Late night post but late night news.
I’ve almost completed the conversion of TETSUO to its MK-8 architecture with the new fixes in. Considering I was only considering what the new version architecture would be on May 1st, that’s a pretty good turnaround time. Tetsuo has now gone from the initial one big monolithic thing that can barely handle the load to a distributed batch processing system, mostly cron tasks, to a fully distributed service-based platform with an api gateway that self-maintains and has a status dashboard.
It’s currently running on my local, but, to test it out as a sanity check before deploying it I plugged its winners for tomorrow into the existing MK-VII winner slots for May 14th:
T1 SYMBOL magnitude accuracy sig_confidence sig_probability
2026-05-14 AGL 10.2773 53.95 90.00 99.08
2026-05-14 PRIM 8.5154 46.88 90.00 94.80
2026-05-14 CBRL 5.9356 47.07 90.00 97.70
2026-05-14 BWAY 5.1599 52.41 100.00 93.54
2026-05-14 SVCO 4.8939 61.57 90.00 97.41
2026-05-14 GOGO 4.8523 46.88 90.00 95.23
2026-05-14 REAX 4.0868 58.02 90.00 94.32
2026-05-14 SMHI 3.1875 70.11 90.00 99.78
2026-05-14 CENX 3.0522 55.29 90.00 96.11
2026-05-14 KLAC 2.7527 63.92 90.00 94.21
That will force the MK7 that’s already live to buy the MK8-selected algorithms at 11 tomorrow. Obviously we’ll be doing more retro testing later once its been in place for a while. I’m thinking of, once all the dust clears from the madness going on with my career, a service dedicated to spitting out accuracy stats so that piece can be largely workless too.
Another rush job unfortunately. I’ll finish the “BUY” component and polish it up until about the weekend and get it deployed unless I run into more issues that need resolved. The new version has explanations for each piece and is designed to be secure while still allowing users who aren’t logged in to see what it’s doing in a more human-friendly format.
Some sneak peaks:
And the status dashboard is a little more informative too:
It needs a cleanup pass to get everything ready for what may be end up being an unfortunate journey of bullshit, but, overall, I’m pleased and couldn’t have done it in that kind of timeline without LLM support.
MK-VII is unique in that every component is a standalone restful API service in addition to being the view layer of an MVC approach so commandline tooling could just as easily come in later if I find a purely UI-based approach doesn’t suit my needs.
I mapped out all the tasks I perform maintaining it and introduced a job/task queue/management subsystem in ORK that provides the orchestration between pieces I need so everything runs in the right order since the whole thing is incredibly stateful. If there were ALOT more compute power it wouldn’t have to be stateful but, this allows the whole thing to run on conventional hardware.
This version also has the scaffolding in place for new models and algorithms for forecasting to be introduced, automaintains symbol blacklists, etc, so, once it’s polished and burned in it should be self-maintaining — this eleviates the problem of always needing to reconfigure the thing to keep wins up so it should be more consistent in its accuracy in addition to being more accurate.
Fingers crossed. I need this to work given how much work has gone into getting it up at this point.