Synth for WordPress
Scope: the WordPress service only. This document covers two WordPress plugins and the services they talk to: Synth Antispam, which classifies comments, and Synth Review Integrity, which keeps a record of a WooCommerce shop's product reviews. Synth Review Integrity has its own sections, starting at Synth Review Integrity: what we receive; elsewhere this page describes Synth Antispam unless it names Synth Review Integrity. It does not replace or cancel the general Synth service documents, which continue to govern the Telegram service: general Privacy Policy and general Terms of Use.
Companion document: Terms of Use — Synth Antispam and Synth Review Integrity for WordPress.
LightApps OÜ, registry code 16712872, Vesivärava tn 50-201, Kesklinna linnaosa, Tallinn 10152, Estonia — trading as Synth Antispam and Synth Review Integrity.
For anything about data processing, write to support@synth.locker.
Synth Antispam is a comment classification service. When you install the plugin on your WordPress site, you remain the data controller for your commenters' data, and Synth acts as a processor — on your instruction, and only to the extent you set by installing the plugin and entering a site key. We do not decide the purposes of processing on your behalf.
One purpose is ours, not yours, and you should know it before you install. Besides returning a verdict on your instruction, we keep a copy of the comment data your site sends and use it to train and improve the Synth spam-classification models. For that purpose we act as a controller in our own right rather than as your processor. It applies to every site using the service unless the site has connected Synth Review Integrity, Synth Antispam itself has no setting that turns it off, and the copies are kept indefinitely. What is kept, and how to have it deleted, is set out under How long we keep it and Training our own model below.
The legal basis for sending the data is yours to choose as the controller. This document does not set it for you.
Exactly and only the following. This list is generated from a manifest in the plugin's own code rather than maintained by hand, so what you read here and what the plugin actually sends cannot drift apart.
| Field | What it is and why it is sent |
|---|---|
| Request format and routing | |
schema_version |
The version of the request format, so the service stays compatible with older and newer plugin versions at the same time. |
plugin_version |
The installed plugin version, for the same compatibility reason. |
surface |
A fixed label identifying that the request comes from the comment form, so future integrations can be told apart. |
object_type |
Whether the submission is a comment, a pingback or a trackback — the last two are machine-submitted and weighed differently. |
| The submission itself | |
content.body |
The comment text as submitted, before WordPress applies its own HTML filtering, so a link that filtering would strip is still visible to the classifier. |
content.author_name |
The display name the commenter typed into the comment form. |
content.author_url |
The website URL the commenter typed into the comment form. On the traffic this plugin was built against this field carries a large share of the spam signal, most of which is not visible in the comment text at all. |
content.author_email_domain |
Only the domain part of the commenter’s email address, used to recognise disposable or throwaway address patterns. The address itself is never sent. |
content.author_id_hash |
A one-way hash of the commenter’s email address combined with a secret value belonging to this site alone. It lets the service recognise that two comments on this site came from the same address without ever receiving that address. Because the secret never leaves this site, the same commenter cannot be correlated across two sites — that is impossible by construction, not merely switched off. |
| Context, without identifying the commenter | |
context.author_status |
Either “registered” (a logged-in user of this site) or “anonymous” (everyone else). Only these two values are ever sent: a commenter who already has an approved comment here is recognised as trusted and skipped before any request is built. |
context.is_reply |
Whether the comment is a reply to another comment. |
context.site_locale |
This site’s configured language, used as a hint for language-specific handling. |
context.client_ip_status |
Either “absent” or “direct” — whether the request carried an IP address at all. The address itself is never sent, in either case, and the plugin never tries to reconstruct a visitor address from proxy or CDN forwarding headers. |
context.has_user_agent |
Whether the browser sent a User-Agent string at all. The string itself is never sent. |
Separately, and not about any commenter: each request carries your site key, the plugin version and your site's domain. That is how the service knows which site is talking to it.
Each of the following is enforced by the plugin's code, not promised by this document:
@ is sent, and
the pseudonym is a one-way hash — never the original value. The salt belongs to your site alone,
so the same commenter cannot be matched between two sites: that is impossible by
construction, not disabled by a setting. Deleting the plugin destroys the salt, after which not
even your own site can link a previously sent hash back to an address.There are two separate stores, kept for two different lengths of time. Reading only the first would leave you with the wrong answer.
