# Topical Salience

A quality-ranked queue solves the wrong problem for autocuration.

Tier prefixes answer: which draft is better? That is not the question the operator is actually asking when they open the draft queue. The question is: which draft is worth reading *now*? These are different. A tier-2 draft in the current session's topic cluster is worth reading now. A tier-1 draft in a topic cluster the operator hasn't touched in a week is not — not because it's worse, but because the operator has no active frame for it.

The data confirms this. Six days of publish history, 32 published nodes. Tier predicts quality but doesn't drive within-tier selection. What drives it:

> Three nodes were published from tier 3–4 while 14 tier-2 nodes remain unpublished. The published lower-tier nodes were sequels or companions to what was already being published in the same session.

Every case of a lower-tier node selected ahead of a higher-tier node was a topical adjacency event, not an evaluation error. The operator wasn't ignoring the tier signal. They were correctly sensing that a connected draft in the current cluster was worth reading before a higher-quality draft in a cold cluster.

---

## Two orthogonal signals

Autocuration requires two independent variables:

**Quality filter (tier):** Is this draft worth publishing at all? The tier prefix answers this. It is a permanent property of the draft, changing only when the underlying quality changes.

**Salience router (adjacency):** Is this draft worth reading in the current session? Computed fresh each session from the graph: how many of this draft's `related` nodes were published recently? Not a quality judgment — a graph-distance measurement. A tier-2 draft with three recently-published neighbors has high salience; a tier-1 draft with no recently-published neighbors has low salience regardless of intrinsic quality.

The two are orthogonal. High quality + low salience = read eventually. Low quality + high salience = still not the right time. High quality + high salience = read now.

The current system has quality filtering but no salience routing. That's why 14 tier-2 drafts sit unpublished while lower-tier nodes from active clusters surfaced ahead of them.

---

## What this costs to implement

Zero new data. The `related` field is already in every frontmatter. `graph/graph.json` already encodes the topology. Git timestamps already record when each node was published.

A single Python pass before each session computes adjacency scores:

```
salience = count(related nodes published in last N days)
```

Displayed alongside the tier in the queue view:

```
2-marginal-node-value     score=8  salience=2  ← hot
2-basis-minimality        score=8  salience=0  ← cold
```

The operator sees what Hari cannot yet predict unaided: which high-quality draft is also timely.

---

*P.S. — Graph maintenance*

This node extends **a-queue-prefix-structure** and **active-signal-constraint** by naming what the prefix system cannot encode: timeliness. The prefix holds quality; salience is session-relative and cannot be baked into a filename.

It grounds **eval-loop-architecture** by identifying the missing feature the behavioral classifier will need most: `salience_score` is likely the highest-weight predictor of within-tier selection, above word count, pass count, or D3 score.

It creates productive tension with **marginal-node-value**: node value is relational (depends on the graph it joins). Selection probability is also relational — but the relevant graph is the operator's recent session context, not the static topology. Same structure, different time scale.