An open-source decision engine. Models and scores tell you what is going on. Ordo decides what to do about it, with rules that have tests and versions, and shows which decisions a change moves before you ship it.
if jev.confidence > 0.85:
refund(order)That line decides who gets money back. It usually has no review and no test, and when the model is updated, 0.85 starts to mean something else.
Jev can tell that a customer wants a refund. It doesn't know your refund window, your amount limits, or how sure is sure enough to pay out without a person looking. That part is a decision, and it belongs in Ordo.
What to do with a refund request, written as a decision table:
config:
name: refund
version: 1.0.0
entry_step: route
steps:
route:
id: route
name: Route the refund request
type: decision_table
inputs: [jev.intent.choice, jev.intent.confidence, order.amount, order.days_since_delivery]
outputs: [action]
rules:
- when: ["*", "*", "*", "> 30"]
then: [HUMAN_REVIEW]
- when: [refund, ">= 0.85", "<= 200", "*"]
then: [AUTO_REFUND]
- when: [refund, ">= 0.6", "*", "*"]
then: [LLM_DRAFT]
default: [HUMAN_REVIEW]
next_step: done
done:
id: done
name: Done
type: terminal
result:
code: OK
output:
- [action, $action]$ ordo exec --rule refund.yaml --input '{"jev":{"intent":{"choice":"refund","confidence":0.72}},
"order":{"amount":129,"days_since_delivery":1}}'
code: OK
output: {
"action": "LLM_DRAFT"
}Jev is 0.72 sure the customer wants a refund. The rules say that is not enough to move money on its own, so an LLM drafts the reply and an agent approves it. To change the threshold, edit the table and run the tests again. The calling code stays the same.
Someone proposes auto refunding from 0.80 instead of 0.85. In the jev-routing rule pack, all 12 tests still pass after the edit. ordo impact runs it next to the last commit:
$ ordo impact refund-routing
impact refund-routing (HEAD → working tree)
309 inputs: 12 tests, 0 captured, 297 boundary probes
LLM_DRAFT → AUTO_REFUND 18 inputs
LLM_DRAFT → HUMAN_REVIEW 3 inputs
…
21 changed · 288 unchangedThe tests did not notice, but 18 requests that used to wait for a person now get their money back without anyone reading the message. Add --inputs with real requests and the count is real customers, before the change ships.