research-loop

name	research-loop
description	Run one cycle of the Empathy Ledger wiki research loop. Lints the wiki, surfaces the highest-priority gap, drafts fill content, queues the result for elder review.

research-loop

Modeled on the ACT Knowledge Ops Loop. Six steps. One cycle per invocation.

When to invoke

• Manually via /wiki-loop slash command
• On a recurring schedule (weekly cadence is the default — overlong cycles produce stale graphs)
• After a new transcript, archive reply, or Trove batch lands and you want it threaded into the wiki

Inputs

• wiki/ directory at repo root
• Optional argument: a specific cluster slug, person slug, or research-thread slug to focus on

The six steps

1. Capture

Check wiki/raw/ for any new captured material since the last loop. If new files exist without wiki/sources/ summaries, generate them. Walk recent git log for thoughts/shared/ additions; offer to migrate any new transcripts or archive replies into raw/ + sources/.

2. Compile

For each new source, identify the people / clusters / country / events it references. If a referenced page exists, append a "this source mentions you" line to its Sources section. If a referenced page does not exist, add it to wiki/research/<cluster>-open-threads.md with a status of pending-compile.

3. Lint

Run node scripts/wiki-lint.mjs --write-report. Read the resulting wiki/output/lint-YYYY-MM-DD.md. Identify the top 5 highest-priority gaps using this priority order:

1. Broken wikilinks from primary cluster or person pages (these block reader navigation)
2. Stubs (< 400b) on referenced pages
3. Missing sources for cited claims
4. Orphans on people/cluster/country pages

4. Sync

Compare wiki/clusters/*.md chapter lists to the chapters arrays in src/cluster-configs.ts. Any divergence is a sync gap — wiki is the source of truth, cluster-configs is the published artifact. If there's drift, generate the diff and offer to write a migration patch (do not apply automatically).

5. Publish

The 10-years SPA reads from cluster-configs.ts + Supabase, not directly from wiki. Verify the dev server is running on port 4173 and the latest cluster-configs build matches what's in the wiki.

6. Learn

• Identify any recurring questions (questions asked across multiple transcripts or research threads).
• For each, check if a wiki/synthesis/<question>.md page exists. If not, draft one.
• For every gap fill drafted in this cycle, run node scripts/wiki-grade.mjs <page> --report to grade it against the v1 anti-fabrication rubric. The grader checks every cited surname, date, place, and Trove citation against the page's resolved sources. Critical issues block the elder-review queue.
• Update wiki/dashboards/research-loop-status.md with: cycle date, gaps closed this cycle, gaps remaining, archive-reply status, grader pass rate.

Output

Three things, every cycle:

1. Cycle log entry in wiki/dashboards/research-loop-status.md — date, what closed, what remains
2. Gap fills — pages drafted this cycle (in wiki/people/, wiki/sources/, etc.)
3. Next-cycle queue — top 3 priorities for the next loop, written to wiki/research/next-cycle.md

Stop criteria

• All broken wikilinks resolve, OR
• The next gap requires elder confirmation that hasn't happened yet, OR
• 10 gap fills drafted in one cycle (cap to prevent runaway sessions)

Fallback

If you can't make progress (e.g., archive replies haven't come back, elder review is pending), append a cycle skipped entry to the dashboard with the reason. Do not invent fill content. Anti-fabrication rule from SOUL.md: empty fields stay empty, "for elder review" tag for inferred claims.

Cultural protocol

• Never publish sacred-tagged content from synthesis/ or decisions/ without elder sign-off via Rachel Atkinson.
• Always cite source via wikilink to sources/.
• Three sacred items: Lucy at Blencoe Falls 1880s, Mulrunji Doomadgee 2004, Fantome Island leprosarium imagery.

Loop cadence

Cadence	Suitable for
On-demand	After a new transcript / archive reply / Trove batch
Weekly	Default rhythm
Monthly	If the wiki is stable and waiting on long archive turnaround