Describe the language. Never the person.
Receipts uses language models to identify patterns in conversations. That carries obligations that don't end when the product ships. This page is the full policy, and the things we hold ourselves to even when nobody is looking.
The plain version.
Receipts reads the words in a conversation and surfaces named patterns like "blame-shifting" or "permission-seeking" with the exact quotes they're drawn from. It does not score people, predict future behaviour, or call anyone an abuser. The interpretation is always yours.
How the analysis works.
When you analyse a conversation, the following happens on our servers:
- 01Messages are segmented and normalised. Timestamps and message direction are preserved; names are replaced with role tokens (Person A, Person B) before anything is shown to the model.
- 02Each message, and its surrounding context, is sent to the Anthropic Claude API and classified for 23 named patterns. We describe each pattern in plain language on the Product page.
- 03Hits are scored by the model and ranked by confidence. Low-confidence matches are shown in a separate "weaker signals" tray, not mixed into the main list.
- 04Every finding links back to the exact quote. No pattern is ever presented without its source text.
Analysis happens on our servers via the Anthropic Claude API. Your conversations are encrypted in transit and at rest, never used to train any model, and never shared with third parties beyond the API call itself.
What the analysis can't do.
- ·It can't tell you whether the other person is abusive. That's a judgement only a human can make, usually in conversation with a professional.
- ·It can't predict what will happen next, or assess danger. It describes what has already been said, nothing more.
- ·It can't pick up tone of voice, facial expression, or silence. It only sees text.
- ·It can't speak for cultures or contexts it wasn't trained on. Sarcasm, irony, and in-jokes are routinely missed.
- ·It doesn't know who is speaking. You can correct the labelling; we cannot verify it.
What we train, and how.
The classifier is trained on a mix of published research corpora (openly licensed), synthetic data generated under strict review, and a small corpus of volunteer-consented conversations from domestic-violence researchers, used with written permission and never retained after training.
When we retrain the model, the list of datasets, the training date, and a summary of changes are published on our model card before the new version ships.
Bias, language, and edges.
Language models reflect the language they're trained on. Ours is strongest on English, weaker on Australian English idioms, and weaker again on bilingual or code-switched conversations. We document known weaknesses rather than hiding them.
- ·We test every release against a held-out set of conversations from culturally and linguistically diverse sources. Failures are tracked in public.
- ·We deliberately tune for false-negative over false-positive. Missing a pattern is safer than inventing one.
- ·We don't use demographic signals (gender, age, orientation) as features. If you tell us the relationship type, we use that to frame the output; we don't use it to score.
Human oversight.
Pattern definitions are written and reviewed by hand before they ship. When we change a definition, the previous version stays in history, nothing is silently rewritten, and you can see which version of the model produced any given analysis.
We're a small team, and we're honest about what that means: we don't currently have a standing clinical or legal advisory board. We consult case-by-case with practitioners we trust when we're drafting sensitive patterns, and we're building a more formal review group as we grow. When that group exists, it will be named on this page.
Red lines we don't cross.
- ·We will never rate a person. No abuser scores, no compatibility scores, no personality profiles. Ever.
- ·We will never generate synthetic messages that could be mistaken for real ones. No "what they probably meant" replies.
- ·We will never sell access to the model to an insurer, employer, or court-facing third party.
- ·We will never ship a feature that uses detection to nudge a user towards leaving, or staying. The decision is theirs.
- ·We will never allow a model output to be presented as fact. Everything is framed as what the language appears to be doing.
Report a mistake.
If Receipts missed something important, named something wrongly, or produced output that felt unsafe, tell us. Every report is read by a human within two business days and logged against the model version that produced it.
contact@receipts.love · or use the "Flag this finding" link inside any analysis.