Provenance & confidence¶
Every claim in PPX has a source.kind recording how the claim came to be,
and a confidence value recording how strongly the provider believes it
represents the subject.
Provenance kinds¶
| Kind | Meaning |
|---|---|
user_stated |
The user explicitly stated the preference. |
observed |
Derived from observed user behavior. |
imported |
Imported from another PPX provider or compatible source. |
inferred |
Produced by a model or rule applied to other data. |
derived_aggregate |
Derived from aggregate patterns. MUST NOT be re-exposed as user-stated. |
Rule: user-facing surfaces MUST distinguish user_stated from all other
kinds. Inferred claims MUST NOT be presented as facts.
Confidence¶
confidence is a value in [0.0, 1.0]. It expresses the provider's
confidence that the claim accurately represents the subject. It is not a
probability of truth in the world.
Rough bands (non-normative):
| Band | Meaning |
|---|---|
| 0.0–0.3 | Weak signal; show cautiously, flag for confirmation. |
| 0.3–0.6 | Moderate signal; surface with provenance and confidence visible. |
| 0.6–0.9 | Strong signal; still not fact. |
| 0.9–1.0 | Essentially confirmed — typically only for user_stated claims. |
Stability¶
Independent of confidence, a claim has a stability level (low, medium,
high) expressing how stable the preference is expected to be over time.
Evidence references¶
A claim MAY reference the evidence that informed it via source.evidence_refs.
Evidence payloads MAY remain private; only the reference and a privacy-safe
summary need be portable.
{
"evidence_id": "ppx:evidence:quiz:12",
"kind": "pairwise_choice",
"timestamp": "2026-03-18T12:10:22Z",
"visibility": "provider_only",
"summary": "User preferred scent A over scent B in a 'date night' context."
}