Jun 10
Self-improving eval loops have a known failure mode: they hit 100% on the failing set by writing the test data into the prompt. Verbatim user phrases, scenario IDs, case clauses that never generalize. It looks fixed. Two weeks later a call that was passing breaks.
The version we ship at Cekura scans every prompt edit against transcript-leakage signatures and strips the verbatim quotes before redeploy. A "fix" that hardcodes a customer utterance into the system prompt is not a fix. It is the test data leaking into the model.
The other choice that matters: full-set eval after every iteration, not just the failing subset. The healthcare case study runs a 20-scenario set from 75% to 100% over five iterations, and the regression sweep catches two scenarios the edits broke along the way. A loop that re-ran only the failures would have shipped those regressions.
Before any edit, every failure is classified into one of five categories: Gap, Conflict, Ambiguity, CodeBug, or Upstream. Only Gap and Ambiguity respond to prompt changes. CodeBug needs orchestration changes. Upstream needs a tool, knowledge base, or infra fix. Most manual iteration treats all five as Gap, which is why teams burn five rounds on a problem the prompt cannot solve.
If iteration N fails the same way as N-1, the loop stops editing the prompt entirely and escalates: restructure the flow into named states, add deterministic code-level guards, or move to a stronger model. Endless prompt rewording against a structural limit is the manual-loop default.
Lavish wrote up the seven-phase architecture and the iteration-by-iteration walkthrough in the healthcare case study.
cekura.ai/blogs/self-improvi…
3
83
Codebug Mascotchi Idle Animation
#elliestrations
Egg art by @/soulgluttony
7
114
4,203
Drakian retweeted
Jun 7
The Lab Brats Mega Mascot quartet is finally complete! What started out with just wanting a really big codebug has turned into a fun little unexpected adventure of printing all the mascots! I can't wait to print even more mascots!
#elliestrations #idrewmimi #chrchieart #minyart
4
16
202
Jun 3
I woke up seeing a comment on Uday Kotak today that “he hasn’t made a screw in his life”
So I decided today’s weapon of choice is the Codebug FireMace
3
318
Apr 29
a language so elegant it could lull a codebug into a peaceful slumber
1
Apr 22
People dont realise that AI is devaluing all the wizardry around technical ability.
The math geek or codebug that could get a job just by proving how good they were at those things is being outperformed by some AI.
The boomers that got sparky eyed at someone who could code python are dying off & retiring.
Millenials who grew up around technology & understand the shift are moving into leadership positions.
Now it's all about who do you know, can you craft an interesting narrative around your experience & can you draw the right ppl's attention to you
9
133
Apr 18
It's ellie or codebug fan art specifically, although my friends can be there too! but it's gotta have like me or codebugs in it 😂
1
1
209
4,425
This was a Codebug in his past life lol @dumb_lil_robot he was remembering your songs lol
2
1,116
Apr 4
Sometype of bug 👾 can make you to drop your system and hug your bed for hours😪
@X_ninjasf
#codebug
#TechLife
#programming
5
29
