Moluscca builds the pricing layer for risks the world hasn't figured out how to price yet. Too new, or changing too fast. Either way, someone has to do the math.

The Problem

Insurance runs on the past.
The world stopped repeating itself.

Every insurance price is a bet that tomorrow looks roughly like yesterday. That works until the thing you are pricing has no yesterday at all. Autonomous AI systems and private space operators are both carrying real, often legally mandated exposure that no one can currently put a fair number on.

Enterprise AI deployments, 2025
0 years
of loss history exist for damage caused by an autonomous agent acting on a company's behalf. Underwriters have nothing to price it with, so most say no.
PSLV-C62, January 2026
16 payloads
lost on one launch failure. None of the private payloads carried insurance, in a market where cover is a licensing precondition.
US flood counties, 2025
2.6%
of homes in counties that received flood payouts actually had flood coverage. The same shape of gap, in a market that has had data for decades.

One. Risks getting worse faster than the models.

Insurers price flood risk on a "1-in-100-year" event. That number is the whole foundation of the price. Except those events now arrive in clusters. When the impossible happens twice in a decade, the model isn't conservative anymore. It's fiction.

And insurers don't just raise prices when the math breaks. They leave. Whole towns, whole coastlines, quietly uninsurable.

Modeled loss curve vs. observed events · the tail is where the model dies
Loss history available at pricing time · new risks start at zero

Two. Risks that are brand new, with no past at all.

An AI agent with write access to a company's systems takes a wrong action: sends the wrong payment, deletes the wrong record, makes a call no human approved. The company is liable. The loss is real. But there is no actuarial history for "agentic AI causes damage," because the category is younger than the policies meant to cover it.

A private company puts a satellite on a rocket and it fails on the pad. A small business faces a threat that didn't exist five years ago. Same pattern: the exposure is real, the cover is often expected or required, and the insurer has nothing to price it with. So the default answer is no.

Both directions, same failure

The price and the reality have come apart. Overpriced, so people go uncovered. Underpriced, so insurers get wiped out and flee. Or refused outright, because no one knows the number. Global flood losses passed $42 billion in 2025. Most of it was never covered.

What We Do

Score. Explain. Deliver.

We build the missing pricing layer, and hand the number to whoever carries the risk. We are not an insurer. We don't carry the risk. We make it pricable.

01

Score

We turn a risk question into a defensible number. Frequency and severity modeled separately, because averaging them together hides the loss that ruins someone.

02

Explain

Every score carries its reasons in plain language. If an underwriter or a regulator can't follow the logic, the score isn't finished. Never a black box.

03

Deliver

Reports, dashboards, and data feeds for the people already pricing this risk blind: the insurer, the broker, the operator who needs cover to launch.

Honest about thin data

New risks have short histories. We use credibility methods that blend an entity's limited record with a broader class average, and we always state how much of a score rests on real data versus assumption. A score that pretends to more certainty than it has is worse than no score at all.

Early, Serious

No product theater.

We are early, and we say so plainly. What we sell today is not a platform but the work itself: a fixed-price Risk Assessment that puts a defensible number on an exposure nobody else will price, with the full reasoning attached. If you build, price risk, or just want to watch this take shape, there is a way in for you too.

Products

Put a number on it.

Tell us the risk you cannot currently price, and we will tell you honestly whether we can help. Or try the tools below.

Claude skill · Works with Claude

Quant Risk Dashboard

Give Claude any spreadsheet and a question. It cleans the file, picks and tests the right model, and returns an interactive risk dashboard: the 1-in-200 loss, capital, layer prices and forecasts, a what-if panel to stress them yourself, every hypothesis tested, and a PDF report and Excel export in one click.

Works onAny CSV or Excel file
Verified160 mathematical checks
You getDashboard · PDF · Excel
$20one-time
Buy now → See a sample output →
Price outside India ·
Sample product · Access controlled

Parametric Pricing Engine

A working demonstration of the engine behind a Risk Assessment. Describe a business and it returns the scored exposure, the indicative cover, and the reasoning behind the number — the same output an underwriter would receive.

AccessApproved business users
IncludedOne assessment
Applies toAI liability exposures
Turnaround
2 to 3 weeks
Lead focus
AI liability
Also covering
Space tech
Scope
Set on a call

How it runs

01

You describe the risk

A call, or a written brief. We tell you honestly whether we can price it, and say no when we cannot.

02

We build the number

Frequency and severity modelled separately, thin history blended against a class average with credibility methods.

03

You get the working

The score, the reasoning, the sources, and a plain statement of what rests on data and what rests on assumption.

