Space Insurance — The Out-of-Model Signal for Launch-Day Risk
Built for space insurance underwriters, actuaries, and brokers: a sealed, externally-adjudicated launch-forecast record offered as a back-testable,…
Space Insurance — the out-of-model signal for launch-day risk (for underwriters, actuaries, and brokers)
The frame, address-first (never a confession): time is not a river — it is an index, and time is the address itself; the input method is Vedic astrology — Jyotisha — the discipline whose five-thousand-year-old claim is precisely that the address can be read. The desk need not accept a word of that: astrology is the signal, JYOTINT is the instrument, the Brier is the calibration — with Information Yield and SITA completing the readout — and the record is source-agnostic arithmetic that recomputes without trusting the operator. This page is written for the market that prices launch risk — space insurance underwriters, actuaries, and brokers — and it sells a back-test, not a belief: a sealed launch-forecast record built to be joined against a desk's own launch database and re-scored independently.
THE BRIEFING — six acts for the desk, at the top of the page (read it or forward it as-is)
The page now opens with a self-presenting six-act deck, the underwriting twin of /space-force's: Act 01 — the one thing: before liftoff, we tell you what is at risk on the day — the sheet prices the vehicle; the day it flies is priced by nobody; this desk reads exactly that residual and publishes the grade. Act 02 — the claim that was paid: BlueBird 7, the one commercially insured asset across all 23 graded launch days — four sealed calls named the catastrophic payload-loss class, the last ~1 day out; total loss; $30M claim paid per the owner's SEC 8-K. Act 03 — the book that flew unpriced: ~$6.2B flew under the graded calls with exactly one asset covered; Peregrine's ~$97M was milestone-paid regardless — the bill arrived, the delivery didn't. Act 04 — the menu: BACK-TEST IT (the per-launch CSV against your own book, zero cost, zero integration) · PRICE THE RESIDUAL (a named failure class before the window, orthogonal to heritage/weather/range — its premium value is the desk's arithmetic, never ours) · RATE THE FORWARD BOOK (experience rating already moves rates — Falcon 9's cover fell 5.5%→3.1% as its record accrued; a graded day-level input is one more documented rating factor). Guardrails carried verbatim: capacity cycles also move rates; government payloads sit outside this market; nothing promises any input lowers any premium. Act 05 — the forward book: 3,000+ launches/yr planned by 2036 while the whole market's annual launch premium is a few hundred million — every launch not priced by a desk is priced by a taxpayer at zero; that gap is the whitespace. Act 06 — the ask: back-test for free (CSV · verifier · /regrade under your own verdicts); if the record survives, the smallest commercial unit is the buyer-adjudicated Proving Round (/services) — sealed as numbers, graded by your adjudication, fee at risk; a failed forward season gets published here at full weight.
The money on the missions — cost of mission, cost of insurance, what was actually paid
The numbers first, all sourced (unpublished missions counted at zero, no two sourced numbers multiplied into an unsourced product): ≈$6.2 billion of publicly documented mission cost flew under the 23 graded calls. $138M of it was destroyed outright: a $108M NASA lunar delivery (Vulcan Cert-1/Peregrine, /advisory/LA-003) lost as the sealed read warned — the read was PUBLIC five days before liftoff, addressed to the launch team by name, saying verbatim that Stage 1 would perform nominally while flagging risk to 'guided placement of the Peregrine lunar lander'; Stage 1 was flawless, the lander died hours later, and ~$97M of the award had been milestone-paid regardless of the failure (only 10% was success-gated, per NASA/Astrobotic review coverage; the signature panel: The bill arrived. The delivery didn't.) — and the $30M-insured BlueBird 7 written off inside the catastrophic payload-loss class the sealed read named about a day out (/advisory/LA-022), claim paid per the owner's SEC 8-K. Another ~$5.6B flew months-to-years late: Artemis II slipped ~17 months across the sealed windows on programs obligating $4.79B in FY2024 alone (NASA OIG puts each launch at $4.1B — a figure the Inspector General told Congress ‘strikes us as unsustainable’); the ≈$1.5B NISAR slipped more than a year for antenna repairs (GAO logged $30M cost growth); Axiom-4 burned four attempts across 14 days; CMS-03 (₹1,589 crore ≈$225M, the Indian Navy's 2019 contract) flew through cyclone duress with a cryogenic anomaly — both inside the sealed spectra. WHO PAID: an insurer paid exactly once — the one commercially covered asset in the whole set; every other bill landed on a taxpayer or an owner. Self-insured is not no-insurance — it is an underwriter charging itself zero, and Congress does not authorize bottomless pits forever. The rail, restated: no backward prevention claim anywhere — what any read would have changed on any mission is unknowable and never asserted; the reading is proven in both directions — risk days 15 of 15, clean days 5 of 5 — and whether acting on a read prevents the loss is untested; the forward season is that test — prevention is the purpose, warning is the promise, the ledger is the proof. The table states cost, outcome, and payer per mission and leaves the arithmetic to the desk.
