ops
Live operational numbers — spend, calls, latency, and the day's allowance, read from the same journals the pipeline writes.
generated 2026-10-24 09:14 UTC
public fresh analyses today
0
resets midnight UTC
settled LLM spend today
$0.0000
backstop at $2.0000
reports published
101
all time, this store
LLM spend per report
$0.0271
all-time spend ÷ reports
fresh analyses are capped at 5 per visitor IP a day, inside one pooled allowance of 50 per UTC day
recent analyses (newest 20)
| started (UTC) | game | outcome | duration | views | cost |
|---|---|---|---|---|---|
| 2026-10-23 19:52 | God of War Ragnarök | done | 0m 53s | 3 | $0.0213 |
| 2026-10-23 19:23 | Portal 2 | failed | 0m 02s | — | $0.0000 |
| 2026-10-23 19:22 | Solasta II | failed | 0m 02s | — | $0.0000 |
| 2026-10-23 19:09 | God of War Ragnarök | failed | 0m 02s | — | $0.0000 |
| 2026-10-22 13:53 | Grand Theft Auto V Enhanced | done | 2m 21s | 5 | $0.0147 |
| 2026-10-22 13:47 | Dispatch | done | 1m 33s | 5 | $0.0340 |
| 2026-10-20 21:51 | Before We Leave | done | 1m 17s | 6 | $0.0785 |
| 2026-10-20 21:04 | Shadowrun: Dragonfall - Director's Cut | done (degraded) | 3m 52s | 5 | $0.0771 |
| 2026-10-20 20:55 | Dying Light 2 Stay Human: Reloaded Edition | done | 1m 41s | 4 | $0.0289 |
| 2026-10-20 20:26 | Wasteland 3 | done | 2m 08s | 5 | $0.0519 |
| 2026-10-19 14:02 | Rust | failed | 4m 54s | — | $0.0179 |
| 2026-10-16 00:57 | Undertale | done | 3m 33s | 4 | $0.0322 |
| 2026-10-16 00:54 | NBA 2K25 | done | 0m 55s | 14 | $0.0142 |
| 2026-10-16 00:51 | Football Manager 26 | done | 1m 33s | 14 | $0.0277 |
| 2026-10-14 23:44 | Dragon Age™ Inquisition | done | 2m 08s | 14 | $0.0402 |
| 2026-10-13 19:03 | Rise of the Tomb Raider™ | done | 3m 29s | 6 | $0.0233 |
| 2026-10-13 18:57 | Remnant: From the Ashes | done | 2m 26s | 6 | $0.0350 |
| 2026-10-13 18:05 | Max Payne 2: The Fall of Max Payne | failed | 4m 56s | — | $0.0162 |
| 2026-10-13 17:27 | METAL GEAR SOLID V: THE PHANTOM PAIN | done | 3m 34s | 3 | $0.0311 |
| 2026-10-13 15:53 | Backseat | done | 0m 21s | 4 | $0.0047 |
by day (last 14)
| day | admitted | refused | LLM calls | cache hit | cost |
|---|---|---|---|---|---|
| 2026-10-23 | 1 | 0 | 43 | 91% | $0.0213 |
| 2026-10-22 | 1 | 0 | 114 | 92% | $0.0487 |
| 2026-10-20 | 0 | 0 | 290 | 87% | $0.2365 |
| 2026-10-19 | 0 | 0 | 57 | 92% | $0.0179 |
| 2026-10-16 | 0 | 0 | 177 | 93% | $0.0740 |
| 2026-10-14 | 0 | 0 | 58 | 88% | $0.0402 |
| 2026-10-13 | 2 | 0 | 219 | 90% | $0.1215 |
| 2026-10-11 | 0 | 0 | 34 | 87% | $0.0214 |
LLM stages (all time)
| stage | model | calls | cache hit | cost | p50 | p95 |
|---|---|---|---|---|---|---|
| classify | deepseek-flash | 3,072 | 91% | $1.7835 | 2.6s | 5.2s |
| classify | deepseek-v4-flash | 3,002 | 92% | $0.9039 | 2.6s | 5.2s |
| compose | deepseek-flash | 56 | 18% | $0.0293 | 2.7s | 4.6s |
| compose | deepseek-v4-flash | 71 | 31% | $0.0236 | 2.7s | 4.6s |
about these numbers
- all days are UTC days; the operator's unlocked runs spend but are never counted as public admissions
- views counts report-page loads — search answers, library clicks, and direct links alike, each once at render; the count is identity-blind by construction (no IP or user-agent stored), so crawlers and the operator are included
- a job's cost joins the spend ledger by its run id, failed runs included; its views join the report-view journal the same way
- the recent-analyses table lists journaled jobs, and the job journal began 2026-08-09: 10 published reports predate it, so they count in reports published but list no row there, and $0.1610 of ledger spend joins no job row
- done (degraded) marks a published report whose narrative the grounding gate trimmed or withheld — the numbers and quotes are whole, the prose is partial or absent, and the report's own panel says which
- cache hit is the provider-side prefix-cache share of prompt tokens, computed over calls measured live at write (post-2026-08-09) — a group with none reads a dash, never 0%
- refusals count the spend breaker's and the search limiter's firings
- p50/p95 are dispatch-to-response wall clock per LLM call over the live-measured calls, retries included
- ledger rows from before 2026-08-09 were flat-priced at write and repriced 2026-08-10 from the archive's recorded cache splits; since 2026-09-24 each call is priced by its rate period (the provider's weekday peak hours at full rate, off-peak at half) and the rows since 2026-09-16 were repriced that way, while rows from 2026-08-16 to 2026-09-16 stay at the earlier flat table; the provider dashboard is billing truth
- token-level splits (prompt, cached, output, thinking) stay in the ledger journal — this page renders the rates and unit costs they produce