The verdict record. The verdict for each comment is stored for 24 hours and then expires automatically. Your site does not control that expiry — it happens on our side. This is the record your site collects its verdict from.
The training record. Separately, we keep our own copy of each request your site sends — the comment text and the author and context fields listed above, exactly as they were sent — together with the verdict our classifier produced and any later correction a moderator of your site makes to it. These copies are kept indefinitely. They have no expiry date, and no scheduled deletion runs against them. A site that has connected Synth Review Integrity is the exception: from that moment no such copy is taken — see Training our models.
They are stored as sent. They are not anonymised: the comment text, the display name and the website URL are kept in the form your site sent them. They are not aggregated and not reduced to statistics. The fields listed under What never leaves your site are not in them either, for the same reason as everywhere else on this page — we never receive those fields, so we cannot store them.
We use the training records described above to operate and improve the Synth service: training and evaluating the spam-classification models that run it, and nothing beyond that. Synth is one platform serving several surfaces — WordPress today, further integrations over time — and the records from all of them are used together: a comment that teaches a model what spam looks like improves detection everywhere, not only on the site it came from. This is the one purpose for which we use your site's data on our own account rather than on your instruction, and we say so plainly here rather than leaving it to be inferred from the field list above.
What this means in practice:
www. prefix counts as the same domain,
another subdomain does not. Disconnecting the shop does not lift the exclusion. Copies taken before
the shop was connected are kept until you ask for them to be deleted, as described below.Having them deleted. Write to support@synth.locker from the address associated with your site, naming the site. We delete the stored copies by site key, so one request covers every copy taken from your site rather than one comment at a time. We are building an operator tool for this deletion alongside the release that starts the collection; until it is in place the deletion is carried out by hand on request, and the address above is the route in either case.
To produce a verdict, the service passes comment content to an external language-model provider
acting as a sub-processor. Part of the list above reaches it:
object_type, the comment text (in normalised form), the display name, the website URL,
the email domain, context.author_status and the site language. What does not
reach it: the per-site pseudonym, the reply flag, and the two presence flags for IP address
and User-Agent. Those stay inside our own perimeter.
Payments. If you buy a paid allowance, the plugin sends you to the payment page of Stripe, which acts as a separate controller for the payment itself. Card details never reach your site, the plugin or this service: they are entered on Stripe's own page. We receive from Stripe only what is needed to attach the purchased allowance to your site key — the fact of a completed payment, its plan and its renewal date. No comment data is sent to Stripe, and buying is never required to use the plugin. Stripe's own terms and privacy policy govern that processing.
Synth Review Integrity. Its data is processed by Amazon Web
Services, acting as a sub-processor in its eu-west-1 region
(Ireland): AWS hosts and stores it, and Amazon Simple Email Service (Amazon SES), in
the same region, sends the dashboard's emails. The Synth Review Integrity plugin causes no request
to a language-model provider: its check for promotional reviews reuses the verdict Synth Antispam
has already given, without a new request. The only part of Synth Review Integrity that uses a
language model is the dashboard's assistant: the model runs in Amazon Bedrock, part of Amazon Web
Services, in its regions inside the European Union, and receives only what is listed under
Synth Review Integrity: the dashboard assistant.
This section is the single maintained place where sub-processing is described. The current sub-processor list is available from support@synth.locker on request, and any change to it is published here, on this page, before it takes effect. Other Synth surface documents point to this section rather than keeping their own copy of the list.
Our own processing happens in the European Union: the service runs in Amazon Web
Services' eu-west-1 region (Ireland), and that is where site records and verdicts are
stored.
The one exception is the Synth Antispam sub-processing step described above. The language-model provider that Synth Antispam uses is established in the United States, so producing a verdict involves a transfer of the listed fields outside the European Economic Area. That transfer takes place under the European Commission's Standard Contractual Clauses, incorporated into our data processing agreement with that provider. The fields excluded from sub-processing — the per-site pseudonym, the reply flag and the two presence flags — are never transferred, because they never leave our own EU perimeter at all.
