Devlog #5 — RUMI Dual-Track Architecture

so as yall know yesterday i finished adding the curiosity pipeline
(wait… you dont know about this? naaa💔go read yesterday’s devlog >:3…WUT ? GO RN)

but yea that was mostly just the foundation

today was about actually refining it and wiring it into the rest of RUMI

basically i ended up bringing a lot of the Track A discovery pipeline into Track B too because like… why wouldnt i 😭

not just straight up copy-pasting it though because then both tracks would end up doing the same thing

instead i reused parts of the architecture while keeping the curiosity systems separate

the biggest addition was making the generated curiosity constraints actually flow into the discovery process itself

which sounds simple until you spend half the day finding out they werent actually being propagated through the pipeline properly 💔

and then came the debugging arc 😭

DNS resolution failures causing discovery runs to randomly die 😾
(like wdym rumi requested for more papers and data and dns forgetting that internet exists)

curiosity constraints not propagating into the pipeline😫
Semantic Scholar 429 rate limits 🥀
Gemini 503 errors 😢
free-tier rate limits everywhere (she makes like 6-7k requests in one run )😺🥀

and somehow RUMI has become so LLM hungry that i had to run her on 13 API keys across 3 different providers just to keep discovery sessions alive 😭 (lwk running on hope and sum duct tape 🥀🥀)

there was genuinely a point where i was about to crash out because providers kept rate limiting, APIs were dying, runs were hanging, and everything was fighting me at once 💔🥀😭

but after all that RUMI finally completed a full dual-track discovery run

topic:

“Why does the universe expand faster than expected?”

results:

Track A Winner
→ Scale-Dependent Effective Gravity

Track B Winner
→ Temporal Vacuum Shear Framework

and this is where things got interesting

Track A and Track B finished with:

0 shared theories <–(🙀)
different winners
different hidden variables
different contradiction sets
different mechanisms
0 runtime errors

RUMI generated:

17 theories
30 mechanisms
22 hidden variables
12 contradictions
12 predictions

from a single discovery session 😭

and honestly the most important number isnt the score

its this:

Shared Theories: 0

because it means Track B isnt just making a slightly remixed version of Track A anymore

both tracks started from the same question, same literature, same knowledge base, and still ended up exploring completely different hypothesis spaces

which is pretty much the entire reason i built the curiosity system in the first place 😭🥀

still more work to do, but today feels like one of the biggest architecture milestones for RUMI so far💔🥀
lwk gonna try to improve GFlowNets tmrw

here’s the updated pipeline diagram :-