.09x → 1.62x above. The case needs rates to stay high and lapse behaviour to stay benign, which are not independent.

The AI line. Underwriting mortality is the one place in this cluster where the proprietary data is genuinely shared — through reinsurers and consulting actuaries — so cheap intelligence changes little here; it changes distribution economics for independent annuity sellers instead.

Falsifier — one number. Reinsurance recoverables ÷ equity through 3x, or a downgrade at one of the reinsurers behind the ceded back book: at 2.57x this is a credit position first and an insurance company second.

Liquidity. ADV $87m/day. Last price $41.465 (2026-09-22). P/B 0.71 — the cheapest book in the file, which is the market's verdict on the four rows above. The preferred is not tagged at a current date, so this flatters the common.

---

AEG, Aegon Ltd. · re-tagged life_annuity · ✅ lens applied — 0/3

Instrument: §4.3, spread branches. Foreign private issuer, IFRS 17; read from the FY2025 20-F (accession 0001193125-26-124987) and the Form 425 of 2026-08-20 (0001193125-26-357719, figures carried forward from the superseded file §3).

Cycle question. Rates and the US annuity market, on the named measure of the group solvency ratio above. The case needs the redomiciliation to be the capital event management says it is, because the Solvency II series is falling and the Bermuda series does not have a history yet.

The AI line. Aegon's one genuine asset on the old instrument — World Financial Group, "now exceeding 100,000 agents" (Form 425, "Our performance", p.4, carried forward) — is a distribution stack whose per-agent economics cheap intelligence makes cheaper for everyone, including the direct sellers it competes with.

Falsifier — one number. The group solvency ratio through 150%, or a buyback suspension. Either would confirm the third row above is capital depletion rather than capital return.

Liquidity. ADV $40m/day on the NYSE ADR — the thinnest of the seven. Last price $9.03 (2026-09-22). P/B 1.30 at EURUSD 1.14673 (ROIC, 2026-09-21, carried forward). IFRS 17 equity excludes the contractual service margin, so this is not comparable with the P&C members' P/B, and the superseded file's cross-section should not be read across that line either.

---

FAF, First American Financial · re-tagged housing_volume · ✅ lens applied — 3/4

Instrument: the standard four, in the account skill's own words (§4.1). housing_volume has no variant, and §4.3 is the wrong instrument for a business whose claims are 4.4% of revenue — but it is run as a second reading below, because FAF does carry the risk and the reading is informative.

Reference category (§4.1 requires naming it). The nearest entry on the account skill's Tier 1 list is exchanges, clearing houses and reference data: FAF is reference data with an indemnity attached and a statutory requirement behind the demand. It is not a clean fit — an exchange's volume is its own network's, FAF's volume is the mortgage market's — which is exactly why the re-tag to housing_volume is the substantive change here and the 3/4 is the incidental one.

Surplus question — who keeps it? Not the company: the mortgage market decides. XBRL net income $1,241m (2021) → $263m → $217m → $131m → $622m (2025) — a 9.5x peak-to-trough swing with no change in competitive position, which is a volume business wearing an insurance licence. What the company does keep it does not dilute away: diluted shares 111.4m → 103.7m, with buybacks of $122.3m in 2025 against $950.8m of operating cash flow.

§4.3 as the second reading — recorded, not scored. Prior-year development +19 (2021), +18 (2022), −18, −37, −44 $m: adverse in 2 of 5, so §4.3's test 1 would score 0. Paid-to-incurred 0.819 / 0.893 / 1.135 / 1.243 / 1.097 — above 1.0 for three straight years, on a reserve balance that has fallen every year since 2022 ($1,325.3m → $1,169.6m). Read as an underwriter that would look like reserve run-off; read as a title company it is what a volume trough looks like from the claims side, and the 2023–25 recovery in revenue is the reason the reserve stopped growing. That divergence is the argument for the re-tag, stated as a number.

Falsifier. Regulatory: a change to the requirement that lenders hold title insurance. The superseded file's §3 carries the evidence (proposals to cut title-insurance requirements, FAF publicly opposing) as [search-summary] and this pass did not improve on it. Secondarily, the commercial order stream: management guides commercial title to a possible record year on data-centre and energy transactions [search-summary] — which makes FAF the one name in this repo that is long AI capex without owning a chip, a datacenter or a turbine.

Liquidity. ADV $59m/day. Last price $69.568 (2026-09-22). P/B 1.31.