No transfer of any kind happens before you enter a site key: an installation with no key sends no request.
Synth Review Integrity data is stored and processed in that same region, eu-west-1
(Ireland). Two things leave the region. The first is a message to the dashboard's assistant,
together with the facts the assistant passes to the model with it: these are processed by Amazon
Bedrock in Amazon Web Services regions inside the European Union and are not transferred outside
the European Economic Area — see Synth Review Integrity: the
dashboard assistant. The second is a dashboard email, which Amazon SES delivers to the
recipient's own email provider.
The measures below are properties of how the service is built, not aspirations:
Data subject requests are addressed to you as the controller, and we are obliged to help you answer them. The plugin implements WordPress's standard personal-data export and erasure tools for the classification records it stores locally on your site.
On our side, the verdict record is removed by the 24-hour expiry described above. The training record is not: it has no expiry, so it is removed only when we delete it. Write to support@synth.locker, naming your site, and we delete every training copy taken from it. Deletion is by site key and therefore covers the whole site at once; we cannot currently delete the record of one individual comment while keeping the rest, and we say so rather than implying a granularity we do not have.
Synth Review Integrity is a separate WordPress plugin for WooCommerce shops. It keeps a record of
how product reviews are created, edited and moderated, and the Synth Review Integrity dashboard at
reviews.synth.locker shows which reviews and actions need attention. As with Synth
Antispam, for the data about reviewers that your shop sends, you remain the controller and we
act as your processor.
The plugin has its own site key. Pressing Get a site key asks the service for that key with an empty request; facts about reviews are sent only after an administrator connects the shop with a pairing code from the dashboard. Sending always happens later, in the background — never while a visitor is submitting a review.
Sent by default: facts, not content. For each product review:
Every request after that also carries the plugin's site key, its version and your site's domain. From time to time the plugin also sends a snapshot of the shop's technical set-up (such as the versions of WordPress, PHP and WooCommerce) and of its own local queue. The plugin's readme lists every field one by one, in its External services section.
Never sent: the reviewer's name, email address, website and IP address; the order number, its contents, amounts and addresses. Orders are read on your site only.
Review text — only with the Owner's consent. The text of a review is sent only after the Owner of the shop's workspace switches on content consent in the dashboard, and only for review versions recorded after that. Switching it off stops new transfers at once; text already received is kept until the end of its term, below. Stored text is shown to the members of the shop's workspace, whatever their role, and it is never included in an export.
Never used for training. What Synth Review Integrity sends is never used to train any model. Connecting a shop also excludes its domain from the training records of Synth Antispam — see Training our models. Synth Review Integrity does not change what Synth Antispam does: if Synth Antispam checks the same reviews, the comment text it sends is handled as described in the rest of this document.
o***@example.com,
shown only to the Owner of your workspace. The full address exists only while a request is being
handled, and is passed to Amazon SES to send you an email. One exception: if an email to an address
bounces or is reported as spam, Amazon SES keeps that address on our account's suppression list,
so that no further email is sent to it.no-reply@reviews.synth.locker.The assistant is part of the Synth Review Integrity dashboard at reviews.synth.locker, not of the plugin: the plugin sends nothing new because of it, and nothing reaches the language model behind the assistant until a person sends it a message.
The assistant answers questions about Synth Review Integrity and about your shop's data in the dashboard. When you confirm it, or ask in your message for it to be sent or saved, it also records a suggestion for the product or a problem report.
It is an AI. The answers are written by an AI language model and can be wrong. The assistant panel says so before you send your first message.
What the model receives, and only once you send a message:
What the model never receives:
Who runs the model, and where. The model is run by Amazon Web
Services in its Amazon Bedrock service — the same sub-processor named
under Further processing. The company that made the model does not
receive your messages or the answers and has no access to them. Amazon Bedrock processes each
request to the model in one of these Amazon Web Services regions inside the European Union:
Frankfurt, Stockholm, Milan, Spain, Ireland and Paris. No request goes to a region outside the
European Union, so the assistant involves no transfer outside the European Economic Area. Amazon
Web Services may keep requests to detect abuse; if it does, they stay in the region where they were
processed, inside the European Union. Conversations and the records below are
stored in the eu-west-1 region (Ireland), like the rest of Synth Review Integrity.
