diff --git a/backend/docs/research/identity-and-ownership.md b/backend/docs/research/identity-and-ownership.md
new file mode 100644
index 0000000..31108d2
--- /dev/null
+++ b/backend/docs/research/identity-and-ownership.md
@@ -0,0 +1,116 @@
+# Identity and ownership
+
+## Problem
+
+The system needs a concept of who the user is to have uploaded files have owners. Owners have permissions on files,
+whereas non-owners do not.
+
+## Design
+
+### Storing users and credentials
+
+Taking inspiration from [OWASP's guidance on storing
+passwords](https://cheatsheetseries.owasp.org/cheatsheets/Password_Storage_Cheat_Sheet.html), Argon2 can be used to hash
+passwords with a user-specific salt.
+
+#### Comparables
+
+Looking at comparable applications for inspiration about which dependencies to rely on and the standards that exist.
+
+##### Django
+
+[Django](https://docs.djangoproject.com/en/4.2/topics/auth/passwords/) uses PBKDF2 with a SHA256 hash by default, but
+allows different hashers to be set by users (incl.
+[Argon2](https://docs.djangoproject.com/en/4.2/topics/auth/passwords/)).
+
+Their [Argon2
+hasher](https://github.com/django/django/blob/517d3bb4dd17e9c51690c98d747b86a0ed8b2fbf/django/contrib/auth/hashers.py#L374-L473)
+[uses](https://github.com/django/django/blob/517d3bb4dd17e9c51690c98d747b86a0ed8b2fbf/setup.cfg#L48-L50) the `argon2-ccfi` [library](https://github.com/hynek/argon2-cffi).
+
+Password hashes are stored as a `VARCHAR(128)` and stores an ASCII string prefixed with the algorithm used and
+containing the hash and the salt as a base64 encoded value. This is handled by [the underlying
+library](https://github.com/hynek/argon2-cffi/blob/e9473c8f0b8b860bb4369d11f5a605a326255f3f/src/argon2/low_level.py#L53-L118).
+
+The hashed value can be split into parts and `argon2-cffi` can [retrieve the
+salt](https://github.com/hynek/argon2-cffi/blob/e9473c8f0b8b860bb4369d11f5a605a326255f3f/src/argon2/_utils.py#L95-L140) for password verification.
+
+##### FastAPI Users
+
+[FastAPI Users](https://github.com/fastapi-users/fastapi-users) uses BCrypt [by
+default](https://fastapi-users.github.io/fastapi-users/12.1/configuration/password-hash/) and does not offer
+alternatives out-of-the-box without [custom
+code](https://fastapi-users.github.io/fastapi-users/12.1/configuration/password-hash/#full-customization) being supplied by adopters.
+
+##### NextCloud
+
+[Nextcloud](https://nextcloud.com) uses Argon [by
+default](https://docs.nextcloud.com/server/19/admin_manual/configuration_server/config_sample_php_parameters.html?highlight=htaccess%20rewritebase#hashing).
+
+#### Dependencies
+
+`argon2-cffi` is a good candidate as backing for authentication; being used by Django, it's likely to be closely vetted
+for quality.
+
+#### Table design: `users`
+
+|Key|Type|Notes|
+|---|---|---|
+|`id`|`bigint`|User ID, primary key.|
+|`username`|`varchar(64)`|Unique username.|
+|`password_hash`|`varchar(128)`|Hashed password, prefixed with algo.|
+|`created_at`|`datetime`|UTC datetime of record creation.|
+|`password_updated_at`|`datetime`|UTC datetime of the last update to the hashed secret, for renewal tracking.|
+|`updated_at`|`datetime`|UTC datetime of last record update.|
+
+The password-storing scheme is largely inspired from Django's, no reason to deviate. Prefixing the algorithm opens the
+door to user customization in the future and to changes in algorithm if need be.
+
+Usernames are unique across the table and should be used to refer to the user externally (that way, no leaking of
+sequential IDs).
+
+`username` are initially meant to be immutable, but there's no harm in having those be updateable. They do need to be
+indexed for searching though.
+
+### Representing ownership
+
+The user table tracks individual users, and the files table tracks file entities. A third table should track the
+relationships between the two. This would give entities flexible ownership (i.e. what if a given file could have
+multiple owners?).
+
+"Ownership" is too rigid a concept to be represented without needed to be modified a bunch in the future. It might be
+best to represent _permissions_ instead such that "owners" have all the permissions on something. This facilitates the
+creation of "shared resources" since users that would get files shared to them would just have reduced permissions on
+those files. 
+
+Permission representation should be flexible such that we can add different permission types along the way. For that
+reason, having a convention such that permissions are stored as a number whose bits represent individual permissions is
+probably best.
+
+Storing permissions as `bigint` would provide 64 different bits that can be encoded as different permissions. In
+principle, we could represent the number as a string and base64 encode it so that the format is more flexible, but
+that's not really necessary (a 64b number should be more than enough to account for all cases of permissions. This also
+allows it to be indexed, making different levels of share and ownership searcheable without too much trouble.
+
+Permissions can be updated.
+
+#### Sample permissions
+
+Some permissions that we'd need could be:
+
+- Can read file;
+- Can edit file;
+- Can delete file;
+- Can share file;
+- Can copy file;
+- ...
+
+#### Table design: `permissions`
+
+|Key|Type|Notes|
+|---|---|---|
+|`id`|`bigint`|Permission entry ID, primary key.|
+|`user_id`|`bigint`|Foreign key to the users table, the user who has the permission set.|
+|`file_id`|`uuid`|Foreign key to the file that the permission applies to.|
+|`value`|`bigint`|Permission value. The bits represent individual permissions.|
+|`created_at`|`datetime`|UTC datetime of record creation.|
+|`updated_at`|`datetime`|UTC datetime of last record update.|