The question that stops the room
Picture a credit committee review at a public-sector bank. A retail loan was declined in March. The customer complained, the ombudsman asked, and now someone has to explain the decision. The team knows an AI model was in the loop. Which one? The one that was live in March — or the one that replaced it in April? Was it the fine-tuned version, or the base? Who signed off on it, and against what test results?
We have watched capable teams spend a week reconstructing that answer from deployment logs, chat threads and one engineer's memory. Everyone meant well. Nobody could prove anything. That is not a competence problem; it is a record-keeping problem, and it is exactly the problem a registry solves.
The same scene plays out in an insurer's claims unit, a defence programme's evaluation board, a space agency's mission-readiness review and a contact centre's quality audit. Different rooms, one question.
What every model record carries
A model in Model Hub is treated as a decision with a paper trail, not a file. Its record falls into six categories.
Identity
Name, version, checksum, named owner, lifecycle state. If two people say “the model”, the record settles which one.
Provenance
The model's licence and model card, and recorded lineage links to the models and datasets it came from.
Evaluation
Benchmark results, fairness reports across protected groups and drift alerts — the evidence behind “it works”.
Policy
Organisation policies on data sensitivity, licences and export control apply to every model; each model's record carries its data sensitivity, jurisdiction and residency region, so the rules are written down, not remembered by the team.
Lifecycle
Model and module versions move through draft, published, deprecated and retired, with who changed the state and when; a module version is published only after it is approved.
Audit trail
Recorded events, exported in a compliance pack with the recent audit events and a PDF report. The record is evidence, not a screenshot.
Why a registry, and not a spreadsheet
Most organisations already have a model inventory. It is a spreadsheet, it was accurate the week it was made, and it is maintained by the one person who cares. It cannot tell you that the model in production is the one in the inventory, because nothing checks the checksum. It cannot tell you the evaluation was run on this version. It cannot stop a deployment.
A registry is different in one decisive way: it sits in the path. The runtime asks Model Hub which model is approved for this agent under this policy, and runs that one. An unrecorded model is not merely undocumented — it is unrunnable. The inventory stops being a report about the system and becomes part of the system.
Where it earns its keep
The same registry, five rooms. Each of these is a real conversation we have had; none of the outcomes below is a metric we are claiming — they are the questions a record lets you answer.
PSU and private banks
Model risk management wants a model inventory with owners, validation evidence and change history. Model Hub is that inventory, kept current by the runtime itself — designed to support the RBI FREE-AI framework’s expectations on model governance.
Insurers and NBFCs
When an underwriting or claims model is replaced, the record shows what changed, what was re-evaluated, and who approved the swap — before the auditor asks, not after.
Defence programmes
Every model’s lineage is on record: what it was trained on, who trained it, and that its composition can be shown rather than asserted. Air-gapped deployment keeps the registry inside the wire.
Space and mission operations
Mission-readiness reviews get a per-version evaluation history for the models advising operators — and a recorded trail of which version was live on which day.
Contact centres and BPOs
Swap a model behind an agent without touching the agent, and keep a receipt of the swap. When a client asks what handled their customers in a given quarter, the answer is a query.
Model routing, explained
When work is routed between models, every routing decision cites the graded runs behind it. The registry is where those runs live.
What we are careful not to promise
An honest boundary is part of the product. Here is ours.
Model Hub does
Record every model in the six categories above, hold a module version as draft until it is approved, apply organisation policies to every model, and keep an audit trail.
Run inside your perimeter — your cloud, your jurisdiction — as part of the AgentAnywhere platform or alongside the models you already serve.
Model Hub does not
Make a model accurate, fair or safe. It records the evidence you gathered; it does not manufacture it. A registry with an empty evaluation field is an honest registry telling you the work is not done.
Certify your deployment. ShepHertz operates a control environment credentialed for SOC 2 and ISO 27001 and independently assessed for HIPAA and GDPR; our alignment to the DPDP Act and RBI FREE-AI is designed-to-support, and informs your own assessment rather than replacing it.
Where Model Hub is today
The commercial Model Hub — evaluation history, fairness reports, drift alerts and the audit trail — is part of the AgentAnywhere platform.
An open edition is one of four projects in our open-source programme; it will cover model identity, versioning, checksums and lineage. It is the earliest of the four: the licence is not yet fixed and the repository is not public. We would rather tell you that plainly than let a roadmap slide read as a release.
Frequently asked questions
What is an AI model registry, and how is it different from a model inventory?
A model inventory is a document that lists models; a model registry is a system that keeps each model's record. AgentAnywhere Model Hub is a registry: every model carries identity, provenance, evaluation history, the policies that apply to it, lifecycle state and an audit trail, and a module version is published only after it is approved; any change sends it back to draft for approval again.
Which questions can Model Hub answer for a regulator or auditor?
The three that come up in every review: which model produced a given output, what that model's known limitations were at the time, and who approved it for production. Each is answered from the model's record in Model Hub.
Does Model Hub help with RBI FREE-AI or model risk management at a bank?
It is designed to support them. Model risk management frameworks and the RBI FREE-AI framework expect a governed model inventory with named owners, validation evidence and change history. Model Hub keeps that inventory current from the runtime itself. This is an alignment to inform your own assessment, not a certification of your deployment.
Can Model Hub run in an air-gapped or defence environment?
Yes. Model Hub runs inside your perimeter as part of the AgentAnywhere platform — on your cloud and in your jurisdiction, including region-pinned deployments, and is designed to deploy air-gapped. The registry and the audit trail stay inside the wire.
Is there an open-source Model Hub?
Not yet. An open edition is planned and will cover model identity, versioning, checksums and lineage; the licence is not yet fixed and the repository is not public. The commercial Model Hub, with the full evaluation history, organisation policies and audit trail, is part of the AgentAnywhere platform; talk to us to see it.
Written by
AgentAnywhere Research
The team that builds the platform and the models
AgentAnywhere Research writes about the platform, the model families and the trust layer we build and run in India. Where a figure is ours, it says what it covers; where something is a demonstration, it says so.