The sample-size question, taken head-on (n=23 vs n=0)
n=23 is small by count — but each observation is a sealed, dated, externally-adjudicated call on a named mission, several carrying nine- and ten-figure stakes, in a class whose own actuarial literature calls the data sparse. Conservative credibility math: 15/15 risk-flags realized in-class puts the one-sided 95% binomial lower bound near 0.82; at n=23 the record earns partial credibility — which is why the offer is a back-test plus a fixed-criterion forward season, never a rate input. And the honest comparator is n=0: to this record's knowledge no other public, sealed, pre-launch, graded launch-day forecast record exists from any consultancy, model vendor, or forecaster — a falsifiable claim: if one exists, name it and the page will link it. The choice is not 23 versus more; it is 23 versus none.
The market's own problem
Launch cover is priced from expected loss plus a volatility loading, with “technical heritage” as the credibility proxy, on sparse data — the industry's own actuarial vocabulary. A heritage rate card is static between flights: it is blind by construction to anything that varies launch-to-launch on identical hardware. Per industry market reports, the 2023 loss year ran to roughly $1.43B in claims against roughly $550M of premium, and the insured share of launches has fallen sharply — a capacity pool hungry for discrimination. Trade press has reported a recent launch-vehicle-caused insured total loss; the sealed record covers that campaign as public record (/advisory/LA-019 through /advisory/LA-022 — the last a catastrophic payload-loss class named about a day out). No insurer, broker, or claim party is named on this page; no past prevention is ever claimed. THE FORWARD BOOK: the US Space Force's published vision plans 'more than 3,000 launches per year by 2036' (Objective Force Design 2040, p.62) — today's market prices a few hundred launches a year (~20% insured, ~$550–650M annual premium); the cadence is set by mission demand, not by who carries the risk — at 3,000 a year every day is a launch day, and the bill for a bad one lands on whoever underwrites it: the market where the day is priced, the taxpayer where it isn't. Day-level risk discrimination is worth more at 3,000 launches a year than at 300. Reliability pays everyone: the provider's rates, the insurer's loss ratio, the government's self-insured book.
The signal
23 graded launch calls across five competing providers — SpaceX 8/8 · Blue Origin 5/5 · NASA 4 HIT + 1 NEAR · ISRO 3 HIT + 1 PARTIAL · ULA 1/1 (n=1, shown not argued from) — each SHA-256-sealed and OpenTimestamps-anchored on public infrastructure BEFORE the event, and graded against NASA / FAA / provider statements where a primary source exists. Zero false all-clears; the NEAR and the PARTIAL kept at full weight (both severity overcalls — the safe-side error, self-punished on the ledger). The launch-only Brier is 0.036 — self-scored, small n, and a base-rate baseline ties the aggregate; the checkable claim is SPECIFICITY: the failure mechanism named in advance, 15 of 15 risk-flag calls realized inside the named class, zero bare scrub calls.
Falcon 9 — the cleanest isolation of the day variable (a page centerpiece)
Falcon 9 is the industry's benchmark of reliability — engineering that has removed the machine-side variable more completely than any rocket in history, and the cheapest launch cover in the market, with per-vehicle rates reported falling in trade coverage. Precisely for that reason it is this record's cleanest isolation test: six sealed Falcon 9 reads, six HITs, the reads varying day-to-day on the SAME hardware while reality followed the read each time — including the Ax-4 natural experiment, four different verdicts on the identical stack (same booster, same capsule, same crew) in fifteen days, varying AGAINST the engineering gradient (attempt 2 rated worse than attempt 1 despite the repairs) and grading 4/4. The anchor row: a LOX-leak/engine class sealed ~44 hours before the static fire found a LOX leak (/advisory/LA-007) — the leak was not public information at seal time. On the vehicle where the machine variable is nearly gone, whatever still moves outcomes is day-shaped — the elimination establishes the signal's granularity, not any doubt about the vehicle: even the safest book in the market carries day-structure a heritage rate card doesn't see.
Anomaly-based, not binary
Every row publishes the failure CLASS named at seal beside the realized anomaly (the mechanism ledger, machine-readable in the launchDeep block of /api/v1/corpus-insights.json): a LOX-leak class before the static fire found the leak; hidden reflown-booster degradation before the provider confirmed a full engine replacement (/advisory/LA-020); a catastrophic payload-loss class about a day before the payload was lost (/advisory/LA-022). The industry's own actuarial literature names its desired direction of travel — anomaly-based instead of binary, environmental tracking replacing historical averages, credibility by component behaviour rather than calendar time. This record is already that shape, arriving from outside the model.