04

You interrogate it

One round of questions. If the logic does not hold up to you, it is not finished.

A finished example

This is the shape of the work. The field note below is the public version of how we approach a market with no pricing data underneath it: mandated cover, willing buyers, and no one able to name a fair number.

FIELD NOTE 001 · SPACE · JUL 2026

The satellite insurance gap: what one launch failure exposed →

Sixteen private payloads, none of them covered, in a market where insurance is a licensing precondition. Read the working.

SAMPLE DELIVERABLE · FULL REPORT

Aerthos Risk Assessment →

A complete assessment in the format you would receive. Opens in a new tab.

What we are not

We are not a licensed insurer, broker, or underwriter, and we do not carry risk. We produce the number and the reasoning behind it. Whoever carries the risk decides what to do with it.

About

A company that starts from what goes wrong.

Most people build from optimism. We build from the quiet question underneath every serious decision: what if this fails, and who pays when it does.

That question turns out to be badly served. The world has gotten very good at pricing familiar risks and almost helplessly bad at pricing new ones. A rocket. An AI system acting on its own. A flood that was supposed to be impossible twice in a decade. The people exposed to these things often cannot get a fair price, or any price at all, because no one has done the work of understanding the risk well enough to put a number on it.

Moluscca exists to do that work. We are early. There is no grand product to show you yet, and we would rather say that plainly than dress it up. What we have is a clear view of a real problem and a disciplined way of chipping at it.

Why the name

A mollusk cannot remove the grit that gets inside it. So it does something stranger. It wraps the irritant, layer by layer, until the thing that threatened it becomes a pearl. That is the closest thing we have to a philosophy. Risk does not go away. But worked at carefully, it can become something valuable: a price, a decision, a bit of certainty where there was none.

The gap is where the pearl is.

Mission

Put a fair, honest number on risks that don't have one yet.

Our mission is narrow on purpose. We are not trying to predict everything or protect everyone.

We are trying to take specific risks that are currently unpriced or mispriced, and build the understanding that lets someone price them fairly.

Fairly matters in both directions. A price too high pushes people out of cover they need. A price too low quietly bankrupts the person carrying it. Getting the number right is not a technical nicety. It decides whether a small satellite company can afford to launch, whether a business can get covered, whether an insurer stays in a market or leaves it.

We measure ourselves by one thing: did we help someone make a better decision about a risk they could not see clearly before.

Vision

A world where being new is not the same as being uninsurable.

Right now, if a risk is unfamiliar, the system's default answer is no. No price, no cover, no way in.

Whole categories of useful, ambitious activity get held back not because they are too dangerous, but because nobody has done the patient work of measuring them.

We want to change that default. We picture a future where a new kind of business, a new technology, a new kind of exposure can be understood and priced within reason, instead of being refused because it has no history.

We will not get there in one leap, and we are suspicious of anyone who claims they will. We start with one risk we can genuinely understand, do it honestly, and earn the right to take on the next. The ambition is large. The method is patient. Both things are true at once.

Research

How we score risk.

The method is old where it can be and new only where it has to be. Most of what makes a risk pricable is careful, unglamorous work, and we would rather do that well than reach for something clever we cannot defend.

A

Frequency and severity, separately

How often a loss happens is one question. How large it is when it happens is another. We model them apart and then combine them, because averaging them together hides the very thing that matters: the rare loss large enough to ruin someone.

B

Honest about thin data

New risks have short histories. We blend an entity's own limited record with a broader class average using credibility methods, and we always state how much of a score rests on real data versus assumption.

C

Explainable by design

If an underwriter or a regulator cannot follow the logic, the score is not finished. Every number we produce carries its drivers in plain language.

Research areas

AI Liability
Pricing loss caused by autonomous agents acting on a company's behalf
Pricing
Credibility theory and thin-data pricing
Core
Frequency and severity modeling
Tails
Extreme value and tail risk
Updating
Bayesian updating under uncertainty
Detection
Fraud and anomaly detection
Contagion
Graph and network contagion
Simulation
Scenario and Monte Carlo simulation
Trust
Model explainability
Go deeper

Every distribution, process, and model we reach for is documented in plain language, with the formula, the shape, and a real use case, in For the Nerds →

Technology

What exists now, and what is planned.

We label these honestly, because the difference matters and most companies blur it.

Now

A structured data layer that turns public and partner data into something queryable. A scoring engine built on the method in Research. Reports and dashboards that put a defensible number in front of a decision-maker.

Planned

A subscription risk-intelligence dashboard for brokers and insurers. A published, citable annual read on the risks we cover. An API for partners who want the score inside their own workflow.

