Skip to content

Capabilities

Everything it does, and what makes each one true.

83 capabilities across 12 areas, each with the thing in the running product that makes the claim checkable — and real screens of the workspace working, on synthetic data.

Bring us a fileThe eight-property argument

Derived, not typed

The completeness of this list is checked against a census generated from the product’s own source and compared byte for byte in its build.

Evidence on every item

Each capability names the array, table, rule set or screen that makes it true — the thing you would open to check it, not a restatement.

Status on every item

78 are in the product. 4 are built and not yet connected and say so. 1 is roadmap.

Real screens, synthetic data

Every image is the running workspace, captured from its fixture build. No live client data, no mock-ups, and the marking is in the pixels.

How to read this page

In the product

Reachable by a person with a seat today.

Built, not yet connected

The implementation exists and is tested; no adapter is bound to it in a running environment, so it does not run. Said out loud rather than folded into the list above it.

Roadmap

Decided, not built.

The second of those three is why this page is worth reading. Four capabilities below are built, tested and not connected to anything that runs — three of the four ways a case posture can be confirmed, and semantic recall. Folding them in with the rest would make four channels look like four running channels. Leaving them out would make real engineering invisible to a buyer entitled to ask about it.

A desk of specialists, not a model with a prompt

No single discipline can value a contested claim. The medicine says whether the treatment was defensible; the law says whether the security is enforceable and who is prosecuting; the counterparty side says how they behave when it is time to pay; capital says what that is worth today. Four readings, composed — and allowed to disagree.

Four named specialist readings of one position — medicine, the case, the carrier and the capital — followed by a composed verdict recommending against funding at model value.
Four Digital Employees read the same file independently, each answering only its own question. The composition names the disagreement instead of averaging it away, and here it declines to fund at model value. Synthetic data.
One specialist expanded, showing its finding and the retrievals behind it as labelled chips.
Open one specialist and read its working: the finding, the retrievals it rests on, and the axes it is permitted to opine on. A conclusion outside those axes is quarantined before it reaches the verdict. Synthetic data.

No number without the retrieval that produced it

There is no training and no fine-tuning anywhere in this product. Valuation is retrieval-grounded: the reasoning works over comparable settled matters pulled from a corpus, and the band is descriptive statistics over that cohort. The corpus is the model, so the retrieval is pinned as carefully as the engine.

The same reasoning run replayed from a recorded trace, each specialist showing the retrievals it made: pages read, docket verified, carrier profile, simulation count and cohort depth.
A mark struck weeks ago, replayed from the retained reasoning artefact rather than from a log of it — the cohort, the comparables, the query and the pinned versions — so the same inputs produce the same answer in front of a committee years later. Synthetic data.

It refuses to price what the evidence will not carry

A product that always produces a number is a product whose numbers mean nothing. Refusing is a capability here, and it is enforced in the schema rather than left to anyone remembering.

A position showing No price in place of a discounted value, with both the expected settlement and the advance rendered as dashes.
The product refusing. 180 comparables and still no midpoint the evidence will carry: the mark is recorded as indicative, recommends no advance, and is excluded from book value rather than assumed to zero. Synthetic data.
A position marked THIN COHORT, with a recovery band drawn across six comparable settlements and a note stating the spread as a proportion of the midpoint.
Six comparables and two hundred must not look the same on screen. The band is shown, the spread is stated as a proportion of the midpoint, and the screen says the width is shallowness rather than volatility — a different fact with a different remedy. Synthetic data.
A position marked WIDE DISPERSION on 214 comparable settlements, with a note that the comparables disagree.
Depth does not rescue dispersion. Two hundred and fourteen comparables, and the band is still too wide to present as a firm number — which is why cohort size alone is never the confidence. Synthetic data.
An investor-relations view stating that the marked value may not be stated as a book value, with the caveats that travel with the figure and the bracket it sits inside.
Whether the book may be presented as a book value is decided by the server and reported to the screen, not judged on it. The caveats travel with the figure wherever it is quoted, and the bracket is published beside the midpoint. Synthetic data.

What they told you, beside what their own data supports

A counterparty’s stated collection rate is the seller’s estimate of the thing being priced. It is worth having. It is not a measurement, and storing it as one is how a book gets marked on somebody else’s arithmetic.

A position header reading NO FILING OF ANY KIND ON RECORD, with a composed recommendation describing it as a high-confidence mark on a deep cohort.
Nothing on file is a finding, not a blank. The position states what is missing rather than defaulting the mark — and a deep, predictable cohort can still be a confident one. Synthetic data.

The controls are in the ledger, not in the process

A gate that reads a field the same statement can set is not a gate. Everything here is enforced where the money moves, so a control cannot be satisfied by remembering to satisfy it.

Concentration lines by counterparty with exposures, shares and no-limit-set markers, above a table of pools with reserve targets.
Concentration measured continuously across five dimensions — attorney, carrier, provider, jurisdiction and line — with the headroom on each. Where no limit is stored the line says so, because an unlimited dimension is an observation and cannot render as a breach. Synthetic data.
A position header reading NO LIEN NOTICE ON RECORD followed by a verified UCC-1 filing date, above a composed recommendation to fund.
The security interest on the face of the position: what was filed, when it was verified, and whether that supports advancing against it. Funding checks the verified filing itself, not a column any statement could set. Synthetic data.

Reconstructable, not merely retrievable

Retrievable means we can show you what we saved. Reconstructable means you can re-derive it. A committee, an auditor or a trustee asking why a mark was struck three years ago needs the second one.

