PUBLIC EVIDENCEGATED
∅No validated public benchmark result has been published yet.
LANGUAGE / RELATION / REASONING / EVIDENCE
Turkish-first
neural model research.
We treat Turkish not as a localization layer added later, but as a starting condition of model research. DEVİM is an open, testable research program spanning data, training, evaluation and computational dialectics.
RESEARCH-FIRSTEvidence over claims
TURKISH-FIRSTLanguage as a starting condition
OPEN RESEARCHProcess and results remain visible
LANGUAGE · TRAINING · EVALUATION
→Devim Model
Turkish-first neural model research.
We study generation, data distribution, training schedules and capability measurement together. Scaling is treated as a variable to test, not an explanation by itself.
RELATION · CONTRADICTION · REVISION
→Computational Dialectics
An independent research track seeking computational and testable dialectics.
We investigate contradiction, mediation, revision, historical formation and multiple perspectives through formal contracts and controlled experiments.
PUBLIC RESEARCH LEDGER
All research →Research & Experiments
From model behavior and evaluators to benchmarks and formal theory — we version what we test, why we test it, and what we learn.
update · devim-model
336M Gate-B: capacity effect not established; architecture-design gate opened
The matched 336M Gate-B evaluation is complete. Scale improved FORM and positive-control readouts but did not establish the preregistered capacity effect.
research · platform
Where is DEVİM now?
Verified current state of the DEVİM Model and Computational Dialectics research programs as of September 15, 2026.
research · devim-model
DEVİM research timeline
From the 43M reference and 110M diagnosis through FORM/CONTENT separation to the final 336M Gate-B decision.
experiment · devim-model
FORM Rescue: can response behavior improve at the same scale?
Without increasing parameter count, counterbalanced response-register supervision produced large improvements in EOS and repetition behavior.
experiment · devim-model
What did a failed QA experiment teach us?
Answering behavior improved partially, but held-out CE regressed and a later control showed that much of the task could be solved by trivial copying.
report · devim-model
Why did language learning at 110M fail to translate into measured capability?
As the ~110M plain Transformer moved from 500M to 1.2B supervised tokens, held-out language modeling improved strongly while the frozen capability measure barely moved.
MEASUREMENT DISCIPLINE
Benchmarks →Benchmarks & Evidence
Only validated results deliberately cleared for publication.
- Model / checkpoint
- Evaluator + dataset version
- Validated run
- Public gate
Results are linked to versions, artifacts and source history instead of being detached marketing numbers.
Inspect evidence →