aims-okf-curator

name	aims-okf-curator
description	Maintain OKF indexes, logs, tags, cross-links, duplicate consolidation, and stale concept cleanup.
disable-model-invocation	false

AIMS OKF curation

Keep tags lowercase kebab-case. Ensure okf/index.md links durable concepts and okf/logs/log.md records meaningful curation changes. Run the adapter with --check to catch broken links and drift.

AIMS paths and commands

• Canonical OKF source: okf/
• Generated Hugo shadow content: content/knowledge/
• Daily report output: content/results/
• Numeric analysis artifacts: data/analysis/*.json
• Generate: uv run python tools/okf_hugo_adapter.py --src okf --dst content/knowledge --clean
• Check: uv run python tools/okf_hugo_adapter.py --src okf --dst content/knowledge --check
• Build: hugo --gc --minify

Guardrails

• Do not let LLM-authored OKF prose become the source of truth for scores, ranks, dates, prices, risk gates, or data availability.
• Do not hand-edit content/knowledge/; regenerate it from okf/.
• Keep custom code small and deterministic.

Qualitative theme curation (#96)

Daily qualitative artifacts (data/qualitative/*.json) are point-in-time snapshots; some of what they surface is durable and belongs in OKF. Run a curation pass monthly (or when a theme visibly persists) once artifacts have accumulated on main:

uv run .agents/skills/aims-okf-curator/scripts/curate_themes.py \
    --qualitative-dir data/qualitative \
    --concepts-dir okf/concepts

The script (delegating to src/aims/okf_curation.py) deterministically clusters recurring theme titles across artifacts, prints promotion candidates with dated artifact and evidence citations plus a ready-to-edit concept skeleton, lists supporting per-instrument stance streaks, and flags retirement candidates. Its output is a proposal for human review — it never writes to okf/ itself. With no accumulated artifacts it prints a recorded empty pass. Promoted theme concepts must carry the qualitative-theme tag and a theme_tokens front-matter list (the skeleton includes both) so later passes can assess recurrence and retirement.

• Promotion criteria: promote a theme only when it recurs in artifacts from at least 3 distinct analysis dates spanning at least 14 days — never a single-day observation. Judge recurrence from market.themes titles and per-instrument drivers.
• Concept content: describe the driver qualitatively; cite the dated qualitative artifacts (data/qualitative/YYYY-MM-DD.json) and the evidence IDs they rest on. Numeric facts stay pointers into data/analysis/, per the guardrail above — never asserted as truth in OKF prose.
• Review: curation output is a human-reviewed pull request only; never auto-merge OKF changes. Record each pass (including "no promotions") in okf/logs/log.md, then regenerate shadow content (--clean), verify with --check, and build with hugo --gc --minify.
• Retirement criteria: archive a theme concept when it has not recurred in any qualitative artifact for 60 days (curate_themes.py flags these from theme_tokens), when the #93 gates repeatedly withheld the entries it rests on, or when it is contradicted by newer artifacts. Retirement follows the stale-concept cleanup flow (reviewed PR plus a log entry); do not silently delete.

OKF primary references

• Google Cloud announcement: https://cloud.google.com/blog/products/data-analytics/how-the-open-knowledge-format-can-improve-data-sharing
• OKF v0.1 draft specification: https://github.com/GoogleCloudPlatform/knowledge-catalog/blob/main/okf/SPEC.md