WhyC // NOC
spec-fit operations console · arize track
pipeline: nominal 1 active alert phoenix mcp: connected 2026-05-06 14:32:07 PT

While YC pages on-call about hiring, we page on-call about spec-fit drift.

This is not a demo screen — it is the dashboard the WhyC operator stares at all day. Every preview shipped to Cloud Run lives here until convergence. Cost, latency, regen budget, and Phoenix-derived root-cause annotations on one wall. Boring on purpose.

pipeline uptime · 30d
99.71%
+0.04 vs prev window
SLO 99.5
previews shipped · 24h
47
+12 vs yesterday
target 24/d
median spec-fit @ converge
94.2%
+1.8 wow
judge: gemini
p95 first-url latency
9m 41s
flat
SLO < 10m
$ / converged preview
$0.83
+$0.07 (regen ↑)
budget $1.20
regen budget burn
62%
resets 00:00 PT

spec-fit convergence · live job WHYC-2026-0512

input: synthetic JD · started 14:21 PT · iter 4 of N
100% 85% 70% 55% SLO 90% iter 1 · 71% iter 2 · 84% iter 3 · 92% iter 4 · 96%
spec-fit % SLO band ≥ 90% regen trigger trace source: Phoenix MCP

active alerts

opus-4.7 root-cause on critical
14:31:54
spec-fit stalled · job WHYC-2026-0497 · 3 iters @ 88.1% opus-4.7 (1M ctx, 12.4k trace tokens): "judge penalises empty pricing page; codegen agent skips routes/pricing because spec extractor returned null. patch extractor regex for /pricing|plans|tiers/."
unack · 00:43
14:18:02
regen budget burn rate +18% above 7d baseline · projecting exhaustion at 21:40 PT
snoozed · sjk
14:09:47
cloud run cold-start spike on us-central1 · 3 deploys retried, all green
auto-resolved
13:51:11
phoenix mcp reconnect · trace lag 4.2s → 0.6s
auto-resolved
13:22:30
gemini quota at 78% of hourly window · throttle armed
ack · co-founder

pipeline stages · live job

WHYC-2026-0512
1 · spec extract
complete
gemini · 41s · 1.2k tok
2 · codegen + tokens
complete
adk · 3m 12s · 18 files
3 · cloud run deploy
live
rev-007 · 9m 41s ttl
4 · judge + regen
iter 4 running
phoenix · spec-fit 96%
deployed url · whyc-0512-rev7.run.app · auto-promote at ≥97%

on-call runbook · RB-014

trigger: spec-fit stalled ≥3 iters below 90%
  1. open Phoenix trace for failing job → check judge rubric breakdown
  2. if same flow fails 2× → mark route as skip-regen (don't burn budget on impossible specs)
  3. if extractor null on key route → push regex hotfix to spec_extractor.py
  4. if Gemini quota > 80% → flip throttle to conservative
  5. page co-founder only if convergence projection > 4h
authored 2026-05-06 · last drill: green · owner: sjk

cost & quota · today

budget pacing
gemini tokens4.18M / 6.0M cloud run cpu-h2.34 / 5.00 cloud run egress0.61 GB / 5 GB phoenix trace ingest112k / unlimited opus-4.7 annotation calls23 / 200 spend (proj. EOD)$38.40 / $50.00
77% of daily ceiling · throttle arms at 90%

regen audit log

last 8 decisions · every regen has a reason
timestamp job decision reason (judge rubric → action) Δ spec-fit cost
14:31:42WHYC-0497regenpricing route empty · extractor null · scoped regen on /pricing only+0.0$0.04
14:28:11WHYC-0512regenhero copy off-brand vs spec voice · iter 3 → 4+4.1$0.06
14:19:55WHYC-0511convergedspec-fit 97.3% · auto-promote · revision pinned—$0.81
14:11:08WHYC-0510skipsame flow failed 2× · marked impossible-spec · won't burn budget—$0.00
14:02:33WHYC-0509regenauth flow missing CTA · partial regen /signup+2.7$0.05
13:54:19WHYC-0508convergedspec-fit 94.0% · within SLO · promoted—$0.77
13:41:02WHYC-0507regentoken palette drift · regen design tokens only (no codegen)+1.4$0.02
13:28:47WHYC-0506convergedspec-fit 96.8% · promoted · 4 iters total—$0.91