Ingest → structure → score → explain → deliver · the loop learns from what happens next
What we do not claim

We are not a licensed carrier, a broker, or an underwriter. We produce risk intelligence. The people we serve carry the risk.

Manifesto

Risk is mispriced, and it is getting worse.

The world produces new kinds of risk faster than the tools for pricing them.

Every year, more of what matters, a new technology, a new business, a new kind of weather, falls outside the reach of the methods built to price the old world. The result is a widening gap between what is exposed and what is covered.

This is not an abstract problem. It is a satellite company that cannot launch because no one will price its risk at a number it can afford. It is a town that was insurable last decade and is not anymore. It is a small business turned away, an operator flying uncovered, a family discovering after the fact that the thing they thought was protected never was.

The usual response to a hard-to-price risk is to refuse it. That is the safe move for the person doing the pricing, and it quietly pushes the cost onto everyone else. We think that is a failure of effort more than a law of nature. Most unpriceable risks are not truly unpriceable. They are unstudied.

We do not believe any single model will ever predict everything. The rare, world-changing event is rare precisely because it sits outside what we have seen before, and we are honest about that limit. But between the fully predictable and the truly unforeseeable sits an enormous middle ground of risk that is simply neglected, priced with data that no longer fits, or not priced at all. That middle ground is where we work.

Our approach is deliberately modest in its steps and immodest in its aim. We take one risk at a time. We build the data no one bothered to gather, apply the method the situation actually calls for, and stay honest about how much we know and how much we are guessing. We would rather deliver one number we can defend than ten we cannot.

We are at the beginning. We would rather be plain about that than perform a success we have not earned. But the direction is clear, and we think it is worth committing to for a long time: to make being new less of a reason to be uninsured, and to treat the careful pricing of uncertainty as work worth doing on its own terms.

The gap is where the pearl is.

Field Notes

What the data already knew.

Short analyses of risks the market has not caught up with, led by AI liability and followed by space tech. We publish only what is real, and we show our working. If a number appears here, we can defend where it came from.

Note 002 · AI Liability · Sep 2026

The policies were written before the software could act on its own

An autonomous agent with system access can create a liability in seconds. The cover a company already holds was drafted for a world where a human made that call. Click to read why that gap matters.

→

Most commercial policies allocate liability by asking who acted and whether they acted reasonably. Professional indemnity asks whether a professional fell below a standard of care. Cyber asks whether an outsider got in. An AI agent that has been given write access to a company's systems and takes a wrong action fits neither question cleanly. Nobody broke in, and no professional was negligent in the usual sense. The company is still liable for what its systems did.

The regulatory direction is not ambiguous. The EU's AI Act places graded obligations on deploying organisations, not only on the companies that build the models, and those obligations phase in over the period most enterprises are actually rolling agents out. India's data protection regime puts the accountability for automated processing on the entity that decides how it is used. In both cases the buck stops with the deployer, which is precisely who currently cannot buy a policy priced for the exposure.

The pricing problem is narrower than it looks. It is not that AI risk is unknowable. It is that the three inputs an underwriter needs are all thin at once: how often an agent causes a loss, how large that loss runs, and how strongly incidents inside one organisation correlate with each other. Incident counts are sparse and inconsistently disclosed. Settled claims that would anchor severity barely exist yet. And correlation is the one that should worry a portfolio, because a fleet of agents running the same model on the same prompt scaffolding does not fail independently.

What we are doing about it: pricing this from the deployment's actual shape rather than waiting for a decade of claims. What the agents can reach, what a wrong action costs at each of those reach points, what human review still sits in the path, and how much of the fleet shares a single point of failure. Frequency and severity modelled separately, thin history blended against a class average, and an explicit statement of how much of the number is observed versus reasoned.

The honest caveat: a meaningful part of any AI liability score today rests on structural reasoning rather than fitted loss history, because the history does not exist yet. We would rather put that on the face of the assessment than quietly price it in.

Note 001 · Space · Jul 2026 · second vertical, expanding next

The satellite insurance gap: what one launch failure exposed

India's private space sector is flying largely uninsured. When PSLV-C62 failed in January 2026 carrying sixteen satellites, none of the private payloads on board were covered. Click to read why.

→

Private Indian space companies have said insurance was never a priority while launches were limited to government missions. Now the market has moved from sovereign missions to private enterprise without a risk framework underneath it, and premiums quoted to small operators can run as high as half a satellite's value, because insurers have no pricing data to work from.

