Borzoi (calico release) variant-scoring remote tool

TOU validation and deployment status (2026-08-16)

calico’s own borzoi_sad.py at commit 5c93582 on baskerville at 544073b, fold 0, reverse-complement ensemble, held resident on the CPU of an NVIDIA GB10 (aarch64, cc 12.1, driver 580.173.02), under kit mode off. Read five notes before deploying. (1) The equivalence gate is the provider: the worker holds the model but reimplements nothing numeric — it imports stock’s own initialize_output_h5, write_snp_len and write_pct out of borzoi_sad and calls baskerville’s own snp_seq1, untransform_preds and seqnn model in stock’s own order, then writes the same sad.h5. It passed: compared against a stock run.sh sad --mode off run dataset by dataset — the right test rather than a byte diff, since HDF5 allocation order and the kit’s chunked writer differ legitimately — all 16 datasets are identical, and the worker’s two requests agreed to the last float. The gate has teeth, which was shown rather than asserted: an earlier draft misread the pinned -u (declared dest='untransform_old') and called untransform_preds instead of untransform_preds1. That worker started, was deterministic across two requests, returned the pinned shapes and the right target_ids, and reported a plausible max|SAD| of 9.34 — while every one of its eight score datasets was wrong (SAD: 1809 of 6385 values, max |delta| 34.66) and the top-ranked track was a different track altogether. A provider can be deterministic, correctly shaped, correctly labelled and completely wrong in its numbers. (2) This kit’s exact mode is the first on this standard that actually runs here — it ships no prebuilt host objects at all and its levers are Python data-path rewrites, not CUDA kernels — and it is byte-identical to stock across three pairs of runs, all 16 datasets. It is still set to off, because no speed advantage could be established on this host: the within-mode spread (33 s) exceeded the between-mode difference (14.6 s) and the host’s load climbed monotonically through the measurement. provider.toml’s [kit] section has every number. (3) [gpu] policy = "none" is a measurement: TensorFlow publishes no GPU-capable aarch64 wheel at any version, so is_built_with_cuda() is False on a card torch drives in the same session. The upside is composition — no VRAM at all, ~10 GiB of host RSS, so variant scoring runs while a folding model owns the GPU. (4) The weights directory must be writable, unlike every other provider here: pysam.Fastafile builds hg38.ml.fa.fai beside the genome on first use. launch.py refuses by name rather than letting the worker die inside pysam. (5) This is not a replacement for run_borzoi_variant_effect at port 8012, and it is not even the same operation — see below. Public publication, cross-user isolation, concurrency, recovery and indels are unvalidated; keep this deployment private.

• Operation: run_borzoi_calico_sad

• Start: python -m tooluniverse.remote.borzoi_calico.borzoi_calico_tool

• Endpoint: http://127.0.0.1:8061/mcp

• Provider configuration: this provider is on the remote-tool standard, so the model runs in its own environment built from a pinned lock, and the weights are declared with a size and a sha256 and verified before the worker starts. Keep provider data, weights, caches and credentials outside Git. Use one worker: the resident worker serves one request at a time by design.

• TOU check: tu doctor --forward http://127.0.0.1:8061/mcp --json

• Private relay: tu serve --share --forward http://127.0.0.1:8061/mcp --name validation-borzoi-calico --workers 1

Non-loopback binding requires TOOLUNIVERSE_API_TOKEN; otherwise keep the server on loopback. The relay requires TOOLUNIVERSE_SERVICE_KEY. Public publication and independent-caller testing were not run.

New-user check: python scripts/remote_validation/setup_skill_preflight.py --implementation borzoi_calico. Add --check-provider-env before launch, --live after launch, and --check-connect-prereqs before sharing.

Not the same operation as the sibling at 8012

	run_borzoi_variant_effect (8012)	run_borzoi_calico_sad (this one)
implementation	borzoi-pytorch, a community port	calico’s borzoi_sad.py @ 5c93582
input	two DNA sequences	variants against pinned hg38
what it computes	ref-vs-alt delta at the central bin	upstream’s SAD, logSAD, D2, logD2
tracks	the port’s heads	6,385, strand-summed from 7,611
untransform	none	upstream’s targets table (-u)
ensemble	none	reverse complement (--rc)
device	GPU	CPU only here
cost	seconds	~255 s per variant

The right reason to reach for this one is that the Borzoi paper’s numbers are SAD over the strand-summed, untransformed, RC-averaged tracks, and that is what it returns. The right reason to reach for the other is throughput.

What a caller sends and gets

{"variants": [{"chrom": "chr1", "pos": 43110773, "ref": "G", "alt": "A"}],
 "top_tracks": 20}

SNVs only, at most four per request (≈17 minutes). Coordinates are GRCh38, 1-based, as in a VCF; a bare 1 is accepted and normalised to chr1 in both the tool and the worker, because baskerville.vcf.SNP prepends the prefix itself and two callers who disagreed about it would otherwise get rows that look like different variants.

The reply names the strongest tracks per variant by |SAD|, each with all four statistics and upstream’s own identifier and description, plus max_abs_SAD. The full 6,385-track vectors stay in the provider’s sad.h5, whose path the reply carries.

One thing the index cannot tell you

The kit pins tensorflow==2.15.1, and on aarch64 that wheel carries no code: a RECORD of six dist-info lines and a Requires-Dist: tensorflow-cpu-aws ==2.15.1 ; platform_machine == "aarch64" that hands the real library to another distribution of the same version. Installed with --no-deps the redirect is never followed, so pip list shows tensorflow 2.15.1 and import tensorflow raises. This provider’s lock keeps the pin and adds tensorflow-cpu-aws==2.15.1 beside it. Its TensorFlow 2.17.1 sibling, enformer_deepmind, needs no such line.

The authenticated 2026-08-16 Platform matrix found all 30 private owner relays online and all 41 operations discoverable. It imported each implementation as unpublished owner draft(s), configured a 120-second timeout, and invoked it through /expert-sessions/{id}/test. This implementation was added after that matrix and was not part of it, so none of that evidence covers it: it has not been exercised through the Platform at all, and everything above was measured locally, on one card, through the loopback endpoint. Public publication, independent-caller authorization/isolation, broad saturation and persistent supervision are unvalidated here.

See the complete setup and verification guide.

Authorize once, then share with one short command

Run once per machine, and again after key rotation:

tu remote login

Then each private share is:

tu remote share borzoi_calico

Override the defaults only when needed:

tu remote share borzoi_calico --name my-borzoi-calico-remote --workers 1

Sharing automatically starts the browser flow when the key is missing, expired, or revoked. A malformed or revoked explicit TOOLUNIVERSE_SERVICE_KEY fails fast instead of being silently replaced; unset or correct it, then run tu remote login. Non-interactive jobs also fail fast. Use tu remote logout to remove only the locally stored key.