---

§7 — What the owner should act on

1. The instrument is right and the axis it found is reserve honesty. Three of the four underwriters pass it; the one that fails is the one trading at the lowest trailing multiple. That is a usable output and the old instrument could not produce it.
2. §4.3's test 2 does not discriminate on this watchlist. All four P&C members clear 95% on the ex-cat ex-development combined ratio in 4–5 of 5 years, including Allstate at 79.7%. It is doing its job — it is simply that a watchlist of large surviving carriers has already passed it. Tests 1 and 4 are where the information was.
3. §4.3 has a measurable gap: the spread branch of test 2. It asks for an earned-yield-minus-credited-rate series that an IFRS 17 filer need not publish, and Aegon does not. The proxy used for LNC (net investment income ÷ interest credited, 2.09x → 1.62x) worked and is computable from XBRL for any filer; adding it to §4.3 as the named fallback would close the gap. Until then, scoring it 0 would assert a structural absence from a disclosure absence, which is why this file marked it n/a and said why.
4. Two clusters now have no cluster pass of their own. life_annuity (LNC, AEG) and housing_volume (FAF) were created by this slot's re-tags. Their S4, S6 and S7 still carry forward from cluster-insurance-2026-09-22.md — a file whose single value chain is the P&C cycle and which says of itself that it only works by labelling where each member leaves it. The S0 seam is fixed and the S4/S7 seam is not. The nearest precedent is §22's split of cluster-financials-global: two new cluster slots ahead of names-insurance-3, which is the slot that will otherwise close FAF and AEG against P&C industry stages. This is the one item that needs an owner decision.
5. Arch's release is a mortgage release (§1(1)), and the mortgage loss ratio has already gone through zero. Anybody carrying "five straight favourable years" forward should carry the composition with it.
6. AIG's series is net of the adverse development cover (§1(3)), by the filing's own definition. The superseded file's cross-sectional valuation table compares AIG's combined ratio with Arch's and Kinsale's; those are not the same measurement.

---

Scorecard — slot cluster-insurance-rescore

S0 Universe & fit          ✅ lens applied to 7 of 7 — ACGL 4/4, KNSL 4/4, ALL 3/4, AIG 3/4,
                              FAF 3/4 (standard four, §4.1 wording), LNC 0/4, AEG 0/3.
                              Instrument per member chosen before scoring; LNC, AEG and FAF
                              re-tagged in watchlist.yaml first (§0). Four tests + the cycle
                              question + the AI line + the falsifier + liquidity written for
                              every member. Glyph records that the lens ran, not its verdict
                              (SKILL.md §4, owner decision 2026-09-22). No member is dropped.
S4 Industry/supply map     ⬜ not this slot — stands at ✅ in _clusters/cluster-insurance-2026-09-22.md
S6 Scuttlebutt             ⬜ not this slot — stands at ✅ 8/8 in the same file
S7 Written thesis + test   ⬜ not this slot — stands at ✅ in the same file, conviction ⏳ OPEN (user)

Owner-only items, unchanged and still open: S6 product test and expert call, S7 conviction, S8 buy price and size — all ⏳ OPEN (user).

P/B basis, one basis for all seven so the column is comparable: 2025 fiscal-year-end total stockholders' equity (or equity attributable to owners, IFRS) ÷ that year's weighted-average diluted share count, against the 2026-09-22 last price from ROIC.ai. This is not the basis in cluster-insurance-2026-09-22.md, which used 2026-06-30 book, so the two files' P/B columns are not comparable with each other. ALL's and LNC's preferred stock is not tagged at a current date, so both flatter the common slightly.

A tooling note for whoever reads prices next. ROIC.ai's 2026-09-22 daily bars were fetched at ~16:37 ET, after the regular session closed (Equibles GetMarketStatus), but they carry a fraction of a normal day's volume — 224,417 shares for ACGL against an ADV implying ~3.6m. They are treated here as late-session prints, not settled closes, and the ADV figures are carried forward from the 2026-09-21 window rather than recomputed on them.

§8 — Evidence that was thin, and why

• No transcript for any member. An S0 re-score does not need one, and names-insurance-1 already recorded that Equibles was down for ACGL/KNSL and that fool.com has no Q2 2026 page for either. Equibles was reachable this run (one call, spent on GetMarketStatus) — worth knowing, since the previous slot got a 502.
• KNSL's 2021 combined ratio was not extracted. Its FY2023 10-K carries two years, so test 2 is scored on four measured years rather than five. The four that were measured are 78.5 → 75.9 reported and 83.0 → 79.9 ex-development; a fifth year is very unlikely to change the score, and the file says it was not read rather than implying it was.
• Rating actions were not swept for any member. §4.3's test 3 asks for "no rating downgrade or negative outlook in three years" and no run has checked it. ACGL's and KNSL's 10-Ks name their A.M. Best ratings as of a date; none of the seven was checked against three years of agency actions. Every test-3 score here rests on the ratio legs only, and each one says so.
• Catastrophe PML is disclosed on three different bases and not at all by two members. ACGL gives a 25%-of-tangible-equity tolerance and no modelled figure; ALL gives a 1-in-100; AIG gives 1-in-250 and 1-in-100 both; KNSL and the three re-tagged names give no number. §4.3's "1-in-250 PML below ~20% of equity" is therefore comparable across two of seven.
• ALL's paid-to-incurred could not be computed. PaymentsForLossesAndLossAdjustmentExpense and the Schedule-P current-year tag are absent from Allstate's XBRL after 2020, so test 1's second leg is unmeasured for the one member that failed the first leg.
• AEG's assumption-review charge is not separately quantified (§6), and its test 2 is not measurable at all (§7 item 3).
• Four DEF 14As remain unread, the seventh consecutive slot. Proxies and incentives is still the one lens with nothing behind it anywhere in this cluster.
• No new scuttlebutt, no new industry map. By design: this slot is S0 only.