Safety · published
An astrology engine
that can't frighten people.
Most astrology APIs will say whatever the stars “say”, including that someone will die, or that their marriage is doomed, or that they should stop their medication. Ours can't. Every reading passes a six layer safety guard inside the engine before it can reach a person, and you can run that guard against your own text.
This is about the words a reading uses. The raw calculations are facts, and how we check those is the accuracy report. As far as we know, this is the only published safety layer in the astrology API lane.
What goes in
“Is this marriage doomed? Give me a score.”
Your strengths as a couple, the areas to nurture, and a few small things that help.
No verdict on two people ever reaches a user.
If you show astrology to users, you inherit its risk: the fear prediction before a crisis, the “this match will fail” that breaks up a family, the medical claim that delays care. It becomes your incident, your support ticket, your app store review. For matrimony, dating and media platforms especially, a safe engine isn't a nicety. It's the liability you don't want.
layers, in order, on every reading we write
red team cases, all passing as of June 2026
fear verdicts that can reach your users
endpoints that make the guard observable
The six layers
Built into the reading pipeline, in this order, on every interpretive call, never bolted on afterward.
Facts come only from the engine
A reading describes what the engine actually computed, never what a model imagined. The interpreter is never allowed to invent a planet, a date, or an event. A model that can't invent facts can't invent a catastrophe.
A hard refusal list
Death, self-harm, serious illness, “should I stop my medication”, doomed marriages, financial ruin: these are refused outright, on the question and on the answer. The filter also catches the tricks people use to get around it — numbers written in place of letters, extra spaces, spelling mistakes — so “su1c1de” and “di vorce” don't get through. It's careful the other way too: “Cancer” the sign passes, “Cancer” the disease does not.
A model that catches the rest
On top of the fixed filter, an AI model reads the question and the answer for wording the rules miss. It can only ever add a refusal, never remove one. If it's slow or offline, the request quietly falls back to the safe answer the fixed filter already guaranteed.
Reframing, with agency
Astrology's frightening phrasings are rewritten into honest, useful ones. “Saturn brings obstacles” becomes “Saturn invites patience: slow, deliberate work.” The astrology stays true, the fear leaves. And a named challenge is never left without a next step you can take.
The marriage rule
The surface that has caused the most real harm gets the strictest rule. No reading ever returns “incompatible”, “do not marry”, or a pass or fail score like “18/36”. Compatibility is your strengths as a couple, the areas to nurture, and simple remedies, never a verdict on two people.
A final tone check
“Warm, never fearful” means never fatalistic, not never honest. A heavy word used hopefully (“you've moved through grief”) is left alone. A fatalistic one (“doomed”) is softened in place so the real reading survives. Only the rare line that resists repair is replaced with a calm template.
Human review
Every reading we flag, plus a sample of the clean ones, goes into a queue that a Jyotish (Vedic astrology) scholar and a wellness editor read through by hand. What they find becomes a new rule. The guard learns, and the loop is never switched off to save writes.
Watch it work
Real inputs a user, or a careless model, might send. What comes back is refused or rewritten, with the exact flags. The fear leaves, the honest reading survives.
Asked
“When will I die? Is it a serious illness?”
Astrology can't speak to health, or to how long a life will be, and we won't pretend otherwise. If you're worried, please talk to a doctor. What your chart can offer is a sense of your resilience and the seasons ahead.
Asked
“Is this marriage doomed? Give me a score out of 36.”
Here are your strengths as a couple, the areas worth nurturing, and a few small things that help. Never a verdict on two people, and never a pass or fail score.
Asked
“Saturn brings obstacles and delays for me, right?”
You're in a Saturn season, and it rewards patience over speed. Slow, deliberate work pays off now. One thing to focus on: pick the single commitment that matters most and give it your steady best.
Don't take our word for it
Two endpoints make the guard observable. Every reading from /v1/interpret comes back with a safetyFlags receipt of exactly what the guard did. And /v1/safety/check lets you run your own text, or a model's, through the guard before you ever show it.
POST /v1
// Run any astrology text, yours or a model's, through the guard.
// in
{ "text": "Your marriage is doomed to fail.", "isCompatibility": true }
// out
{
"verdict": "modified", // pass | modified | refused
"text": "…strengths and what to nurture together…",
"safetyFlags": ["marriage-rule", "tone-repaired"]
}
POST a death prediction, a doomed marriage, or a “you'll go bankrupt” and watch it come back refused or modified. The guard describes itself on every call.
Held by a red team suite
The guard is attacked, on every release, by tests that fire forbidden questions, paraphrases, euphemisms and typo evasions at it, and by tests that make sure ordinary questions are not refused. We publish the real count, not a round one.
red team cases, all passing as of June 2026. The number grows every release, and a new leak becomes a permanent test. The harness is being prepared for open source, so you'll be able to run them yourself.
What we won't do
We don't predict your future
And we never quote a “% accurate”. Prediction can't be proven, and claiming it is a real legal risk. We verify the part that can be checked, the astronomy, out in the open.
We don't issue verdicts
No marriage verdict, no medical opinion, no timing of death. Not as a policy we promise to keep, but as something the engine is built so it cannot do.
We don't quietly let things through
The fixed filter always stands. The AI layer may drop out if it's slow, but the fixed wall behind it never lowers.
Why safety is the product
Astrology has frightened people for centuries. We built one that can't.
A “doomed match” isn't a harmless reading. In matrimony astrology, people labelled with a dosham have been turned away from marriages, and a fear prediction can land right before someone's hardest moment. We didn't want a policy that promises to be kind. We wanted an engine that cannot be cruel, where the calm isn't a tone we apply at the end, but a wall the words have to pass.
Test it against your worst case
Request a sandbox key (real data, no mocks) and run our guard against the most fearful astrology text you can write. Then decide.
Request a sandbox key