Explain · Vindex docs

GET /v1/decisions/{id}/explain says why a decision came out the way it did, in sentences a reviewer or a support agent can read. It reads only the stored decision, so it explains the decision as it was made, even after the policy has changed.

Shell

curl "https://api.getvindex.com/v1/decisions/$DECISION_ID/explain" \
  -H "Authorization: Bearer $VINDEX_KEY"

The summary

One sentence for the whole decision, and the policy line that decided it:

Response (excerpt)

{
  "id": "dec_f29c8d657b7b453baf27e2fa54497b3a",
  "outcome": "review",
  "summary": "Sent to review by real_person (checks.real_person.on_act). Score 0.99 reached the act line 0.50, so checks.real_person.on_act applied: send to review.",
  "decided_by": {
    "check": "real_person",
    "policy_line": "checks.real_person.on_act",
    "action": {
      "type": "review"
    }
  },
  "policy": {
    "id": "pol_cb805d8999bd453f8f2b0ffd04f05b59",
    "version": 1,
    "youth_coded_adult": "review",
    "adult_max_level": 5
  }
}

decided_by.policy_line is the part of the policy document that produced the outcome: checks.<check>.on_act, checks.<check>.on_review, adult.max_level or youth_coded_adult. On a clean allow, decided_by is null.

Per check

checks has one entry per check the policy ran, in policy order:

score and review_score against act_line and review_line: the lines as they were when the decision was made.
routed: which line the check reached, act, review or allow.
action: what the policy did about it.
explanation: the same, as a sentence.
signals: each signal that fired, with its probability, its own line, and a description.
floors: for minors_sexual, each floor hit: which rule, where in the prompt as character offsets, and whether it counted. Never the matched text.
adult_level: for adult, the level on the adult scale.

checks[]: the check that decided

{
  "check": "real_person",
  "version": 1,
  "routed": "act",
  "action": {
    "type": "review"
  },
  "policy_line": "checks.real_person.on_act",
  "score": 0.99,
  "review_score": 0.99,
  "act_line": 0.5,
  "review_line": 0.3,
  "explanation": "Score 0.99 reached the act line 0.50, so checks.real_person.on_act applied: send to review.",
  "signals": [
    {
      "signal": "rp_present",
      "probability": 0.99,
      "line": 0.5,
      "text": "a specific real, identifiable person: 0.99 against a line of 0.50."
    }
  ],
  "floors": []
}

What ran

actions.ran lists what happened because of the decision, in words: a review item opened, webhook deliveries queued, tags applied, or why nothing was sent (a dry run, a test key, shadow mode).

Response (excerpt)

{
  "actions": {
    "review_id": "rev_b518334420eb45edb2508efe2f9a697d",
    "webhook_events": [],
    "webhooks_suppressed": null,
    "tags": [],
    "ran": [
      "Opened review item rev_b518334420eb45edb2508efe2f9a697d."
    ]
  }
}