A real port measured against its reference — Karpathy's nanochat versus the .NET implementation that chased it.
Has the .NET port reached the Python reference?
36 of 36 capabilities fully present · baseline nanochat (Python) → target NanoChat.NET · 36 compared
assessment v35 · 2026-08-02 · authored match
Audit update against fixed NanoChat .NET specification v2 at commit 43a4027: F35 is aligned through the named resumable speedrun and measured BPB/base-CORE sweep aggregation. All 36 source capabilities are now aligned; there are no Partial, Different, or missing source gaps, plus 2 intentional .NET additions.
The work queue: capabilities the baseline has and the target does not.
None.
Started, not finished — present but not equivalent.
None.
Not defects. Parity is directional, so what the target adds beyond the baseline shows up here rather than counting against anyone.
Audit update against fixed NanoChat .NET specification v2 at commit 43a4027: F35 is aligned through the named resumable speedrun and measured BPB/base-CORE sweep aggregation. All 36 source capabilities are now aligned; there are no Partial, Different, or missing source gaps, plus 2 intentional .NET additions.