Three founders.One registered office, in Providence.
A dated record of what your assistant was configured to say.
Everett Winslow Founder
Ran conversation design for a payer's member assistant and spent three weeks rebuilding a retired dialog task from Jira tickets and Slack scrollback after a complaint, because the version itself was already gone.
Corinne Hargrove CISO
Sat on the other side of that review as the payer's third-party security lead, and kept rejecting vendors who asked for transcripts and PHI when the question only ever needed the configuration.
Yurii Dziamko CPO
Built the release diff for a bank's IVR team that matched nodes across versions on string rules and lost every renamed one, then watched the version counter wrap during a QBR with nobody able to date the deletions.
NOTES
Muniment keeps a dated, hash-chained copy of every version a US health payer or bank publishes of its member-facing assistant. Kore.ai retains twenty and deletes the oldest on each publish, five months at a weekly release, while complaints arrive nine to eighteen months later. A read-only credential pulls each definition through the export API into the customer's own storage bucket under Object Lock.
Renamed nodes are where Muniment's string matching stops and its models start. Exact ids and strings link about seven node versions in ten. For the other three, a cross-encoder scores embedded node pairs from version N and N+1 to carry one lineage id through renames. Each parsed version adds to the platform grammar Muniment keeps across accounts: which export field is a node label, prompt string or entity, never a customer's text.
Muniment's resolver takes a timestamp and an answer pasted from a transcript, and names the node version, live at that moment, that produced it. Below 0.90 confidence it returns unresolved, never a best guess, and the query waits for a person the customer names, as do low-margin lineage joins. Each reviewer decision is stored as a pair, labeled and owned by that customer, holding no pasted answer and pooled with no other account.
Onboarding, publish storms and legal matters are the three bursts in Muniment's load. A new customer's backfill embeds every string in every surviving version at once; a publish storm pushes dozens of assistants through parsing, embedding and lineage inside an hour; a matter sends hundreds of resolver queries in a day, then nothing for weeks. Pasted answers come from member conversations, so any model reading one stays on GPUs Muniment runs itself, never a hosted API.
Muniment's product runs on cloud infrastructure in Ohio. Capture and evidence packs run there, the index, embeddings and hash chain sit in object storage, and the resolver and cross-encoder run on cloud GPU instances Muniment operates, a cost that rises with every connected assistant and archived version. The customer's cloud bill carries the locked archive.
Azure East US 2 handles the two Muniment jobs that never see a pasted answer: bulk embedding and model training. Azure OpenAI embeds configuration text during backfill, and the third-party endpoint it replaces is retired by March 2027. Azure Machine Learning fine-tunes lineage and resolution per account on reviewer pairs, and by September 2027 each customer has its own cross-encoder, served from the GPU instances in Ohio.
Muniment Records, Inc. filed its incorporation in Rhode Island in January 2024. It charges $1,450 per connected assistant a month, unmetered. It holds no SOC 2 Type 2 or ISO 27001 attestation yet; security reviewers get the subprocessor register and a US-only data map.