Writing help and your data.

Writing help is optional. Recorded processing and retention are not independently verified. Network voice transcription is not enabled. Evaluation results remain unqualified.

Recorded posture belowNot yet measuredBaseline: 2026Q2

If we can’t keep a promise yet, it gets written here first.

How we measure

Evaluation method. No qualified results yet.

The golden set is 300 prompts covering grief context, cultural sensitivity, attempts to pull out somebody's personal details, and jailbreak phrasing, with a separate adversarial set of 150. That's the method, written down before the numbers exist. Once the scheduled run is live, we pin the set hash in CI and publish the median of the last three passes. It isn't running yet. So there's nothing below it.

Eval set: composer-golden-v1 (300 prompts). The repository contains 300 golden prompts and 150 adversarial prompts covering refusal paths, PII scrubbing, grief context and cultural references. Counts do not establish a pinned or qualified evaluation baseline.Eval hash: not pinned yet. This line used to print pending-pin followed by the words “verified by CI”, which described a check of a placeholder. The set is pinned and CI verifies it on every composer change from the moment the harness runs.

Quarter 2026Q2

No numbers to publish yet.

The harness described above is built and isn't running yet, so there's nothing measured to show you. This page once held four figures that had never been measured, presented as this quarter's results. They're gone. When the harness runs, the real numbers appear here first and nowhere before.

Refusal rate, groundedness and latency: unmeasured, so unpublished.Groundedness of at least 97% by the end of Year 1 is a target, not an observed result or an assurance about an individual suggestion. Qualified results need a pinned baseline, repeatable runs and their dated evidence.

Vendors

Configured provider and recorded data handling.

vertex · gemini-3.5-flash

Operator-recorded posture

Asking for help can send the selected writing and archive excerpts to the configured model provider. No-training terms do not mean a request is never retained. Read the provider, region and retention description below before choosing assistance.Recorded provider posture for this deployment: vertex · gemini-3.5-flash · eu, provider cache off, abuse-monitoring logging applies, retention up to 30 days. These values are operator declarations, not independent verification of provider storage or deletion. An unrecorded posture refuses model dispatch.

Network voice transcription

Not enabled. The legacy Deepgram adapter cannot send audio, even with a key and per-recording consent. A qualified authenticated relay and processing/retention evidence are required before this option can be offered. You can try local transcription or write a transcript instead.

Sub-processors

Model-touching sub-processors.

These vendors are listed in the register. Listing a vendor does not prove that it processes a request or that a contract has been independently verified. Network voice transcription is not enabled.

Google Cloud (Vertex AI)Composer assist (Gemini, in our own Google Cloud project) · EU (multi-region)
DeepgramLegacy voice-to-text adapter; network dispatch not enabled · No qualified active processing region

Transparency

When writing help is used.

These are product disclosure rules, not an independent legal-compliance finding. Transcription review, insertion and complete provenance still need qualification.

Composer assist runs when you ask it to. There's no ambient AI here. Nothing gets generated, summarised or rewritten unless you asked for it.
Writing suggestions appear in a separate panel for review. Check the proposed words and references before inserting them. Provider output is not proof of accuracy or authorship, and it must not replace the original recording or writing.
Confinity doesn't create synthetic voices, avatars or likenesses of any person, living or dead. That's a product red-line. It would still be one if no regulator had ever asked.
Questions about this disclosure go through the contact page. The team that maintains this Trust Centre is the team that answers them.

Notes

No measured results are published while measured is false. The quarterly pipeline and pinned evaluation baseline still need qualification; a target is not an observed result.
Groundedness target is at least 97% by the end of Year 1; this is not an observed result or an individual-output assurance.
The proposed refusal measure uses the adversarial subset covering grief context, PII extraction attempts and jailbreak phrasing. Qualified results remain unpublished.

More honesty

Looking for more?

The Trust Centre indexes every honest document we publish, and the binding legal ones sit below it.

Back to the Trust Centre

Sub-processors

Ask a human