Home About Research

Mardenic's research is small-scale by design. We build one system — Noema — from scratch, measure it against fixed evaluations, and report material results with their limits. Failed approaches remain part of the record instead of being edited out of the story.

97.7M Alma 2 parameters, trained from random initialization
600k base-pretraining steps complete; 19.66B token presentations
90/90 founder interview answers approved for constitutional design
0.1.0 founder-approved governing specification; curriculum and evaluation not started
Alma 2 base pretraining — 600,000 / 600,000 steps Complete
Protected model copy — strict weights-only reload Verified
Constitution founder interview — 90 / 90 answers Complete
Constitution contradiction audit — 6 / 6 rulings Complete
Governing specification 0.1.0 — four artifacts founder-approved Approved
Reviewed training curriculum and frozen behavior evaluation Not started
Alma 2 constitutional behavior and promotion Not evaluated
Alma 3 — proposed 307.1M-parameter scale-up Not training
001 July 2026 Ongoing

Project Northfall: building Noema from nothing

A from-scratch model program with original tokenizers and randomly initialized weights, attributed FineWeb-Edu pretraining text, and founder-led constitutional design. Alma 2 base pretraining is complete; constitutional behavior remains untrained and not evaluated for promotion; the founder-approved specification now gates the curriculum and evaluation work ahead. Read the program page.

002 July 2026 Controlled prototype

Noema-SearchBot: retrieval without pretending the model knows

A focused, host-mediated retrieval prototype combines a Mardenic-controlled index with scoped web search and fetch. A limited bridge smoke test passed; Alma 2 still cannot autonomously call tools, and retrieved material is treated as evidence rather than authority.