How ranking works

Every candidate item gets eight scores, each between 0 and 1, computed before you see it. Your own weights combine them into one score, and the same numbers explain the pick.

flowchart TB
    accTitle: From signals to an explanation
    accDescr: An item is scored on eight signals, each is multiplied by your learned weight, the products add up to the score, and the largest products become the explanation.

    item["Candidate item"] --> signals["8 signal scores"]
    signals --> weights["Times your weights"]
    weights --> score["Sum = ranking score"]
    weights --> why["Largest parts = the reason"]

The signals

Signal	What it measures
semantic	How close the item is to the main topics in your interest graph (importance-weighted best cosine match)
reading_depth	How far you're likely to read it, judged from other items of the same source and length. The item's own history is left out.
suggestion	Higher for items that came from discovery
explicit_feedback	Thumbs, saves and skips. A skip only counts if you viewed the digest.
source_trust	How your reading of this source has gone recently. New sources start neutral.
content_quality	Length, links, citations and code, adjusted to how much you finish at that length
temporal_context	Freshness over long, medium and short horizons, lower for the fifth story on the same event
novelty	Distance from what you already read

From signals to a score

• Contribution = your weight for a signal × the signal's value.
• The contributions add up to the item's score, which is then clamped to the range 0 to 1. Because of that clamp, the stored score can differ slightly from the sum.
• Missing evidence leaves a signal at a neutral 0.5. It doesn't appear in the explanation and doesn't push your weights either way.
• The explanation names the largest contributions, in plain words. It never claims a reason the numbers don't show.

How your weights change

Weights start from sensible defaults and are trained on pairs from the same digest: the item you opened against one you skipped. Comparing items shown side by side is fairer than counting raw clicks, because both had the same chance to be seen. Updates are batched per reader so one session doesn't swing the ranking.

More context: Architecture → How ranking learns and Feedback mapping.