Evidence and provenance
What is stored with every answer, so it can be trusted, checked and corrected.
AppThentic never stores a bare claim. Every fact about a rule carries its provenance: where it came from, who said it, and when.
What is kept with every rule
| Field | Example |
|---|---|
| Rule | Consumer refund window is 30 days |
| Locations | refunds/policy.py:42, support macro “Refund eligibility” |
| Reason | Payment partner settles chargebacks within 35 days; 30 keeps every refund inside that |
| Owner | Head of Finance |
| Source of the reason | Answer from the owner, 21 Jun 2026 |
| Supporting evidence | PR #1182 description, Jira FIN-208 |
| Status | Approved |
| History | Who changed what, and when |
Evidence is separate from the model
Facts are not stored in a model’s weights. They live in Organization Memory inside your infrastructure. That has three consequences:
- Every answer can cite its source. You can open the PR, the ticket or the owner’s answer it came from.
- A wrong answer can be corrected or deleted on its own, without retraining anything.
- Two disagreeing sources can be kept side by side, instead of being averaged into one confident sentence.
See om for how the interviewing model uses Organization Memory without holding facts itself.
Kinds of evidence
| Kind | Meaning |
|---|---|
| System | What the code, config or data actually does |
| Written | What a doc, ticket, PR or commit says |
| Answered | What a person told AppThentic, in their own words |
| Inferred | AppThentic’s reading, not yet confirmed by anyone |
First-hand and second-hand
What a person knows directly and what they were told are kept apart. If the person asked says “I don’t know. Priya told me this”, AppThentic records exactly that: this person does not know the original reason, and attributes the rule to Priya. It does not turn the attribution into a fact. It asks Priya.
Questions every answer can answer
- Who told us this, and when?
- What evidence supports it?
- Was it approved, and by whom?
- Is it first-hand or second-hand?
- Does the code still match the explanation?
An answer from the rule’s owner outranks a written source, and a written source outranks an inference. None of them overwrite each other.