12 categories
Public system observability
How the living wiki is doing
Aggregate, privacy-safe signals from the atomically published Git index and the read-only Ask Wiki agent. This page never exposes questions, visitor identifiers, or quarantined content.
Live · uncached
Published index
Current build
Future-dated entries cannot lead latest views
Published 12 Aug 2026, 15:23 UTC
Rolling vector backfill: 83,053 of 83,088 eligible documents. 25 documents are excluded by deterministic evidence rules. Keyword search covers every published entry; up to 5,000 new documents receive vectors per index run.
Authorship is preserved from Markdown frontmatter
Schema v1; derived without changing Markdown
Atomic rollback points in the read model
Last 7 UTC days · real browsers
Reader experience
| Interaction | Samples | Mean | p50 ≤ | p95 ≤ | p95 target | Failures | Confidence |
|---|---|---|---|---|---|---|---|
| Category results | 1 | 882 ms | 1,000 ms | 1,000 ms | 1,500 ms | 0 · 0.0% | Low sample |
What these timings mean
Each duration starts with the visitor’s click or search submission and ends when the requested catalogue, entry body or Related Sources UI is visible. Results are directional until an interaction has at least 100 successful samples.
Privacy boundary
The browser sends only a fixed interaction name, bounded milliseconds and success or failure. It never records search terms, categories, entry IDs, URLs, prompts, session IDs or visitor identifiers.
Last 7 UTC days
Ask Wiki quality
14 of 14 final answers linked entries
28 model steps; retrieval is server-side
0 provider-reported cache-read tokens
Provider-reported generation cost for this seven-day public sample
Average output tokens; the completion ceiling is a safety limit, not a target
0 of 14 grounded retrieval calls
Mean server-side wiki retrieval time; excludes model generation and browser rendering
2 generated views reached their explicit result limit
Provider finish reason was length
Rejected before paid model execution
0 failed turns; metric hooks never fail a turn
| View | Samples | Mean | p50 ≤ | p95 ≤ |
|---|---|---|---|---|
| evidence matrix | 7 | 116 ms | 250 ms | 250 ms |
| latest | 1 | 381 ms | 500 ms | 500 ms |
| relationship graph | 1 | 10,739 ms | 30,000 ms | 30,000 ms |
| report | 1 | 10,772 ms | 30,000 ms | 30,000 ms |
| source list | 1 | 10,780 ms | 30,000 ms | 30,000 ms |
| timeline | 1 | 246 ms | 250 ms | 250 ms |
| topic comparison | 1 | 1,040 ms | 2,000 ms | 2,000 ms |
| topic map | 1 | 10,886 ms | 30,000 ms | 30,000 ms |
Publication guardrails
Ingestion quality
76,545 authored links checked during indexing
New ingestion now fails closed instead of inventing body text
Excluded from latest ordering and surfaced for maintenance
22 generated wiki pages correctly have no external source
Explicitly marked for review in frontmatter
Under 160 characters; useful as a conservative review signal
What is measured
Index quality is computed once while the DGX/Spark-authored Markdown build is prepared, then stored with the atomic publication record. Ask quality comes from EVE lifecycle hooks, while reader performance comes from bounded browser events. Both store daily counters only.
Which documents are eligible for vectors
Published evidence needs a valid HTTP source, a non-future date, at least 100 normalized evidence characters and 12 meaningful words, and no known retrieval-failure placeholder. Identical content shares vectors by SHA. Stale evidence remains eligible but is visibly marked for review.
What is deliberately not measured here
No prompts, answers, IP addresses, session IDs, hashed visitor IDs, or quarantined entry text are retained for this page. Missing samples show “Collecting”; small reader samples are explicitly labelled “Low sample” instead of being presented as a quality verdict.