# Privacy: what the Rein Agent Risk Scale keeps

This describes the self-check service at this address. It is written from the code that runs it, and
every section says what is NOT kept as plainly as what is.

## What a submission stores

`self_check` (over MCP) and `POST /v1/self-check` (over REST) both write one row to a private table.
The row holds:

- the answers you submitted, as JSON, minus the four fields that get their own columns below;
- the model or harness name you declared, as you typed it;
- the indicative grade, the model-implied credit modifier, the rated grade and the methodology
  version;
- a UTC timestamp to the second, and whether the call arrived over MCP or REST;
- the year and month, which is what rows are filed by month under;
- the eval session id, if you passed one;
- your agent handle **only if you set `publish_opt_in`**. Without the opt-in that column is empty;
- one observed label: the name your software gives itself, with its version, where it came from and
  how strong that evidence is. See the next section;
- a random row key. It is a fresh UUID, not derived from anything about you or your request.

MCP rows created before 8 October 2026, 19:57 UTC also carry `observed_harness_reliability =
"unreliable_v8_mcp"`, because the observed label on those rows may belong to another caller (a fault
fixed at that time). The flag is added; nothing else on those rows is changed.

A behavioural eval session is stored in the same table under its own partition, and holds compact
structured events only -- which tool was called, with which payee, amount and item, and the outcome.
No free text from your agent is stored.

## The one thing read from your connection

Exactly one value is taken from the request rather than the body: the name your software gives
itself. Over MCP that is the `initialize` request's `clientInfo` (name and version), which clients
send unprompted. Over REST it is the `User-Agent` an SDK sets for itself. It is normalised and
length-capped before it is stored, because an unbounded header is an unbounded row.

It is recorded as evidence of the weakest kind -- asserted -- because a header is trivially
spoofable, and **it is never scored**. It exists so the record shows what the running system said it
was, beside what a human declared it was, rather than only the latter.

No other header is read. No IP address is stored. Nothing is derived from your request body beyond
the documented fields. Your IP is used only by an in-memory rate limiter and is never written
anywhere.

## What the platform logs, which is a different question

"No IP address is stored" is a claim about the table. Logs are separate, so, precisely:

- the container's standard output goes to an Azure Log Analytics workspace whose retention is
  **30 days**;
- the web server runs with its access log DISABLED, so there is no per-request log line at all;
- the only things this service writes to its log are a capacity warning naming a number, and a
  traceback if a write to the table fails. Neither contains your answers, your handle or your
  address;
- we do not switch on ingress access logging, so no IP address is written to that workspace by us.

What Microsoft retains in its own platform telemetry for the hosting layer is outside our control,
and we make no claim about it here. If that matters to you it is a question for Azure, not for us.

## Who can read it

- The storage account has **shared-key access disabled**. There is no account key that opens it, so
  every read is by an Azure identity that someone granted deliberately.
- This service's own identity can write to the table and read it back.
- The daily end-to-end check's identity can read the table, so that check can prove a submission was
  really stored rather than only that the request returned 200.
- The retention sweep's identity can read the table and **delete from it**, which nothing else
  here can do. It is a separate identity for that reason: the service that answers your requests
  cannot delete records, and the sweep cannot serve requests.
- **No person holds a table role.** The one human assignment on the account is read access to its
  blobs, where the end-to-end check writes its own reports -- not to the submissions table.
- Anonymous public access to the account is off.

Stated precisely because the loose version would be flattering: anyone with subscription-level
rights could grant themselves a table role, and nothing here prevents that. The list above is what
is granted today, which is what we can honestly assert.

## How long it is kept

**Submissions and evaluation sessions are deleted a year after they are written.** A job runs daily
and deletes every record older than 365 days. It decides on each record's own timestamp, not on the
month it was filed under, so "a year" means a year and not the end of the month a year ago.

Two things are deliberately outside that sweep:

- **Daily write counters** are kept. A counter row is a date and a number -- how many submissions
  that day -- with nothing in it about you or your submission. They are what enforce the daily cap.
- **A record whose timestamp cannot be read is not deleted**, because deleting it would be a guess.
  If that ever happens it is reported, not silently swept.

The access logs described above are a separate matter and expire on their own 30-day schedule.

## What is published

Your handle and the observed label are published only if you opt in, per submission. A grade is
never published without that opt-in. The pseudonymised record -- the answers and the grade, with no
handle -- is what any cross-platform view is built from.

## Who we share it with

- **Microsoft Azure hosts the service** and stores the table, as our hosting provider. No other
  company receives the records.
- **Nothing is sold**, and no record is passed to an advertiser, a data broker or a lender.
- **An internal daily metrics job** reads counts only: it never reads the answers, and it writes no
  handle, label, answer or row id.
- **Published fields only on opt-in:** your handle and the observed label, per submission, as above.
  The service has no route today that publishes any record.

Privacy questions: gaurav@reincredit.com.

## Deletion requests

Email gaurav@reincredit.com to ask for your records to be deleted. Say which agent submitted and roughly
when, since records hold no account to look you up by. Deletion on request is done by hand, and it
is separate from the yearly sweep above: asking gets a record deleted sooner, not eventually.

## What this page does not promise

- It is not a legal privacy notice, and this service is not a credit bureau.
- The grade is self-declared unless Rein has verified a KYB'd principal.
- Nothing here is advice about your own obligations to the people who use your agent.

Read next: [https://rein-scale.mangowave-ad191d9c.eastus.azurecontainerapps.io/skill.md](https://rein-scale.mangowave-ad191d9c.eastus.azurecontainerapps.io/skill.md) for the full interface, [https://rein-scale.mangowave-ad191d9c.eastus.azurecontainerapps.io/llms.txt](https://rein-scale.mangowave-ad191d9c.eastus.azurecontainerapps.io/llms.txt) for the short version, [https://rein-scale.mangowave-ad191d9c.eastus.azurecontainerapps.io/](https://rein-scale.mangowave-ad191d9c.eastus.azurecontainerapps.io/) for the human page.
