Coverage and performance
Release-scope coverage
100% of the 0.1 NR release scope is covered: 774 of 774 audited OAI function rows. Of these, 757 are directly bound and 17 use reviewed safe treatments where exposing the raw OAI function would be unsafe or misleading.
The broader upstream inventory also contains internal, platform-owned, deployment-specific, unsafe-lifetime, and later-release functions. Those rows remain recorded for engineering traceability, but they are not silently counted as missing from the deliberately scoped 0.1 product.
Three support tiers
| Tier | Public exports | What it means |
|---|---|---|
| Supported workflows | 42 | Named link, stack, scheduler-policy, and independent-cell owners exercised by release examples |
| Supported granular/shared capabilities | 476 | Manifest-backed operations and the value/type closure required to call them |
| Reference-only inventory | 1,189 | Importable pinned-OAI values and codecs outside the first-release workflow contract |
| Default NR wheel | 1,707 | 272 public leaf modules in 10 root namespaces |
The first two tiers are the supported 0.1 contract. Reference-only imports can still be useful for inspection and focused codecs, but they do not imply an end-to-end workflow, a standards-conformance claim, or deployment support.
Supported workflow boundaries
| Boundary | What remains OAI code | Deliberate limit |
|---|---|---|
| Granular PHY | Selected coding, modulation, estimation, reception, decoding, synchronization, or transform operation | One operation is not a complete radio procedure |
NrPhyLink |
Reviewed SISO PDSCH/PUSCH transmit and receive chains | Link level; no MAC or cell acquisition |
NrPhyStack |
OAI MAC grants plus supported real waveform and receiver paths | Preconfigured one UE; not a softmodem |
UeGnbLink |
Persistent OAI UE/gNB MAC and copied FAPI exchange | PHY outcomes are abstracted |
| Scheduler policy | OAI state, validation, HARQ, and grant construction around a copied decision | Policy seam, not replacement of the whole scheduler |
| Independent cells | One isolated owner set per cell | No shared RF, interference, handover, or coordination |
Performance
These are warm steady-state results from the release benchmark host. The table shows the median (p50) and p95; each row compares an identical workload across direct OAI C and the public binding call. The repository retains p99 and every raw sample for engineering analysis without crowding the release summary.
| Workload and timing boundary | Original pinned OAI | Python bindings |
|---|---|---|
| Connected gNB MAC: one loaded scheduler slot | 5.08 us p50 / 6.26 us p95 | 235.21 us p50 / 249.47 us p95 |
| FR1 SISO PUSCH: estimation through demodulation for one frequency-domain slot | 185.81 us p50 / 194.54 us p95 | 985.10 us p50 / 1032.57 us p95 |
The binding measurements include the safety/isolation boundary observed by a Python caller. Owners remain persistent across all timed slots; a new worker is not started per slot. The gNB MAC row is below 1 ms through p95. The PUSCH row is 985.10 us at p50 but exceeds 1 ms at p95, so it does not support a blanket sub-millisecond claim.
Rows perform different telecom work and must not be compared with each other. The release evidence uses warm-ups, multiple trials, stored samples, fixed geometry, semantic result digests, CPU-affinity records, exact source/wheel hashes, and explicit call boundaries. The repository retains the full raw benchmark evidence for audit without publishing it in the user website.