What a desk can do with it — the back-test path
The question is not belief — it is whether an out-of-model signal discriminates. That is checkable arithmetic: (1) the per-launch CSV at /dataset/jyotint-analyst-table.csv — one row per call: sealed date, sealed probability, verdict, Brier term, seal hash, per-row verify command — built to join against a launch database and re-score independently; (2) the zero-dependency verifier: node verify-jyotint.mjs --manifest seal-manifest.json --ots seal-manifest.json.ots (exits non-zero on any drift; anteriority rides on the platform publication timestamps before each event — the Bitcoin anchor proves the sealed text unchanged since); (3) /launch-ledger for the verbatim adjudicator quotes; (4) /regrade to re-score the record under the desk's own probabilities, verdicts, and selection. Machine surfaces: /api/v1/graded.json, /api/v1/by-provider.json, /api/v1/corpus-insights.json, /seal-manifest.json (+ .ots), /grading-ledger.json, /dataset/jyotint-sealed-corpus/corpus.jsonl — all CC-BY, CORS-open.
Delay & scrub economics
Scrubs are not claims — but delayed launches break the alignment between written premium, exposure, and reserves, a reserving pain point the industry's own actuarial literature names. A day-level read is native to exactly that horizon: it speaks to schedule-slip exposure per window. And day-level probabilistic inputs are already normal launch culture — every count carries L-1 weather odds; this record is a read on the non-weather residual of the same day.
Engagement — quiet, 1:1
Request the back-test pack via [email protected]. On the operator side the standing posture is already insurance-denominated: a Launch Seat is posted at exactly half the mission's launch-insurance premium (a formula off the buyer's own number, never a sticker), with a clean make-good — a covered call graded MISS earns the next sealed advisory free (see /services). The ceiling, stated plainly: an augmenting input for a named decision-maker — never a go/no-go, never a review-board artifact, never a rate recommendation; nothing on this page is insurance, actuarial, or investment advice; the aggregate score is self-scored and a base-rate baseline ties it — the specificity axis is the independently checkable part.
The live ops floor · launch lens
The page embeds the record's live track-record map scoped to the launch desk — every sealed launch call pinned on the globe it happened over, verdicts adjudicated, replay and lenses driveable, seals verifiable in-map. Not a screenshot: the same live instrument Mission Control runs.
The Underwriting Seat — flagship, Seat №1
Sealed day-reads on the launch book a desk writes — priced in points on the line, graded like the public record: 5% (5 points) of the desk's net line premium per mission read, fixed at commissioning off the broker's indication and owed whether or not the desk binds; floor $50,000; $500,000/yr retainer that read fees credit against; billed on the market's own premium clock (5% at commissioning, 95% at L-30) with re-seals free as T-0 slips; fee-at-risk via a 20% holdback per read earned on the graded outcome — a clean MISS voids it and the next read is free (season variant $1.8M/yr · 24 missions · 25% rebate trigger); per-read seal election — public seal on the anchored record or hash-only private seal, and private reads never enter the public score. Governed by TERMS_02INS; the desk remains the underwriter of record; never a rate, bind, decline, or reserving recommendation.
Thirteen on-page instruments, every number from the published data files
Detection matrix (15/15 risk-flags realized in-class · 5/5 GO reads clean · 0 false all-clears · the two severity overcalls marked and kept) · lead-time distribution across the 23 calls · calibration profile (Spiegelhalter Z −0.48, Cox slope 0.96, bias-in-the-large −0.018 — whole record, n=92, /api/v1/measurement-profile.json) · launch-only Brier 0.036 beside the whole-record Murphy decomposition (uncertainty 0.1018 − resolution 0.0204 + reliability 0.0143 = 0.0958; no launch-scoped decomposition is published, so none is drawn) · the honest-tie figure (the record's own limitations file concedes a base-rate baseline ties the aggregate; the divergence axis is information per call — median 6.8 bits, launch median 19.9) · Ax-4 divergence chart · dose-response curve with gradient-breakers marked · anti-con curve (~1 → ~12 falsifiable elements per call while accuracy held) · acknowledgment-gap bars (sealed 4 days before NASA's Artemis-II confirmation, 2 days before Blue Origin's engine-swap confirmation) · confidence-ladder map (the sealed phrases' frozen probabilities, /grading-ledger.json) · premium-clock timeline · CSV join-schema · an in-browser verifier that recomputes a seal without anything leaving the tab.