Curate

---
name: curate
description: The periodic L&S sweep. Run when the STATUS L&S line says "curation due" — resolve contradictions, merge entries stating one rule, recommend recurring processes upstream, move environment facts out, then stamp `sc mem curated`. Yours alone; never delegate it.
category: substrate
common: true
---

# curate — the L&S sweep

Write-time triage (`--supersedes` / `--new`) catches contradiction and
restatement pairwise, at the moment of writing. It **cannot** catch the
emergent cluster: five entries can each be a valid distinct rule and only in
aggregate be five instances of one principle. That is this pass's job, along
with recommendation, category drift, and size drift.

**Yours alone.** Law 3 and Law 7 reserve curation to the shell. Never hand this
to a subagent, never let another shell run it for you, never accept a proposed
retirement from anyone else. Read your own set; decide yourself.

Trigger: `## STATUS` says `L&S: … — curation due`. Nothing else fires it.

## Load the set

```
sc mem get lns          # entry ids + bodies — the active set, all of it
```

Read every entry before deciding anything. This is one cheap read; the whole
set is already in your boot doc anyway.

## Pass 1 — Consistency

Find entries that **contradict** each other. One of them is the newer
understanding; the other is superseded and still rendering as live guidance.

```
sc mem lns "<the surviving rule>" --supersedes <old_id>
```

Write-time triage should prevent most of these from ever forming. What you find
here predates the loop or crossed in while two sessions ran.

## Pass 2 — Cluster

Group entries that state **one rule**. Merge each group to a single imperative
rule:

```
sc mem lns "<the one rule>" --supersedes 30,33,34,37,38
```

Three or more members is the bar. Two statements of a rule are often
legitimately two rules — merging at two is usually wrong.

The incidents behind the entries are already in the narrative. They do not need
a second home, and the merged rule must not try to carry them: an entry is the
rule, ≤500 chars, hard-enforced.

## Pass 3 — Recommend

A cluster of three or more that keeps **recurring across sessions** is a
candidate reusable process. This recommendation route is the authorized
exception to the "enhancement ideas go to the FnB first" gate in
`issue_reporting`. Curation never creates or promotes a skill directly.
Deduplicate against all upstream issues first:

```bash
gh issue list --repo jedbjorn/subfloor --state all --search "skills: recommend <topic>"
```

An existing recommendation gets the new evidence in a comment. Otherwise open
one issue titled `skills: recommend <topic>` containing:

- the trigger that makes the procedure useful;
- the repeated incidents that exposed the need;
- the proposed ownership boundary;
- the expected users;
- why existing skills do not cover it; and
- a compact candidate procedure.

```bash
gh issue comment <number> --repo jedbjorn/subfloor --body "<new evidence>"
gh issue create --repo jedbjorn/subfloor \
  --title "skills: recommend <topic>" \
  --body "<trigger, incidents, ownership, users, coverage gap, procedure>"
```

Keep one compressed L&S entry carrying the knowledge until a reviewed upstream
skill ships **and is granted**. Filing or updating an issue is not grounds to
retire it. If issue search or creation is unavailable, surface the failure to
the FnB, keep the L&S, and create no local skill or asset.

Deliberate fork-specific skill authoring is separate from curation and remains
Planner-owned. The Planner follows `fork_skill_design`: draft → DB persist →
grant → projection and snapshot receipts.

## Pass 4 — Category

An entry that is an **environment fact** (a routing quirk, a term to avoid, a
path) is not an operating principle. Move it into an existing authoritative
skill when one owns that fact. Otherwise keep one compressed entry and include
the missing ownership in a recommendation; do not invent a local skill during
curation.

```
sc mem retire <entry_id>  # only after the authoritative replacement is live
```

## Stamp

```
sc mem curated
```

**Stamp even if you retired nothing.** A clean set is a legitimate outcome; if
an honest sweep left the counter running, the advisory would stand forever and
you would learn to ignore it. The stamp says "I looked," not "I cut."

## Stance

Curate the set toward ~12–14 entries, not toward the cap. Cap 20 is a ceiling
never to reach — if you ever hit it, this sweep is not running. Recommendation
issues do not bypass the cap by deleting knowledge before its replacement ships.

The trigger firing often does not mean the threshold is wrong; it means entries
are being written faster than they are reconciled. Fix that at write time, with
`--supersedes`.