Evidence — why you should believe this

Every assertion Tangent makes that is not a plain count carries an evidence object: a root-cause finding, an Ask-in-English answer, and — as they arrive — predictions, release verdicts and agent findings. It is one shape, rendered by one panel, and exportable, so a screenshot cannot lose the part that says how sure to be.

The rule it enforces: a number without its evidence does not ship.

What it contains

Section	What it answers
Claim	The one sentence this evidence supports.
Confidence	high · medium · low · insufficient, with the reason in plain words — and the p-value, q-value or interval where one genuinely exists. insufficient is a first-class outcome: it means the data cannot support a claim, and no number is computed from it.
What was tested	The method, how many comparisons were made, and the multiple-comparison correction applied (scanning many segments guarantees false positives; the correction is why you can trust "significant").
What contributed	The drivers, as a signed share of the change, each with its own corrected q-value.
Data sufficiency	The minimum that applied, what was observed, and why it fell short if it did. Thresholds are shown on the surface they gate — never hidden in configuration.
What would change this	The settings, extra data or checks that could move the conclusion.
Sources	Exactly which events, properties and tables were read, over which date range and timezone, how many rows, how many statements. The Sources footer is always visible.
SQL	Every statement that ran, verbatim.
Caveats	The specific limits of this claim, not boilerplate.

Where it appears today

• POST /api/analytics/insights/explain → evidence (kind root_cause)
• POST /api/ai/ask → evidence (kind ai_answer)

Both are additive: the existing response shapes are unchanged, evidence is a new key.

Reading an AI answer's evidence

A generated query is never high confidence. Its basis reads, verbatim: "the query matched the workspace schema and ran; whether it answers the question as intended is for the reader to check against the query shown." That is the point of showing the SQL. When the query ran but returned no rows for the range, confidence is low and the basis says so.

Exporting

The panel copies a Markdown rendering (always works) and downloads the JSON (best effort in sandboxed viewers). Both carry the Sources footer and the caveats.

The contract

The normative definition is docs/EVIDENCE-SPEC.md; the machine-readable schema is docs/evidence.schema.json. Each producer has a contract test against it.