COMPASS immune-response remote tool#
TOU validation and deployment status (2026-08-16)#
The working CPU deployment ran the official all-cohort COMPASS checkpoint, converted to safetensors plus data-only preprocessing, on the official GIDE sample. In the pinned Torch 2.10.0 environment, the safe artifact and upstream checkpoint had zero difference in class probabilities and all 44 concept scores. This is execution-equivalence evidence, not a clinical-accuracy benchmark. Authenticated private Platform import and owner testing passed on 2026-08-16; public publication and independent-caller authorization/isolation remain untested.
Operation:
run_compass_predictionStart:
python -m tooluniverse.remote.immune_compass.compass_toolEndpoint:
http://127.0.0.1:7003/mcpProvider configuration: set
TOOLUNIVERSE_REMOTE_DATA_ROOT,COMPASS_SAFE_MODEL_DIR, andCOMPASS_DEVICE=cpu. The live service loads safetensors/JSON/data-only NPZ and never callstorch.load.TOU check:
tu doctor --forward http://127.0.0.1:7003/mcp --jsonPrivate relay:
tu serve --share --forward http://127.0.0.1:7003/mcp --name validation-immune-compass --workers 1
Non-loopback binding requires TOOLUNIVERSE_API_TOKEN; otherwise keep the server on loopback. The relay requires TOOLUNIVERSE_SERVICE_KEY. Public publication, independent-caller testing, and clinical accuracy remain unvalidated.
New-user check: python scripts/remote_validation/setup_skill_preflight.py --implementation immune-compass. Add --check-provider-env before launch, --live after launch, and --check-connect-prereqs before sharing. Live preflight checks exact MCP discovery only; it does not run or validate a model. The pinned relay SDK is not on PyPI and currently requires authorized GitHub repository access plus a configured SSH key, so a working local MCP server does not by itself prove that a new operator can share it.
The authenticated 2026-08-16 Platform matrix found all 30 private owner relays online and all 41 operations discoverable. This implementation was imported as unpublished owner draft(s), configured with a 120-second timeout, and invoked through /expert-sessions/{id}/test. Across the set, 38 unique operations passed return-schema and semantic validation; the three USPTO operations returned exact provider HTTP 403 and remain credential-blocked. Public publication, independent-caller authorization/isolation, broad saturation, and persistent supervision were not tested.
See the complete setup and verification guide.
run_compass_prediction performs real inference with the official COMPASS
architecture and a provider-approved model artifact. The public service never
uses torch.load: it reconstructs the reviewed architecture and reads weights
from safetensors plus data-only preprocessing arrays.
Provision a model#
The official checkpoint is a whole-object PyTorch pickle. Run the converter
only in an isolated administrator environment after reviewing the source and
pinning its digest. The official repository checkpoint used for validation is
example/model/finetuner_pft_all.pt at upstream revision
0e5c87665247e3a300f28282c8bbcc14e26973bd, with SHA-256
fc83d7b5eac3697bcd9d117acefa35b56cf4c4c3c5880de367c85a870aef4b0b.
python convert_checkpoint.py finetuner_pft_all.pt safe-compass \
--expected-sha256 fc83d7b5eac3697bcd9d117acefa35b56cf4c4c3c5880de367c85a870aef4b0b \
--upstream-revision 0e5c87665247e3a300f28282c8bbcc14e26973bd
The converter emits model.safetensors, preprocessing.npz, and
metadata.json, with cross-checked artifact digests. It is not imported by the
live service.
Deploy#
export TOOLUNIVERSE_REMOTE_DATA_ROOT=/srv/tooluniverse/data
export COMPASS_SAFE_MODEL_DIR=/srv/tooluniverse/compass-safe
export COMPASS_DEVICE=cpu
python -m tooluniverse.remote.immune_compass.compass_tool
Input is one sample in the official COMPASS TSV/CSV layout: Index, then
integer cancer_code (0 through 32), then TPM genes. The table must be inside
the provider data root and contain all 15,672 model genes. The server listens
on http://127.0.0.1:7003/mcp; a non-loopback bind requires
TOOLUNIVERSE_API_TOKEN.