The serving implementation is pinned upstream TensorFold. See ENGINE for its build, qualification limits and historical release provenance.
| Path | Purpose |
|---|---|
config/tensorfold.rev |
one reviewed upstream engine revision |
config/tensorfold-settings.txt, scripts/engine_settings.py |
the pinned engine's setting names; launcher forwarding and legacy refusal |
docker/zig/Dockerfile, scripts/build-native.sh |
native CUDA pair and runtime library checks |
scripts/run-zig.sh, scripts/serve.sh |
per-rank launch, identity, assets and operations |
config/prod-zig*.env.example |
pack-specific candidate profiles; local copies are ignored |
scripts/pack_engram.py |
raw-byte Engram weight-shard packing; NumPy, no inference engine |
scripts/check_native_assets.py |
native pack/image/profile identity and rank-asset checks |
scripts/check-public.sh, check_secrets.py, redact.py |
publication scanning and credential-safe output |
bench/, tests/ |
HTTP clients and recipe-tool checks |
docs/, results/ |
current instructions and explicitly historical measurements |
scripts/campaign/ |
historical analysis tools; no serving dependency |
The old Python inference family and staged server are retired. The older native export and its build-time tools remain in engine/ for the generator-equivalence audit. Independent clients, security checks, results and required asset generators remain.