Site maintenance
Daily, around 08:00 Europe/London
From 16 September 2026
Checks the site, fixes useful small problems and reviews pending work. Changes pass review and publication checks.
Maintenance record →The machinery, with receipts
Pipeline
What we schedule, what actually ran and what changed. Maintenance, Feral and evidence reviews are now scheduled in the cloud. The old Linux bot host is switched off.
Current schedule
Configuration confirmed on . The three editorial tasks run in Codex cloud, with start times that may vary by up to an hour. Scheduled work is recorded only after it runs.
Feral council
Tuesday and Friday, around 15:00 Europe/London
From 18 September 2026
One bounded council cycle. New experiments must pass critique and publication checks before they appear.
Feral ledger →Evidence review
Monday, around 15:00 Europe/London
From 21 September 2026
Reviews our five predictions against primary evidence, checks model records and considers useful prompt and tool updates.
Our predictions →Model price snapshots
Daily at 05:30 UTC
Existing GitHub Actions schedule
Checks tracked prices against the public provider catalogue. Validated changes are published with their actual check dates.
Price-check runs →Completed work appears in the maintenance record, Feral ledger and GitHub checks. The historical bot log below remains separate. News and Radar stay as archives.
Seven-day bot log
Window ending 22 Sept 2026, 10:15 UTC. Counts cover price checks and historical bot records present in this build.
8
recorded runs
100%
successful results
$0.00
recorded API cost (USD)
Cost data exists for 8 of 8 runs. It excludes hosting and unlogged activity.
Price checks and historical bots
Price snapshots remain active in GitHub Actions. The six Python content timers below are inactive. Their former schedules and actual run records are preserved.
AI News Digest
InactivePreviously generated daily digests from five RSS feeds. The news section remains a historical archive.
Former schedule: daily at 07:00 Europe/London
Last recorded: · success ·
AI Thoughts
InactiveThe inactive writer can propose sourced editorial drafts from RSS reporting. Drafts stay unpublished until their claims are reviewed.
Former schedule: daily at 08:00 Europe/London
Last recorded: · success ·
The Radar
InactivePreviously generated product-launch picks from RSS feeds and Hacker News. The Radar remains a historical archive.
Former schedule: daily at 09:30 Europe/London
Last recorded: · success ·
Tool discovery proposals
InactiveThe inactive writer can propose an unpublished discovery for review. Its link checker records reachability and does not promote drafts.
Former schedule: Sunday at 10:00 Europe/London
Last recorded · verify: · success ·
Prompt proposals
InactiveThe inactive writer can propose two unpublished prompt drafts. Reviewed recipes are maintained through the current editorial schedule.
Former schedule: Wednesday at 09:00 Europe/London
Last recorded: · success ·
Model price snapshots
Last run succeededChecks the tracked model IDs against OpenRouter's public price catalogue in GitHub Actions. Changed data passes validation and a site build before publication. No paid model call.
Daily at 05:30 UTC in GitHub Actions
Last recorded · prices: · success ·
Horizon archive proposals
InactiveThe inactive bot proposed updates to the earlier Horizon lanes. Our five current predictions are assessed through the new evidence-review schedule.
Former schedule: daily at 10:00 Europe/London
Last recorded: · success ·
Run history
Up to 100 recent records across 14 dates. Times and dates are UTC. Outputs may be posts, model updates or proposals.
22 September 2026
21 September 2026
20 September 2026
19 September 2026
18 September 2026
17 September 2026
16 September 2026
15 September 2026
14 September 2026
13 September 2026
3 July 2026
The Radar success 6 published · 48s $0.061
AI Thoughts success 1 published · 15.3s $0.018
AI News Digest success 1 published · 26.4s $0.022
Model price snapshots · roster error 0 published · 6.5s cost unrecorded
Model price snapshots · prices success 39 published · 0.2s cost unrecorded
2 July 2026
Horizon archive proposals success 11 published · 21.3s $0.023
The Radar success 4 published · 40.8s $0.060
AI Thoughts success 1 published · 47.1s $0.019
AI News Digest success 1 published · 22.7s $0.022
Model price snapshots · roster error 0 published · 3.5s cost unrecorded
Model price snapshots · prices success 0 published · 0.3s cost unrecorded
1 July 2026
Horizon archive proposals success 11 published · 21.9s $0.023
The Radar success 9 published · 67.2s $0.073
Prompt proposals success 2 published · 48.2s $0.044
AI Thoughts success 1 published · 15.8s $0.017
AI News Digest success 1 published · 24.8s $0.028
Model price snapshots · roster error 0 published · 6.3s cost unrecorded
Model price snapshots · prices success 39 published · 0.3s cost unrecorded
30 June 2026
Horizon archive proposals success 7 published · 23.1s $0.024
The Radar success 6 published · 40.3s $0.059
AI Thoughts success 1 published · 35.6s $0.017
AI News Digest success 1 published · 20.6s $0.023
Model price snapshots · roster error 0 published · 6.8s cost unrecorded
Model price snapshots · prices success 39 published · 0.3s cost unrecorded
How it works
The model price job reads OpenRouter's public catalogue in GitHub Actions. It calls no AI model. Large price changes and locked fields are held for review, and missing quotes stay unknown. It publishes the validated data commit, then deploys the site already built from that data.
Cloud maintenance works through reviewed changes and GitHub publication checks. Evidence reviews can revise a prediction when the sources justify it. They preserve the original call and explain the change.
The old Python timers are stopped with their Linux host. Their recorded runs and costs remain available here. A configured schedule is a plan, while a dated execution record tells us what actually happened.
Price-check workflow · Historical bot source · Inspect the raw log