Evidence-led roadmap
This is a list of open engineering directions, not promised delivery dates.
| Direction | Acceptance needed |
|---|---|
| Direct HD/raw wall-clock proof | A changed candidate or justified new method, quality held fixed, enough preregistered whole blocks against both references; retain failed trials. Do not rerun the same failed measured trial until it looks favorable. |
| Native candidate promotion | Fresh actual-prefix FATE, native/extended/GPU/fault/soak, source/installed CLI/MCP/consumer and install/update/rollback checks after performance passes. |
| Wider performance coverage | Independent source ranges, more real media, multiple Apple Silicon machines and fully disclosed cache/load/quality controls. |
| Cleaner public reproduction | A distributable small corpus with license/source hashes and a bounded standalone reproduction recipe; full archived lab inputs are currently external. |
| International diagnostics | Translate remaining runtime diagnostics while preserving stable error codes and the original Bengali reports. |
| Platform breadth | Design and verify resource/process ownership backends before claiming Linux, Windows or broader Mac compatibility. |
| Long-term reliability | More fault/storage/overnight evidence and dependency update acceptance, rather than copying older receipts. |
Public 0.6.1 focuses on safe source reuse, understandable setup and transparent evidence. Use the issue templates to propose a small reproducible improvement. See known limitations, benchmark decision and contributing.