Producing evidence rather than reading it? → EMIT
For the parties that assume or price the risk: insurers, reinsurers, MGAs and coverholders, lenders, acquirers and sponsors, and the boards and aggregators accountable for a delegated portfolio they cannot fully see.
Is that evidence sufficient for the risk I am taking?
What you buy now: a reading of an insured's own records against the open covenant schema, delivered to the underwriting desk as a dated findings report at inception and, where agreed, monthly across the term. Arkaya's analysts, supported by AI agents, do the reading. In development: automated collection from an engine at the insured that conforms to the schema. No engine has yet passed the conformance programme, and no date is promised. The report shows what the records support for the underwriting question and where further evidence is needed.
Once a conformant engine runs at the insured, it produces continuous, first-party evidence of its AI estate; Arkaya reads that evidence against the underwriter's question at inception and, where agreed, across the term, and reports what it supports. A reading can sit alongside the questionnaire, on records rather than self-report. The underwriter prices it; Arkaya does not.
The covenant schema is class-agnostic. AI liability is the first class written against it, because the underwriting asymmetry is largest there. Cyber, technology E&O, D&O and operational resilience follow.
The exclusions are based on standard language that Verisk's Insurance Services Office made available in 2025. Insurers can attach them to commercial general liability policies.
Arkaya's hypothesis is that continuous evidence of an insured's AI estate could help insurers assess whether, and on what terms, to offer cover. The underwriting decision remains with the insurer.
A machine-readable schema is only as good as the evidence that feeds it. Evidence produced at human cadence, the periodic questionnaire and the point-in-time attestation, cannot describe an estate that changes between checkpoints. Machine-readable underwriting needs evidence produced at machine cadence: captured continuously, at source, as the estate moves. The decision stays human and is recorded; the observation runs at the speed of the thing observed.
Machine-readable underwriting requires machine-cadence evidence. The decision stays human; the observation does not.
A premium reflects expected loss, uncertainty and the cost of capital held behind the policy. Where a risk cannot be characterised, that uncertainty may affect price, terms or capacity. Better evidence can reduce the information gap; the insurer decides what, if anything, changes in price or terms. At Lloyd's, every syndicate's capital requirement is calculated on an internal model. Our hypothesis: evidence that narrows the distribution could reach the capital held, not only the rate.
The conformant engine produces the evidence; Arkaya reads the record against the underwriter's question and reports what it supports; the underwriter prices it. The calibration layer sets out, as hypotheses, where a governance evidence record might bear on rating, retention, limit and capital. The thresholds and the rate response are the carrier's. The schema and the conformance test are Arkaya's; the calibration, the position and the rating are not.
Each answer sits on one of four rungs of the evidence ladder, set by its provenance and not by Arkaya. The rung caps the evidential weight the desk can give it; the desk decides the weight.
Two insureds, same sector, same limit sought. One produces continuous evidence of its AI estate that the underwriter can check independently; the other answers a periodic questionnaire. The underwriter sees the difference; what follows is the underwriter's decision.
Illustrative. It shows what each underwriter has to read, not an outcome. The carrier sets cover, limit, retention and terms against its own book; nothing here is a quote or a promise of cover.
The read can go either way. Evidence that exposes a weakness may justify tighter terms. The reader's gain is a better-informed decision, not a cheaper one.
The claim is falsifiable and held as a hypothesis: an evidenced counterparty obtains measurably better terms than an otherwise identical unevidenced one. The result is pending. The experiment is a carrier pilot, and the invitation is open to the desks that can run it. If the evidence does not move the terms, the thesis fails on its own stated ground.
An insurance holding company board, an MGA aggregator or a general partner may be accountable for coverholders, portfolio companies or captives it cannot fully see between periodic reviews. Each underlying entity or system produces its own record. The reader reads those records across the portfolio, obligation by obligation. No aggregate state is asserted across obligations: the portfolio view is a set of records, each read on its own terms, not a score for the whole.
Steps one and two, the Protocol and the first record, are the emitter's and sit on the EMIT page.
The deliverable is Arkaya's dated findings report on the insured's records, read against the underwriter's question at inception and, where agreed, re-read monthly across the term. The records stay the insured's.
The record is the insured's. Portability is a schema property: evidence generated against the schema moves with the insured across engines, insurers and capital events without re-basing. No vendor, and not Arkaya, holds it captive.
The pilot is the §05 experiment run on a live desk: the floor at inception, the term monitored, the terms offered compared against the unevidenced book. If the evidence does not move the terms, the thesis fails on its own stated ground.
Detailed scope, implementation timelines and pricing sit in the post-NDA package: working session, sprint map, term sheet.
So the reader does not have to infer the distinction.
The factory mutual system has conditioned premium on engineering-surveyed controls since 1835. Telematics repriced motor in both directions. Cyber desks already buy external ratings, whose opaque scores the no-score discipline refuses. Continuous governance evidence extends a 190-year habit; it does not ask the market to form a new one.
Your governance tools help you manage AI and record what it does. Arkaya reads what those records support for a particular decision, identifies the gaps and states the limits of the evidence.
The discipline that calms a sophisticated reader is the explicit statement of what the architecture does not claim. Each line is enforced structurally, not asserted.
One buyer can hold more than one role: an insurer is a subject to its group board, a supervisor of its own coverholders, and a pricer at its desk.
The Subject sits on the EMIT side, which has its own page; Supervisor and Pricer read.
The reader determines sufficiency, price and terms.
Arkaya does not.
Or request the Reader Brief by email:
Resilience Capital is built.
Not asserted.