Muniment archives every version your AI assistant publishes
Your bot platform keeps twenty versions and deletes the rest, so the date legal names is the one that is gone. We copy every version out the day it ships, into a bucket you own, dated and hash-chained. Name a date and paste an answer from a transcript, and our models find the node version that produced it or say unresolved.
No card. Revision history supplied on request.
20
user-created versions retained before the oldest is auto-deleted
$1,450
per connected assistant, per month
3 in 10
node versions renaming breaks, matched by an embedding model and never auto-confirmed
Who it fits
Go count your surviving versions before you read the rest.
Open the version list for your busiest assistant. Find the oldest one still there. The date on it is the date your evidence starts, and for most estates it is far more recent than anyone in the room expects.
This is for you if
- ✓You publish weekly or faster, and the oldest surviving version is newer than the oldest matter still open on your desk.
- ✓Somebody has already asked what the assistant was configured to say on a date, and the answer took two weeks to assemble out of Jira tickets and Slack scrollback.
- ✓Your counsel will accept a stated absence. Your counsel will not accept a reconstruction worded as what the flow most likely said.
This is not for you if
- ✕Your assistant has published fewer than twenty versions since launch. Your history is whole. If your platform documents a definition export, take it yourself and spend the money on something you need.
- ✕Your platform has no documented definition export. A read-only credential has nothing to pull, and we will say so on the first call.
- ✕You are shopping for QA, linting or release tooling. We record what you published and we never score it.
- ✕Your compliance office requires a Business Associate Agreement from every vendor regardless of scope. We hold no PHI, so we will not sign one.
Figures
Twenty versions, one publish a week, one complaint that lands next year.
Twenty releases is five months at one publish a week and ten weeks at two. Count what your busiest assistant still holds, then find your oldest open matter. Those two dates are the entire argument.
- 0120user-created versions Kore.ai retains, the oldest deleted the moment the twenty-first is created
- 025 monthsroughly what twenty versions buys an assistant publishing once a week
- 030.90confidence below which the resolver returns the word unresolved instead of a match
- 04$1,450per connected assistant per month, with captures, queries and packs uncounted
Installation
What the rules settle on their own, and what stops for a person.
Five steps. The last two are branches: an embedding model proposes the match, a person confirms it, and one branch refuses to answer.
Steps 01–03
- 01Connectyou issue a read-only export credential and your cloud team turns on Object Lock in compliance mode on a bucket you own
- 02Captureeach publish event pulls the whole definition through the platform's export API into Parquet and JSON, with a SHA-256 and a manifest hashing the one before it
- 03Askyou name a timestamp, paste the answer string out of the transcript, and give it a matter reference
Steps 04–05
- 04Above 0.90the pack comes back on its own: the node definitions live at that instant, the raw export, its hash, and the manifest chain that dates it
- 05Below 0.90the query returns unresolved and waits for a named reviewer on your side, whose confirmation is written back as an audited row
what you get
What comes back when you name a date and paste an answer
One evidence pack, dated and hash-chained, out of a bucket you own.
- asked about
- 14 March 2026, 09:41 UTC
- node version
- billing_dispute_v4, published 2 March 2026
- resolver score
- 0.94, above the 0.90 floor
- export
- parquet and json, 2.3 MB, pulled at publish
- object lock
- compliance mode, 7 years, your bucket
- sha-256
- 9f2c41ab7e05d38c6b1f04a927e5cb83d1c7fa6029b45e18
- chain
- manifest links the publish before it, unbroken to first capture
A drawn example, not a customer record. The hash and the chain are what a reviewer checks; the rest is what they read.
Scope
What we refuse, and what lands in your bucket.
The first of these is a job your platform does better than we ever will, and we are not going to pretend otherwise.
What's in scope
- 01Publish-triggered capture of every assistant you connect, backfilled at connection with whatever versions the platform still holds.
- 02Parquet and JSON against a published versioned schema, valid_from and valid_to on every row, written to a bucket in your own AWS account.
- 03Evidence packs against a matter reference at any volume, the review queue, and a verification script anyone can run against your bucket.
What we don't do
- 01We do not restore a version into your platform. Kore.ai's own documentation describes a Restore option that replaces the in-development version with an earlier definition, and a platform-native restore beats an outside archive at that job.
- 02We hold no transcripts, no recordings and no member records. We can tell you what the assistant was configured to say on a date, never what it said to a named person.
- 03We do not test, lint or score a flow. No sandbox, no regression suite, no warning that a prompt reads badly.
Questions
Six answers you would rather have now than after the security review.
Does this replace Kore.ai?
No. It sits beside it on a read-only credential, we hold no write scope, and we will refuse a credential that carries one.
Can you recover a version the platform already deleted?
No. Whatever the cap removed before you connect is gone, and the first capture report tells you exactly where your archive can begin.
Will you sign a Business Associate Agreement?
No. We take no feed that produces PHI and we hold none, so you get that answer before you connect rather than after.
What happens when the resolver is unsure?
Anything under 0.90 comes back as unresolved, in that word. We never return a best guess, and we do not warrant that a match is the node that in fact produced an answer.
What does Kore.ai's Trust Center list?
SOC 2 Type 2, PCI DSS, ISO/IEC 27001:2022, GDPR, CCPA, EU AI Act and DESC badges. It does not list a HITRUST or HIPAA badge. That is a question for them, not a claim from us.
And what does Muniment hold?
None of these badges yet. A security review gets the subprocessor register, the read-only scope, the schema, and a plain answer on location: what you paste is read only on our own GPUs at AWS in Ohio, and models are trained on Azure in Virginia.
Rates
$1,450per connected assistant, per month.
One flat rate per assistant. No seat count, no session meter, no volume table.
- Publish-triggered capture of every assistant you connect, and backfill of whatever versions still exist on day one
- Resolver queries and evidence packs, at any volume a matter needs
- Reviewers, auditor accounts and read-only users, none of them counted
- The published schema and the verification script, under an open license
- The first capture and the coverage report are free and run before any contract.
- Month to month for the first six months. You can cancel on 30 days' notice and no annual commitment is asked for.
- If you leave, the Parquet, the raw exports and the manifests are already in your bucket, and the schema stays published.
Run the count first. Then send us the number.
A work email and a company name. Tell us whether you want a walkthrough or a first capture, and we will do that one and not the other.
We'll come back within one business day with a straight yes or no.
- ✓ The first capture and the coverage report are free and run before any contract
- ✓ What comes back is the earliest surviving version of each assistant and the date your archive can start
- ✓ If that date is older than every matter you have open, say so and we will leave you alone