- the nightly briefing behind
get_daily_report— what changed since yesterday, and what is worth looking at; - anomaly detection — this value is outside its own recent behaviour;
- the daily insight — a model-authored reading of the day, with a recommended action.
When a person asks a question, a wrong answer is a wrong answer. When the product volunteers a claim, a wrong claim is also an interruption the person did not choose — arriving with the authority of having been selected as worth saying.
It never invents a cause
A correlation observed in the data is not an explanation. Where no governed context event covers the window, the assertion says context missing and stops. A context event may name the metric it concerns. When it does, a claim about a different metric will not quote it — and a claim that could not compare says so, rather than treating “no metric named” as a match.Every claimed number carries its provenance, and an absent one is not zero
A figure in a proactive claim carries the pull that produced it and how fresh it is. A measure that was not collected is reported as absent — never rendered as0, which is a value somebody could act on.
A comparison names both of its terms. “Up 12%” without saying against what is
not a comparison.
Silence is a state, and it is disclosed
A detector that cannot yet fire is not a quiet detector — it is an absent one, and the surface says so. A z-score computed over a window that includes the point being tested cannot exceed(n−1)/√n. With the default threshold, that makes the detector
mathematically unable to fire below eleven observations, and unable to reach
the error severity below twenty-seven. The bound is derived from the
configured threshold, not hard-coded: move the threshold and the number moves
with it, on every surface that discloses it.
So a project with nine days of history is not told “no anomalies”. It is told
the detector cannot speak yet.
Confidence is derived, never self-declared
A confidence level chosen by the author of a claim measures nothing — and that applies to a model exactly as it applies to a person. An insight names its evidence, and the publication gate refuses a reference the server did not measure for this project over this window. Confidence is then computed from the rows carrying the members the claim actually cited, restricted to the insight’s own period — not the project’s rows at large:
The reading is the band of the limiting term — never a blend. Completeness
cannot buy back staleness. And
unmeasurable is a legitimate reading, not a
level: an insight whose evidence cannot be scoped says so instead of scoring
itself.
Model-authored prose carries one marker
Where a model wrote the sentence, the surface says so — once, not on every clause. Cited data carries no marker, because the data was not authored. A proactive claim stays a springboard, never a verdict: it points at something worth opening, and the opening is where the answer is defended.Publishing, retracting, and leaving the platform
Over MCP:
get_daily_insight_readiness says whether there is anything worth
publishing, preview_daily_insight composes it, publish_daily_insights
publishes it, and retract_daily_insight withdraws it.
Next Steps & Cross-References
Anomaly Detection
The z-score method, its baseline window, and the bound behind the disclosure.
Agentic Daily Insights
The nightly run that produces the claim.
Agent Rules
Rule 5: context events are candidates, never causes.
Renders, Dossiers and Sharing
The same two-person rule, applied to a figure leaving the platform.