Re-checked in the background
A share of the answers code confirms is decided again, after the response, by the careful AI the rules were learned from. A rule whose answers drift is switched off on its own, and you are told.
For developers building workflows clients can trust
Your clients trust code, not LLMs. Give them both: Jev answers in a fifth of a second, a rules engine confirms every answer, and anything it can't confirm comes back to you. One API call.
The check is code, not another AI.
small-within-window · 0.2 sneeds-person-review · 0.2 sHow it works
While a gate learns, a careful AI decides each request by your policy. DataSnack writes the questions Jev will answer about each request, and rules over Jev's answers, and tests every rule against your history.
Switch the gate on: Jev reads each request in a fraction of a second, and the rules must reach the same answer, firmly, before it goes out. No AI checks the AI.
What code cannot confirm comes back unanswered, with the reason, for your own AI or a person. Their answer, sent back, is repeated for that request and the gate learns from it.
POST /v1/gates/refund-emails/decide
{"input": {"customer_tier": "plus",
"email": "The $40 lamp came last
week and I've changed my mind…"}}
{ "output": {"decision": "approve"},
"decided_by": {"tier": "checked",
"rule": "small-within-window",
"answered_by": "jev-1.14.0"},
"ms": 190 }Built on Jev
Jev, TypeSafe's fast model, answers typed questions about text in about a fifth of a second, for a tiny fraction of a large model's price. On its own it is right most of the time and wrong without warning. Checked by rules learned from a careful AI, it is right when it answers, and says so when it cannot.
| On 150 refund emails it had not seen | Answered | Agreed with the careful AI |
|---|---|---|
| Jev alone, every email | 100% | 85% |
| Jev alone, only when it is 90% sure of itself | 53% | 99% |
| Jev, confirmed by DataSnack's rules | 73–79% | 99–100% |
Our own test, on synthetic refund emails: rules learned from 250 emails the careful AI decided, then scored on 150 more against its decisions; the range is two runs. Your numbers depend on your decision; the dashboard shows them from your first day.
It keeps checking
A share of the answers code confirms is decided again, after the response, by the careful AI the rules were learned from. A rule whose answers drift is switched off on its own, and you are told.
When Jev releases a new version, your recent requests are read with both and compared. If answers would change more than reading again changes them, nothing moves until you decide.
Every rule and every config is versioned. Roll a rule back to shadow, or the whole gate back to learning, in one click.
Three ways to use it
For decisions made from text: emails, tickets, notes. Jev reads, rules confirm. Jev is included in every plan.
For decisions made from typed fields: rules answer in microseconds, the same answer every time. Deterministic, end to end.
Already a TypeSafe customer? Read with your own key from Settings; your gates use it and TypeSafe bills you directly.
Using it
Each gate is an API endpoint from the moment you create it, and a tool on your workspace's MCP server. You integrate once, while the model is still deciding. As rules take over, the request and the response stay the same; the response says what answered.
POST /v1/gates/ad-approval/decide
Authorization: Bearer dsk_…
{"input": {
"price_monthly_usd": 249,
"term_months": 36,
"market": "US-CA",
"channel": "meta",
"flight_end_day": 20261220,
"legal_line_present": true }}
{ "output": {"decision": "reject",
"problem": "price"},
"reason": "The approved price is $299 a month.",
"decided_by": {"tier": "rule",
"rule": "price-mismatch",
"version": 2},
"ms": 0 }"mcpServers": {
"datasnack": {
"url": "https://datasnack.ai/mcp",
"headers": {
"Authorization": "Bearer dsk_…"
}
}
}
tool decide_ad_approval input the gate's input schema result output, reason, decided_by
Rules replace model answers behind the same endpoint. Nothing in your workflow changes when a rule goes live, or when one is rolled back.
Every answer names the rule and version that gave or confirmed it, and on a text gate what Jev read. A rule is code evaluated from memory, in a Go rule engine.
Rule answers need neither our database nor a model provider. If Jev cannot be reached, a text request comes back unanswered within two seconds, with the reason, for your workflow to handle.
Where it fits
Write the policy, send real requests, and switch it on when its rules pass their tests. One API call. Free while in early access.