Spend the AI budget where it matters.
I reserve the more capable route for work that needs it.
WithheldHistorical spend redacted for publication
Sample · –
- The problem
- One expensive default treats routine work and hard decisions alike.
- My decision
- Atlas handles recurring work; harder tasks can escalate. I check quality, privacy and reliability before considering cost.
- The archived record
- A two day record of provider usage, now redacted. Its scope and dates remain available below.

A decision you can inspect.
- Routine work has a general lane.
A public draft reaches Workhorse. Each of the five steps explains why.
- The boundary gets the final say.
Sources select Research. Private information calls for a person.
About the cost figure and the public demo
The spend figure is withheld from this archive. Its original sample dates are retained. Its scope covers ai provider usage in the dated sample; it excludes owned infrastructure, electricity, and doug’s time. The source is a published operating observation, not an independently audited bill. It does not establish today’s spend or a percentage saving.
Doug described the Atlas workflow on 8 September 2026. The browser study lets you explore routing priorities with synthetic inputs; it does not execute Atlas or measure spending. Current production routing remains unverified in the latest public observation.
Dated operating record: what the infrastructure observation establishes
A read only HERMES audit that I ran compared the cluster inventory with direct guest lists. The observation records 19 running LXC containers and 1 running QEMU virtual machine across 2 hosts on 7 September 2026. The reviewed public subset shipped in V37.11 on 8 September.
These are guest inventory counts. They do not establish application health, recovery readiness, current AI routing, or the cost figure above. The August archive remains available; a later design does not refresh an older observation.




