Engine room

live · 30s

The serving infrastructure behind this site’s assistant, deliberately overengineered and left visible. Every number below is real and current; every answer links to its own trace.

All lanes nominal: the canary is green, the queue is idle, 302 answers served so far.

Request pipeline · all time

since 2026-09-05

  1. Requests

    670

    arrived at the edge

  2. Admission

    304

    rate + token budgets

    ↳ 320 shed

  3. Cache

    889

    answered without a model

    ↳ 31 dedup reattached

  4. Guardrail

    10

    scoped refusals

    ↳ 97 failed open

  5. Generate

    302

    durable workflow, ladder armed

    ↳ 62 degraded to raw KB

  6. Verify

    302

    citations resolve

Time to first token

p50 593ms · p95 8.20s · 304 samples

target 2.5s0ms–409ms: 83 answers409ms–819ms: 125 answers819ms–1.2s: 8 answers1.2s–1.6s: 3 answers1.6s–2.0s: 6 answers2.5s–2.9s: 8 answers2.9s–3.3s: 37 answers3.3s–3.7s: 10 answers3.7s–4.1s: 1 answer4.1s–4.5s: 1 answer5.3s–5.7s: 2 answers5.7s–6.1s: 2 answers6.1s–6.5s: 1 answer7.0s–7.4s: 1 answer8.2s–8.6s: 1 answer8.6s–9.0s: 4 answers9.0s–9.4s: 4 answers9.4s–9.8s: 1 answer9.8s–10.2s: 1 answer10.2s–10.6s: 1 answer12.3s–12.7s: 1 answer12.7s–13.1s: 1 answer13.1s–13.5s: 1 answer13.9s–14.3s: 1 answerp50 593msp95 8.2s014.7s

Goodput

100.0% · 302 ok / 0 failed

Full answer

p50 3.97s · p95 16.25s

End-to-end through guardrail, retrieval, generation, and verification, measured inside the workflow.

Ledger · all time

302 answers served for $0.05 in model spend ($0.00016 per answer), 341,691 tokens through the gateway, 920 requests absorbed by the cache and dedup layers before spending anything.

Recent answers

timeoutcomettfttotalcited
09-11 08:40glm-5.3-flash8.20s15.98s✓trace →
09-11 08:40glm-5.3-flash9.18s16.59s✓trace →
09-11 08:34glm-5.3-flash10.25s21.91s✓trace →
09-09 18:41glm-5.3-flash12.43s15.35s✓trace →
09-07 05:00glm-5.3-flash13.47s23.10s✓trace →
09-06 17:52glm-5.3-flash8.88s18.95s✓trace →
09-06 16:48refused: off topic7.19s7.51s✓trace →
09-06 16:48glm-5.3-flash2.93s3.36s✓trace →
09-06 16:48glm-5.3-flash2.81s3.46s✓trace →
09-06 16:48glm-5.3-flash2.88s5.23s✓trace →

Register

primary model
zai/glm-5.3-flash
model override
none
prompt hash
45b79cdd7ce5
policy hash
6bb17bbe0908
breaker · primary
closed
breaker · fallback
closed
eval gate
27/32 · 2026-09-06
canary
up · 2.4s · 00:29 UTC
daily token ceiling
0 / 2,000,000 (0.0%)
kill switch
off

Counters accumulate for the site’s lifetime; latency distributions cover the last 500 answers; traces are kept for seven days and never contain visitor message content. The architecture behind this page is the subject of the writing.