The regulatory forcing function is already here: authorised private operators must demonstrate adequate insurance as a precondition of their operating permit. No proof of cover, no license to fly. Which means the missing piece is not demand. It is the data layer that would let an underwriter price a mission with confidence, the same way weather data lets them price a hurricane.

The gap, stated plainly: mandated cover, willing buyers, and no one able to name a fair number. That is exactly the shape of problem Moluscca exists for.

More notes in progress

We publish when the working is defensible, not on a content calendar.

Writing

Longer thinking.

Longer pieces and full assessments. Each one names the method underneath it, so you can see which part of the engine it exercises.

Roadmap

Where this goes.

Intent, not promises. Dated loosely on purpose, because anyone who gives you exact dates this early is guessing and hiding it.

Now

Prove the method

Run the scoring method on the first risk, with real partners, on real cases. Publish the working.

Next

Make it repeatable

Turn the method into a product that doesn't need us in the room, and publish our first public risk read.

Later

Second risk, same engine

Add the next vertical on the same scoring engine, and open the data layer to partners through an API.

Community

Join early.

No product to sell you, no mailing list to farm. Just a way to keep the right people close while this takes shape. Tell us how you fit and we will keep you in the right loop.

You're in. We'll keep you in the right loop.
Join Beta

Work with us early.

For teams sitting on a risk they currently cannot price. Book a call below, or write to us first. If you already know which product you want, the scope is on Products.

Book a call directly

Available Sat · Sun · Mon
Calendar is taking a moment to load. If it does not appear, email moluscca.support@gmail.com and we will find a time.
or write to us first
Received. If we can genuinely help, you'll hear from us.
Contact

Reach us.

Simple and direct. No form theater.

For the Nerds

The working library.

Every distribution, process, and model we reach for, in plain language, with the formula, the shape, and where it's used. Built for people who want to check our math, and for anyone learning this the way we did.

The six methods we actually use

Every score we produce is built from these six. Each one below says what it does in plain language, and how much real data currently sits underneath it. Where the honest answer is "not much yet", it says that.

Frequency modelling

Partly backed
How often a bad event happens.

We model the number of incidents in a period separately from how bad each one is. For a fleet of AI agents that means asking, plainly, how many times a year one of them does something that creates a liability.

Evidence
Backed by real data for mature classes like motor and property. For AI liability the incident counts are still sparse and inconsistently reported, so we lean on disclosed incidents, regulatory filings and operator logs, and we say so on the face of the score.

Severity modelling

Partly backed
How large the loss is when it happens.

The size of a loss follows a different shape from its frequency, and averaging the two together hides the case that actually ruins someone. We fit severity on its own so the large-loss end stays visible.

Evidence
Severity distributions are well understood in general. What is thin for AI liability is the settled-claim data that would anchor the scale, so a meaningful part of the severity picture currently rests on analogy to professional indemnity and cyber.

Credibility theory

Well backed
Blending a thin history with the class average.

When one company has only a short record, you do not trust it on its own and you do not ignore it either. Credibility weights that company's own experience against the average for similar companies, in proportion to how much real data it actually has.

Evidence
This is the most defensible part of the method and the reason thin-data pricing is possible at all. The mathematics is century-old actuarial practice; what varies is the quality of the class average we blend against.

Bayesian updating

Well backed
Revising the price as evidence arrives.

We start from a stated prior, then move it as real outcomes come in. Every revision is traceable: you can see what the number was, what evidence moved it, and by how much.

Evidence
The machinery is sound and the priors are written down rather than hidden. The honest caveat is that a prior is a judgement, so we publish it and let you argue with it instead of burying it in a model.

Extreme value theory

Thin data - stated
Modelling the worst outcomes separately.

The ordinary range and the catastrophic tail are not the same distribution, and fitting one curve to both understates the tail badly. We model exceedances over a high threshold on their own terms.

Evidence
Well established in flood and catastrophe work with decades of observations behind it. For AI liability there are almost no observed tail events yet, so tail estimates here are structural reasoning rather than fitted history, and we label them that way.

Monte Carlo simulation

Well backed
Running the model many times to show the spread.

Rather than reporting one number and implying certainty, we run the whole model thousands of times over plausible inputs and show the distribution of outcomes that comes out.

Evidence
The technique is exact; its output is only ever as good as the inputs it samples from. Where an input is an assumption rather than an observation, the simulation inherits that, which is why every assessment states the split.
Why we label the evidence

A method being mathematically sound and a method being well fed with data are two different things. Most of the risks we work on are new enough that some of the inputs are reasoned rather than observed. Saying which is which is the difference between a number you can defend and a number that just looks confident.

404

This page doesn't exist. Statistically speaking, most things don't.

Back to safety →