About Muniment

Muniment is built for whoever owns a member-facing assistant at a US health payer or bank. A member complains, and compliance asks what the assistant was configured to say that day. The platform usually cannot answer. Kore.ai's own versioning docs cap retained versions at 20, and each new publish deletes the oldest. At a publish a week that is five months of history. Complaints arrive nine to eighteen months later. So a team spends two or three weeks rebuilding a deleted flow from Jira and Slack, and counsel gets a hedge.

We are a data product, not another platform. A read-only credential lets us watch the publish feed. On every publish we pull the whole assistant definition through the platform's own export API and file it as a dated record, with the raw export and its hash. Each manifest hashes the one before it, so a missing capture or a later edit shows. The files land in an S3 bucket the customer owns, under Object Lock, in the region their cloud team picked. Compliance gets an evidence pack: name a date, get the flow that was live, the export it came from, and the chain that dates it.

The hard part is not storage, it is identity. Tasks get renamed, nodes move, flows get copied and edited. Exact id and string matching links about seven node versions in ten across a year of releases. The other three are the ones a lawyer asks about: the flows that got rewritten are the ones that got something wrong. So we embed every prompt string, node label and answer, and a cross-encoder scores pairs of the form (node in version N, node in version N+1) to carry one lineage id through the renames. Paste an answer from a transcript and the models find the node version that produced it. The model is not allowed to conclude anything. Below 0.90 confidence the resolver returns unresolved, and the pack says so in that word. A stated absence holds up in a matter; a confident guess is the exposure we exist to remove. Unresolved cases go to a queue where a named person on the customer's side decides.

Those decisions build a labeled corpus of reviewer decisions matched to node versions, owned by the customer who produced it and pooled with nobody. One asset does carry across accounts and it is ours: the platform grammar, which export field is a node label, which a prompt string, which an entity, and how an export changes shape between platform releases. It holds nothing belonging to any customer, and grows with every version we parse. The compute is bursty in three shapes. Onboarding embeds every prompt string in every surviving version at once. Publish storms push dozens of assistants through parse, embed and lineage inside an hour. Matter bursts throw hundreds of resolver queries into one day, then nothing for weeks. The strings customers paste come out of member conversations, so no model that touches one may run on a hosted model API or leave the United States. Those models, the resolver and the cross-encoder, run on EC2 GPU instances we operate on AWS in us-east-2 (Ohio), beside the index, the embeddings and the hash chain in S3, and that bill grows with every connected assistant and archived version. Microsoft Azure in East US 2 does the two jobs that never see a pasted string. Bulk embedding of configuration text during backfill runs on Azure OpenAI under zero data retention, and the third-party endpoint it replaces is gone by March 2027. Azure Machine Learning fine-tunes lineage and resolution per account on reviewer pairs, never pooled, and by September 2027 each customer gets its own cross-encoder trained there and served back on EC2 in us-east-2.

We are early, and the price says so. One rate: $1,450 per connected assistant per month. Captures, versions, storage, resolver queries, evidence packs, reviewers and readers are all uncounted, so an estate publishing three times a week pays what one publishing twice a year pays. Light users subsidize heavy ones. First capture and coverage report are free, before any contract. The backfill is free, however large. The first six months are month to month. We hold no SOC 2 Type 2 or ISO 27001 attestation yet. What goes in front of a security review today is a plain map of where each thing runs, all of it in the United States, the read-only scope, the subprocessor register, and a schema and verification script anyone can check an archive with.