REPO REVIEWARCHITECTURE & SECURITY Download PDF

Data handling statement

What happens to your repository

For your security reviewer: where your code goes, how long each piece stays, who can read it, and what we do not yet guarantee. Every retention period below is enforced by configuration, not by a person remembering to delete something.

Version
1.0
Last reviewed
16 November 2026
Stage
Private alpha
Region
AWS Sydney (ap-southeast-2)

In short

  • Read-only access through a GitHub App you install and can revoke at any time.
  • One copy of one branch, with no history, deleted one day after the run.
  • Reports deleted 30 days after the run, or at once when you ask.
  • Everything stored in Sydney. Parts of your code are sent to one model provider, Anthropic, and nowhere else.

1In one table

WhatWhereKept forWho can read it
Access to your repositoryGitHub, through the Architecture Automation GitHub App you installuntil you uninstall itthe service, read-only, one repository per token
The per-run access tokenAWS Systems Manager, encrypted, one parameter per rundeleted after the clone; expires within the hour regardlessthe clone task only
A copy of your code (one branch, depth 1, packed)S3, private bucket, Sydney1 day, then deleted automaticallythe analysis task
The working copy during analysisthe analysis container's own diskthe length of the run (under 3 hours); the container is destroyed afterthe analysis task
Parts of your code sent for analysisAnthropic's Claude model API (section 4)per Anthropic's termsAnthropic, as processor
The documents producedS3, same bucket30 days, then deleted automaticallythe reviewer; you, through links once released
Run log and diagnosticsS3 beside the documents; CloudWatch Logs30 days (S3); 14 days (CloudWatch)the operator only
The run record: repository URL, branch, commit, status, ratings, finding countsDynamoDB, encryptedfor the life of your accountthe operator; you, through the portal or API with your key
Fixed scores: the seven ratings and the health verdict for each state of your code, so a rerun of unchanged code reports the same scores. Numbers only, never codeS3, same bucket, under a hash of your repository URLfor the life of your accountthe analysis task; the operator
Your email address, if you signed up for a free SnapshotDynamoDB, on your keyfor the life of the key's recordthe operator; the service, to send you the key and the release
Network log: each HTTPS connection the analysis made or tried, by host name and time, never its contentCloudWatch Logs30 daysthe operator only
Audit trail: who issued, rotated or revoked a key, started, released, withheld or deleted a run, or was handed links to its documents, and whenCloudWatch Logs1 yearthe operator only

2The path your code takes

  1. Access. You install the GitHub App on the repositories you choose. Nobody hands us a password or a long-lived token. You can see and revoke the installation in your own GitHub settings at any time, and revocation takes effect at once.
  2. Clone. When your run's turn comes, the service asks GitHub for a token valid for one hour, read-only, for that one repository. A short-lived task clones the requested branch at depth 1 (no history, no submodules, no git hooks run), packs it, uploads it to the private bucket and stops. The token is deleted.
  3. Analysis. A second task, in a private subnet with no inbound access whose only way out is an allow-list of hosts (section 3), unpacks the code, builds a structural graph of it and runs the analysis. The analysis reads files and sends the parts it needs to the model API. The documents are written to the bucket; the task and its disk are destroyed.
  4. Review. By default, a person reads every report before it is released to you, and nothing is sent to you or anyone else until then. A key can be set up with you to release a report automatically when every check passes; a report with any failed check always waits for a person.
  5. Release. You receive links to the documents. Each link expires after 7 days; the documents themselves are deleted 30 days after the run. Download what you want to keep.
  6. Deletion. The packed copy of your code is deleted one day after the run, whatever happens to the run. The bucket is versioned, and earlier versions are deleted on the same schedule. You can delete a run's documents yourself at any time from the portal, or ask us to: either removes every version of every file the run left, at once.

3Storage and transport

Region
Sydney (ap-southeast-2) for every stored item. Code sent to the model provider leaves Australia: see section 4.
Encryption at rest
S3 server-side encryption on the bucket; DynamoDB encryption; access tokens and the GitHub App's private key as SecureString parameters, the key under its own KMS key.
Encryption in transit
TLS everywhere. The bucket refuses any request not made over TLS.
No public access
The bucket blocks all public access. Documents are reached only through expiring signed links or the operator's console, which requires sign-in.
Outbound allow-list
The analysis can connect only to the model API and the AWS services that run it (the container registry, logs, and Parameter Store for the model credential), checked by host name at a filter it cannot bypass, and to our own storage bucket through a private endpoint that refuses every other bucket. Every connection it makes or tries is logged by host name.
Least privilege
Each part of the service has its own role. The clone task reads only run tokens and writes only packed repositories; the analysis task reads packed repositories and writes documents, and the code under analysis runs without those AWS credentials at all.

4The model provider

The analysis is done by Anthropic's Claude. To analyse your repository, the model reads files from it, so portions of your source code are sent to Anthropic's API and processed in Anthropic's infrastructure, which may be outside Australia.

  • Today (private alpha): the service runs on a Claude subscription, used only on the operator's own repositories and those of customers who have agreed to that in writing.
  • Before general availability: the service moves to Anthropic's commercial API, whose terms state that inputs and outputs are not used to train models by default. This statement will be updated with the date.

5What we do not keep

  • Your code after one day.
  • Any access token after the clone.
  • Your GitHub identity: the App acts as itself and never signs in as you.
  • What a run cost, in or beside your documents.

6What we do not yet guarantee

Stated plainly, because your reviewer will ask.

Network during analysis
The analysis may run parts of your repository's own code (its test suite, for example). Its HTTPS connections are limited to the allow-list above and it can reach nothing else inside our network, but name lookups are not filtered, and the allowed hosts (the model API, AWS services) are shared services that code holding its own credentials for them could use to send data to its own account. This matters only if your repository's own code is hostile; it never exposes our credentials or your access token, which is already deleted.
The model credential, during the alpha
On commercial API terms the credential is held by a separate process and the analysis only ever has a placeholder for its own run, so code from your repository cannot read it. During the alpha's subscription it would be within reach of that code, so during the alpha your repository's own code is not run: its tests are reported as not run.
Free sign-up
Free sign-up is for public repositories only. A private repository is set up with you directly, so nobody who merely names it can request a report on it.
Commercial model terms
See section 4.
Certification
The service holds no SOC 2 or ISO 27001 certification.
Your own AWS account
Running the analysis inside your AWS account, so the code never leaves it, is planned for enterprise customers and not available yet.

7Your controls

  • Revoke access at any time: uninstall the GitHub App, or remove a repository from it.
  • Delete a run: from the portal, or by asking us, which removes every version of its documents at once. The run's record (repository URL, branch, commit, ratings) stays, marked deleted; ask us and we remove that too.
  • Ask what we hold: we tell you every run, its date, its repository and branch, and what is still stored for it.