Exception review queue — Ho Hum (read-only view)

Human-in-the-loop desk for Better World. Agents draft; Ho Hum decides. This page is a read-only snapshot of open items. Live data: data/review/queue.json · API: GET /v1/review/queue (default ?status=open; use ?status=all for history).

How Ho Hum resolves (offline — no write API)

1. List open items: python3 scripts/list_review_queue.py or curl -s 'http://127.0.0.1:8787/v1/review/queue'.
2. Open the cited source (org JSON, claims index, trust flags).
3. Decide approved or rejected.
4. Edit data/review/queue.json by hand — set status, fill reviewer_notes.
5. If data must change, edit the cited data/orgs/*.json with citations only, then run python3 scripts/lint_data.py and python3 scripts/lint_honesty.py.

There is no POST /v1/review/... approve endpoint and must not be one without auth. Agents must never auto-clear trust flags to close a ticket.

Open items

1. exc_20261001_actra_donate — high / open

Reason

org_actra now has website + donate_url on file, but trust.website_reachable=unchecked, donate_url_status=none, http_checks=[], and flags still include WEBSITE_NOT_ON_FILE / HTTP_CHECK_PENDING / WEBSITE_UNVERIFIED. Human HTTP verification (or intentional mark) still needed before recommending a giving path. Do not invent CE.

Recommended human action

HTTP-verify (or intentionally mark) org_actra website + donate_url; update trust.website_reachable, donate_url_status, http_checks[]; clear WEBSITE_NOT_ON_FILE / HTTP_CHECK_PENDING / WEBSITE_UNVERIFIED only with evidence. Then set this queue item status=approved|rejected with reviewer_notes. Do not invent CE or URLs. Offline edit: data/orgs/org_actra.json + data/review/queue.json.

2. exc_20261001_actra_claim_shallow — medium / open

Reason

Contested presentation risk: claim eval_org_actra__0 (WELLBYs per $1,000 = 37, unit WELLBY_per_1000_USD, period HLI honourable) is verified=true from HLI's honourable-mention list but confidence=low and org summary calls it a shallow prediction. Human should confirm UI/API never presents this as HLI Top Charity CE, and that honourable_mention tier stays visible.

Recommended human action

Confirm recommend/compare and org pages never present claim eval_org_actra__0 (HLI honourable / low-confidence WELLBY) as Top Charity CE; keep honourable_mention / shallow labeling visible. If presentation is already honest, approve with notes; else reject any elevation and fix copy. Do not invent CE. Offline edit: data/review/queue.json (+ org/claim docs if needed).

Recently approved (context)

Not actionable — listed so agents do not re-open without new evidence.

• exc_20261001_path_ein (org_path) — Resolved by enrichment_pass 2026-10-01: PATH EIN 91-1157127 from official finances page; matched EO BMF → irs_status=exempt_501c3_eo_bmf. See data/qa/enrichment…
• exc_20261001_living_goods_ein (org_living_goods) — Resolved by enrichment_pass 2026-10-01: Living Goods EIN 20-5010527 from official donate page; matched EO BMF → irs_status=exempt_501c3_eo_bmf. See data/qa/enri…
• exc_20261001_gw_tcf_irs_claimed (org_givewell_top_charities_fund) — Resolved by enrichment_pass 2026-10-01: GiveWell funds (Top Charities / All Grants / Unrestricted) use Clear Fund EIN 20-8625442; matched EO BMF → irs_status=ex…

Agent rules

• Never auto-clear trust flags (WEBSITE_*, HTTP_CHECK_PENDING, EIN_UNKNOWN, IRS_TEOS_LOOKUP_PENDING, etc.) to make a recommendation look clean.
• Never invent CE, EINs, donate URLs, or IRS statuses while “helping” clear the queue.
• Leave status open until Ho Hum edits the JSON.
• Surface open exceptions when recommending the same org_id.

CLI: python3 scripts/list_review_queue.py · Demo panel: /demo/ section “Exception review”.