Stats

Live instrumentation for the human+AI moderation experiment. Cached for up to 5 minutes; each row names the research question (RQ1–RQ8) it feeds.

Articles by review state

Metric	Count	Feeds
Articles	778	RQ8
Never reviewed by the pipeline	57	RQ5
Review fresh (≤30d)	537	RQ5
Review stale (>30d)	184	RQ5
Drifted from pinned Wikipedia revision	310	RQ8
Parked (moved / deleted / needs human)	3	RQ4

Annotations by status

Metric	Count	Feeds
Formalized	10,511	RQ8
Partial	5,307	RQ8
Not formalized	21,927	RQ8
Articles awaiting count backfill	3	RQ8

By-provenance totals are deliberately not computed here (they would require parsing every annotation blob per request); the human-provenance signal is read from the events below.

Annotation events (event type × actor)

Event	Actor	Last 30d	All time	Feeds
`add`	`human`	0	1	RQ1 RQ5
`add`	`pipeline`	1,125	6,442	RQ1 RQ5
`delete`	`pipeline`	6	96	RQ1 RQ5
`modify`	`human`	0	7	RQ1 RQ5
`modify`	`pipeline`	910	6,226	RQ1 RQ5
All events		2,041	12,772	RQ1

Human signals

Signal	All time	Feeds
Endorsements (human agrees with AI)	0	RQ2
Human modifications	7	RQ1 RQ2
Rejections (human veto / tombstone)	0	RQ1 RQ4

Reader flags

Metric	Count	Feeds
Open flags	2	RQ4
Resolved: fixed	0	RQ4
Resolved: dismissed	0	RQ4

Revisions & patrol

Metric	Count	Feeds
Revisions: edit	32	RQ5
Revisions: pipeline	3,067	RQ5
Revisions: proposal-approved	6	RQ5
Revisions: proposal-decided-ai	6	RQ5
Revisions: seed	949	RQ5
Human edits awaiting patrol	32	RQ4
Human edits patrolled	0	RQ4

Pipeline runs

Kind	Runs	Articles	Errors	Tokens	Cost	Feeds
new	7	355	302	1,068,242	$71.06	RQ8
review	88	1,326	468	24,611,628	$1277.74	RQ8
wp-update	72	7,103	221	0	$0.00	RQ8
all runs	167	8,784	991	25,679,870	$1348.80	RQ8

AI proposals (propose-then-approve)

Metric	Count	Feeds
Pending (awaiting a human)	1	RQ2
Approved	12	RQ1 RQ2
Rejected	3	RQ1 RQ2
Decided by AI (auto-resolve)	6	RQ1 RQ2
Stale (target vanished undecided)	0	RQ5
Acceptance rate (approved ÷ decided)	80%	RQ2
Mean time to decision	19.0h	RQ4

Reading a zero: every instrument on this page has shipped, so a 0 in any row means that instrumentation is broken, not that nothing happened — investigate before celebrating. Median time-to-first-human-touch is omitted: it needs a per-annotation pipeline-event × human-event self-join that is not a cheap single SQL pass on D1. RQ3 (AI dissent / ladder blocks) is measured in the runner's decision sidecars, not in this database; RQ6 (inter-generation agreement) and RQ7 (confidence calibration) are deferred pending their sample/field (docs/research-plan.md).