How long we keep it.
A suggestion or a problem report keeps your message as you wrote it, a short description written by the AI, the dashboard page you had open, and the shop and the finding it concerns, if any. A problem report also keeps technical details copied from the shop's records in Synth Review Integrity: the versions of the plugin and of WooCommerce, the state of the plugin's queue, when the plugin last reported its status and, for a finding, its last recheck. A reason you give under Not really, with any comment you type, and the text you type under Other when asked what brought you to Synth are kept the same way, for 12 months.
To have a workspace's conversations and records deleted sooner, write to support@synth.locker, naming the workspace.
Not used for training. Nothing you send to the assistant and no answer it gives is used to train models — not by Synth, and not by Amazon Web Services or the company that made the model.
One script. The assistant panel runs on one script of the dashboard's own,
served from reviews.synth.locker: no analytics, no advertising and no third-party
resources. The rest of the dashboard works without it.
Our role for suggestions, problem reports, the reason you came and usage events. For your shop's data that the assistant reads, and for your conversations with it — your messages, the answers and the record of what it looked up — nothing changes: you remain the controller and we act as your processor. Suggestions and problem reports, your answer to what brought you to Synth, and usage events are different: they are Synth's own records of how people use the dashboard, and for them we act as a controller in our own right rather than as your processor. Our legal basis for them is our legitimate interest in improving Synth, under Article 6(1)(f) of the General Data Protection Regulation (GDPR). The Synth team reads them and uses them only to improve Synth.
Everything is stored in Amazon Web Services' eu-west-1 region (Ireland), encrypted at
rest.
| What | How long |
|---|---|
| Review text | Only with the Owner's consent. Kept while the review is part of an open case and for 30 days after the case is closed; a review that is in no case, 30 days. Erasing the review removes it. |
| The journal | Review versions, check results and the record of actions: 12 months, then they expire. The journal is append-only — no part of the service is able to change or delete a journal record — which is what makes it usable as evidence. |
| Current records | The current facts of each review, findings and cases: for as long as the workspace exists, unless erased. |
| Backups | Point-in-time backups, 30 days. After a restore, every erasure is applied again, so erased data does not come back. |
| Exports | Export archives, 7 days. Each download link is valid for 15 minutes. |
| Links in emails | 15 minutes, single use. |
| Service logs | 30 days, without review text, email addresses or links. |
| On your site | Records waiting to be sent: at most 7 days. An erasure request waits until it has been sent. |
| The assistant | Conversations with the dashboard assistant and the records it keeps have their own terms — see Synth Review Integrity: the dashboard assistant. |
What we keep about a person's reviews is erased with WordPress's privacy tools. When you erase a person's personal data under Tools → Erase Personal Data, the plugin finds that person's product reviews on your site by their email address and sends us only the numbers of those reviews and a random request identifier — never the email address. Under Export Personal Data, the plugin describes what it sends about each of that person's reviews. While the shop is disconnected, the request waits on your site and is sent once the shop is connected again. The store owner can also erase a review from the dashboard, on the page of that review — for example while the shop is disconnected. If neither works, write to support@synth.locker naming the shop and the review numbers, and we erase them.
Erasure removes a review's facts, its text and its evidence of purchase; the review then appears as erased in the dashboard and in exports. A marker of the erasure is kept, so that the review cannot be recorded again from a late or repeated event, or brought back by a restore from backup. One limitation remains, and we state it as it is: the append-only journal keeps salted digests of the erased reviews in storage until the 12-month TTL; they are never output. Backups keep erased data for up to 30 days.
Disconnecting a shop or uninstalling the plugin is not erasure. Disconnecting keeps the shop's history with its workspace. Uninstalling deletes, on your site, the plugin's local queue and the secret value behind every digest and pseudonym; what the service already holds stays until it is erased or expires. Copies of the same reviews that Synth Antispam received, if it is installed, are outside this erasure: its verdicts expire after 24 hours, and its training copies are deleted on request, as described under Training our models.
When our processing changes, this page changes with it — not on an annual schedule. The date at the top is the date of the last substantive change. Changes to the sub-processor list are published here before they take effect.