Where is DEVİM now?

Status summary

This page is the verified research state as of September 15, 2026. DEVİM is moving on two separate tracks: Turkish-first neural-model research and the independent Computational Dialectics program.

DEVİM Model

The controlled scale experiment used a plain causal Transformer at approximately 336M parameters.

At Gate A, the larger model improved several readouts but did not clear the preregistered capacity threshold. Gate-B training later reached its frozen target of 36,755 optimizer steps and 1,200,012,349 supervised tokens.

Gate-B evaluation is now complete. In the matched comparison:

• breadth macro: 110M 0.3364583 → 336M 0.4333333
• breadth delta: +0.096875, below the preregistered +0.20 requirement
• minimum 336M breadth family: 0.15, below the preregistered 0.45 requirement
• FORM EOS: 0.8333333 → 0.96875
• FORM repetition: 0.1666667 → 0.0416667
• V1.8 macro: 0.44125 → 0.44125
• V1.8 competencies above chance: 4 → 4
• positive-control macro: 0.7142857 → 0.8571429

Formal decision: GATE_B_CAPACITY_EFFECT_NOT_ESTABLISHED_OPEN_ARCHITECTURE_DESIGN_GATE with capacity_effect_supported: false.

The correct interpretation is that 336M improved several readouts, especially FORM and positive controls, but did not establish the preregistered capacity effect. This result does not establish that Transformers are incapable of the target capability. It shows that scale alone is not the demonstrated explanation and opens the controlled architecture-design/comparison gate.

Computational Dialectics

This program remains independent from DEVİM model training. The current phase is D1-B Gate B — Preregistration / Learned Pilot Protocol.

Gate A v2 and the Gate-B v5 design are frozen. PyTorch implementation v3 passed strict CPU/static/runtime preflight 22/22, and a fresh independent implementation review returned PROCEED.

The remaining blocker is preregistration. Data splits and hashes, seeds, optimizer/LR schedule, update budget, checkpoint rule, metrics, decision margins, diagnostics, abort rules and compute ceiling must be frozen before any optimizer step is authorized.

What we still do not know

We have not established that Transformers are insufficient for the target capability, that an alternative architecture is superior, that DEVİM performs dialectical reasoning, or that architecture alone explains weak transfer.

Next gate

For the DEVİM Model, the next real gate is the architecture-design gate: instead of adding parameters by default, compare simpler and stronger alternative explanations under the same data, evaluator and compute discipline. For Computational Dialectics, the next gate remains preregistration freeze followed by independent review.