A dialogue titled Adjust the advance, with an editable amount and four selectable reasons, noting the decision is written to a hash-chained audit log.
The system reasons; the underwriter decides. The reason is captured as a coded record written into a hash-chained audit log and cannot be rewritten afterwards — which is what makes the record answer the question a trustee actually asks. Synthetic data.

The book marks itself

A position whose status is checked when somebody remembers is a position that is wrong between checks. Staleness here is measured from the last time a posture was confirmed, not the last time a row was touched.

A daily queue titled What needs me today, showing items past their confirmation cadence with the exposure riding on each, and a note that the ordering is by stated precedence.
The servicing day, ranked by exposure against drift — and staleness measured from the last time a posture was CONFIRMED, not the last time a row was touched. A case nobody has checked in six weeks is a risk even when nothing about it has changed. Synthetic data.

The arithmetic underneath, done properly

Deterministic, integer-minor-unit money with no clock and no randomness anywhere in it. Every model is versioned, every run is digested, and the same inputs give the same answer in any language.

A portfolio view showing capital deployed, marked value, coverage, a share on thin cohorts, a written-off column rendered as a dash, and concentration lines by counterparty.
The book to the portfolio manager. Note the fourth column: the written-off figure is a dash with the reason it is withheld, not a zero — a withheld figure and a measured zero are different facts and the screen refuses to flatten them. Synthetic data.
A table of two pools with their opening dates, position counts, deployed capital, reserve against target, marked value and share on thin marks.
Pools, and what is held against each. A pool with no reserve ledger renders as withheld with its reason rather than as a zero balance — the distinction a controller is paid to make. Synthetic data.

Deletion, inside a ledger that cannot be edited

An append-only spine and a real erasure obligation are not usually compatible. Here they are, and the mechanism is the reason.

Seven seats, and each one lands somewhere it can work

A workspace that shows everyone the same screen is a workspace built for whoever specified it. Each seat here has its own landing surface, its own navigation and its own composition of every screen it can reach.

An administration view listing each role with the surface it lands on and the surfaces in its navigation rail.
What each seat gets, read from the same table that decides it at runtime rather than from a description of it — so the page cannot drift from the product it describes. Synthetic data.
An underwriting queue of eight positions sorted by materiality beside an open position showing billed charges, discounted value, expected settlement and advance.
The underwriting bench. Every position awaiting a decision, sorted by materiality, each labelled with how many comparable settled matters its mark rests on — before you open it. Synthetic data.
A position with 47 comparables showing billed charges, a discounted value at a stated percentage of face, an expected settlement in months, and an advance.
A mark the evidence carries: a band, an expected settlement date, and an advance struck against the discounted value rather than against face. Synthetic data.

It explains itself, in place, to the seat reading it

A product with this many refusals in it has to say why, where the refusal happens. Help here is filtered on the server by surface and by seat, so an underwriter’s browser never receives the controller’s prose.

A table of positions the model could not price confidently, each with its face value, midpoint, spread, comparable count and the reason it was withheld.
The marks that may not be presented as confident, separated from the ones that may, each with the ground it was withheld on — thin cohort, wide dispersion, or no price at all. Synthetic data.

What it takes to run it

A product sold to a firm has to be administrable by that firm, and operable by us, without either one reaching into the other.

The EliAI MRF sign-in screen, showing the workspace name resolved from the hostname and a list of demonstration accounts marked as not real.
Sign-in resolves your firm from the address before a credential is typed, and every seat lands on the surface its role actually works from. Synthetic data throughout.

Why this list can be trusted to be complete

A hand-written inventory has one guaranteed property: it agrees with itself. Nothing reads the code, so it cannot disagree with the code, so it drifts — and the first thing that can falsify it is a customer.

So the denominator here is not typed by a person. A generator in the product repository extracts every number in a capability census from a named array, union, enum or decorator in the source, and the build regenerates it and compares it byte for byte. Landing a specialist, a document kind, a ledger event or a help article without regenerating the census is a failed build.

Measured on the transcription below: 43 dimensions and 539 declared capabilities in the source — 4 specialists, 10 owned reasoning axes, 17 document kinds, 11 cross-document reconciliation rules, 14 entity corroboration checks, 36 indexed fields on a comparable, 54 ledger events, 27 named authorisation capabilities, 13 separations of duty, 18 governed settings and 57 help articles.

That census is evidence, not a catalogue, and it is not published. It also counts HTTP routes and database models, which are plumbing a buyer neither sees nor wants enumerated. What is published is the 83 items above: the ones a person uses.

Census figures transcribed by hand from EliAI-MRF PR 337, master on 2026-09-23, because the two repositories do not share a build. Where they disagree, the generated census is right and this page is stale.

If any of this is unfamiliar, it is the wrong list to start with

This page assumes you already commit capital against evidence somebody has to read. If you are not sure that is you, the shorter question is whether your desk looks like this one.

Who it’s for — and who it isn’tA worked valuation, step by step

Pricing

Priced against the book, not against the feature list.

Nothing above is an upsell tier. Every capability marked as being in the product is in the product, for every seat entitled to it. What shapes a number is the size of the book, the seats, and how much of the corpus has to be built before the first mark is worth striking.

What shapes a quoteGet a number

  • Every capability listed as in the product is included; none of them is a module.
  • Seats are per person, and the seven roles are not priced differently from one another.
  • Quotas on positions and documents are set per tenant against a declared ceiling.
  • There is no published list price for this edition; the pricing page says what shapes a quote.

Bring us a file and we will show you the reasoning on it.

Not a demonstration tenant with our data in it — one of your own matters, with the specialists reading it and the band they arrive at, including the part where it declines to price something.

Bring us a file

EliAI™ MRF is generally available. Screens on this page are the real workspace on synthetic data.