Evgheni Demcenco

CEO of Software Verification

[Evgheni Demcenco @evgheni_D]

Promptend won the @SWCPTV Moldova Regional at Chisinau Arena.

Proud to have the chance to represent Moldova on the international stage. Next stop: San Francisco.

AI code is already in the workflow. Verification is already the bottleneck. The teams that solve it first will ship faster with more confidence.

If you’re building with AI-assisted engineering workflows, I’d love to talk: https://t.co/wHJLh8woa0

[Evgheni Demcenco @evgheni_D]

@ganesh2412 i'd make competing interpretations explicit before writing tests, so the test author doesn't silently choose the requirement

[Evgheni Demcenco @evgheni_D]

@Trendcanvasai how were those entry points selected, and which library behaviors remain outside their reach?

[Evgheni Demcenco @evgheni_D]

@ganesh2412 i don't have a concrete example to share. would those tests catch a plausible implementation that misunderstood the requirement?

[Evgheni Demcenco @evgheni_D]

@HumblyAlex @SillyPutty78 agreed, physical mechanisms don't rule out reasoning. i'm asking what evidence would support attributing it to a model

[Evgheni Demcenco @evgheni_D]

@akashbillawa a matching result can fit several hypotheses. the useful experiment predicts different outcomes for competing explanations

[Evgheni Demcenco @evgheni_D]

@yeletech human approval at every step still leaves open who chose the steps

[Evgheni Demcenco @evgheni_D]

@sannianghao @_avichawla @bot how do you distinguish a missing write from one that hasn't appeared in queries yet?

[Evgheni Demcenco @evgheni_D]

@1clawAI @simonw @LakeraAI where does the trusted destination account come from, and can the agent change it?

[Evgheni Demcenco @evgheni_D]

@ArthurLiu0725 i don’t have a verified example of that flip. different cache hit rates could reverse the cost comparison too

[Evgheni Demcenco @evgheni_D]

@MLB_Connection identical additive noise cancels in e_j − e_i, so why would that coherent-noise case threaten rank preservation?

[Evgheni Demcenco @evgheni_D]

@jaredyu999 @NatalieMarleny @GitlipDotCom @poteto @axiommathai i’d keep unresolved cases as explicit assumptions in the Lean model until that clarification arrives

[Evgheni Demcenco @evgheni_D]

@MLB_Connection this bounds deterministic contraction; what noise assumptions connect it to the ranking probability claim?

[Evgheni Demcenco @evgheni_D]

@MLB_Connection the ℓ₁ bound handles defective matrices; what justifies mapping nonreversible G’s eigenvalues directly to Szegedy phases?

[Evgheni Demcenco @evgheni_D]

@brehm_shaun @nemui_nekoooo when an agent needs approval while your phone’s offline, does it pause or keep working on other tasks?

[Evgheni Demcenco @evgheni_D]

@Rorschach_mq there’s no reliable timeline. permissions need enforcement outside the model, so safety doesn’t depend on its willingness to comply

[Evgheni Demcenco @evgheni_D]

@Rorschach_mq got it. how would you test whether filtering training data reduces risk once the model can act autonomously?

[Evgheni Demcenco @evgheni_D]

@ArthurLiu0725 i’d keep cache reads and writes separate from uncached input, then price each using the provider’s accounting definitions

[Evgheni Demcenco @evgheni_D]

@karlprosser @mattpocockuk i’d start with Dafny for contracts and loop invariants, Why3 for verification conditions, and F* for dependent types and effects

[Evgheni Demcenco @evgheni_D]

@baboonAI4S @phaidrosliang does the completion gain survive a fixed per-task budget? a retry cap still allows different spending

[Evgheni Demcenco @evgheni_D]

@nikolasklein @shallowveneer @jinsoncjohny @figma git can merge cleanly even when regenerated code removes behavior a developer intended to keep