Sparse runtime no longer has to scan every silent neuron.
Rust Implementation Surface
Rust is no longer the experiment. It is the implementation frontier.
The Rust lane is now a serious parallel implementation surface for INSTNCT, not just a faster forward-pass port.
v5.0.0-beta.1 achieves 24.6% peak next-character prediction (best seed; mean varies) with smooth cosine-bigram fitness and 1+9 jackpot selection — full Python parity.
What the lane already validated
Runtime shape is no longer the fragile part.
The Rust lane now matches the Python reference on language prediction. The next frontier is computational tasks and deeper architectures.
The readout surface is load-bearing, not a cosmetic add-on.
Zero-theta collapse confirmed the need for a real relay floor.
Single-window or single-seed stories are no longer trusted by default.
Implementation status
The library is already beyond forward-pass demo status.
- Owned topology, reusable state, CSR cache, and propagation live inside one coherent library surface.
- Genome save/load and checkpoint bundles allow resumable, adversarially testable longer runs.
- Learnable int8 projection and SDR table are already integrated into the live lane.
Frontier now
Python parity achieved. Now pushing beyond.
This is where the page should feel like an implementation frontier, not a diary.
24.6% peak (Python parity)
Smooth cosine-bigram fitness + 1+9 jackpot selection broke the old 17-18% ceiling. Matches the Python 24.4% reference.
Performance is not the blocker
Runtime decisions such as SoA layout, compact indices, and deterministic benchmarking are already directionally settled.
Addition learning
80% accuracy on 0-4 + 0-4 from empty network (83 edges). First proof of real computation, not just frequency matching.
Public interpretation rule
Edge efficiency matters, but it does not count as a mainline breakthrough on its own.
Go deeper
The website frames the lane. The wiki keeps the detailed validation record.
Use the full Rust Implementation Surface in the wiki for detailed checkpoints, archived experiments, chronology, and carry-over findings.