Skip to content

Scoring and history

Trend Finder combines current public evidence with deterministic signals and, when configured, validated AI analysis. Scores prioritize review; they do not certify a forecast or factual claim.

Each topic carries six 0-100 dimensions:

Factor Question
Relevance Does the topic match the current AI and creator scope?
Novelty Is the idea meaningfully early or differentiated?
Evidence strength Is there enough usable evidence to support inspection?
Source diversity Does the signal appear across distinct source roles?
Niche fit Does it match Creator Lens focus and audience?
Creator potential Can it become a useful piece, workflow, product, or experiment?

Named contribution rows, quality caps, corroboration, freshness, lifecycle, saturation, risk flags, and visibility bands make the final decision more inspectable. Source-local metrics never replace cross-source evidence.

A hidden gem is low-volume and high-novelty, not low-quality. It must clear minimum opportunity, novelty, and evidence gates, then receives a continuous gem score for ordering.

Action verdicts such as act now, monitor, review, or ignore summarize the current evidence and risk posture. The next-step copy must remain consistent with the verdict and visible confidence.

Lifecycle uses the bounded sequence unknown, whisper, builder, creator, and saturated. Movement compares the stable topic identity with prior local snapshots so renamed or reclustered topics can retain continuity when the match is strong enough.

Daily series, sparklines, convergence, velocity, risers, coolers, and Story Log entries describe observed run history. They are not guarantees of future attention.

Trend Finder can write bounded next-run predictions and later grade them when a target run arrives. Calibration summaries measure how those calls performed over time. Script-only historical backtests are a separate workflow and are not the same as Engine Replay or normal future-run retros.

Analyst output is parsed against a structured contract. Invalid output gets a bounded repair attempt and then falls back deterministically. The fallback can cluster, score, and produce creator copy without claiming an AI-analyzed run.

Always pair the score with the current provenance labels and the public evidence linked from the topic.