Five NVIDIA Nemotron agent roles, through the Nebius Token Factory API, plan, investigate, read and explain. NVIDIA Parabricks, as a Nebius Serverless AI job on an H100, aligns the reads and calls the variants. Between them, Traceome's own code holds the three approval gates, checks every call and output, and writes every step into one Journey that the researcher and the agents both read. Code copies, the model decides, code checks.
The researcher brings a question, a paper or cited variants, and raw reads. Traceome's code runs the Run: it calls Nemotron for the judgements that change from question to question (which reference to plan for, which record a loose variant name means, how to explain the findings), Parabricks for the heavy compute, and public sources for identity. Nothing consequential happens without an approval, and every arrow is recorded.
Every agent output is a proposal, never a fact, until code verifies it and the researcher approves it.
pbrun germline 4.7.05begin).Four bounded outcomes per Work item: observed, not carried, no VCF record, unresolved. Coverage is measured per item R19; no function or causality is claimed R9.
29/29 findings equal across models; 2/2 planted bad IDs unresolved; 337/337 files re-verified R5 R4.
A GUI for researchers, so no technical skill is needed: it replays recorded Runs and starts and follows live ones. A CLI sits underneath it R18.
Three approvals, proposals never facts, research use only. The public replay needs no key.
The work splits in two. NVIDIA Parabricks does the alignment and variant calling; our own code reads the VCF, hashes every file and assembles the report. NVIDIA Nemotron makes the judgements that change from question to question.
Whatever a source returns is copied into the record exactly as the source wrote it, never generated or rewritten by the model. Each source's original answer is kept beside the typed result copied from it.
An agent chooses which tool to call and what a record means. When it names a value, it must cite the record or the quote it took it from.
Code checks every citation against the record. The researcher approves, and the Journey keeps all of it.
This is what lets human and non-human samples run through the same pipeline: one set of source tools, and no rules that hold only for one species.
Each role is NVIDIA Nemotron 3.5 Super VL through the Nebius Token Factory API and is offered only the tools it needs. The two roles that read results are offered no tools at all.
Inspects the inputs, settles the species and reference genome through NCBI, picks a Parabricks recipe and resources, and compiles one proposal.
inspect_input_metadataresolve_taxondiscover_assembliesinspect_referencelist_compatible_analysis_recipescompile_proposalsearch_journey→ one Parabricks proposal on GRCh38
Turns each loose identifier into exact Work items, one per alternate allele, or leaves it unresolved. It can ask the researcher a typed question.
fetch_refsnpfetch_eva_variantfetch_assembly_reportlocate_variantsearch_journeyrequest_investigation_information→ rs4149056: 2 Work items (T>A, T>C)
Reads the VCF window around each target, as observed by code, and raises concerns such as a nearby variant.
no tools · reads the observed table only→ nearby variant: 2 records within 100 bases
Explains the frozen outcomes in plain words. It may cite only facts in the approved basis.
no tools · approved basis only→ "T>C carried (0/1), T>A not carried [3b84b6f1…]"
Reads the paper the Run found and stored, and pulls out the variants it cites, each from a quoted span. Live: rs429358, rs7412 and rs1801133 from Mullins 2020 R12.
outline_papergrep_paperread_paperfilter_paper_rows→ variants and their build, each with its quote
We started on NVIDIA Nemotron 3 Ultra. It worked in simple cases, but as the steps grew in complexity it fell into loops: runaway replies, retries and gateway timeouts.
NVIDIA Nemotron 3.5 Super VL, at a lower price point, showed no drift across roles and clean tool calls, so one model runs the whole pipeline. Most of our loop guards are no longer needed.
Qwen3.5-397B-A17B on the same Token Factory API separated model issues from backend and pipeline issues. On the same inputs it reached the same findings R5.
Settings: tool_choice: auto, parallel tool calls, explicit max_completion_tokens of 245,760, recorded with every request.
A general agent keeps a chat. Traceome keeps the Journey file: append-only, immutable facts that both sides can query while the Run is happening.
journey.jsonlAppend-only. Every model request and reply, tool call, source retrieval, Attempt, approval and publication is one typed fact. Run H: 1,834 facts of 65 kinds, 30.7 MB.
artifacts/sha256/<digest>Every file is stored under the SHA-256 of its bytes and re-verified when the Journey is opened. Run H: 337 files, 0 mismatches.
Each interval's close binds its journal of external contacts by digest and size, so the record of who was contacted is itself checkable.
Each figure comes from one retained development Run and its post-Run review. Formal evals over repeated Runs are pending.
A coding agent may well find the variant. What the researcher can't see is which reference or allele it used, whether a missing record was read as "not carried", or how to fix one step without starting over.
| A general coding agent | Traceome | |
|---|---|---|
| Starts from | a VCF someone already made | raw FASTQ, then an approved Parabricks job on a Nebius H100 |
| What is kept | a chat transcript | a typed, append-only Journey R14 |
| You approve | a shell command or a file edit | the exact scientific object, bound by its digest R2 |
| A finding is | a sentence | one of four bounded outcomes, tied to its VCF record and sources R3 |
| "Not in the VCF" | easy to write up as "not carried" | always "no record" R9 |
| When unsure | it guesses, or asks in the chat | only that item pauses, with a typed question R17 |
| Fixing one step | re-prompt and run again | amend at the gate, or fork from a closed point R10 |
| Swapping the model | the answer is whatever that session wrote | the same 29/29 findings R5 |
| Who can check it | someone who reads code and logs | the researcher, statement by statement |
Replies were cut at the provider's default 8,192 tokens. We caught it when planning replies ended at exactly 8,192, held that Run, and made the reply limit explicit.
On Nemotron 3 Ultra, one reply held thousands of copies of a single tool call, and long requests timed out at the gateway. Streaming didn't help: with forced tool use every call arrived in the final chunk. We took control of the turn in code and moved to Super VL with the model free to choose its tools.
An earlier explainer contradicted its own correct findings. Now code fills in the claim words, each genotype is checked against the VCF row it cites, and it may cite only approved facts. It is still our weakest part.
For a public C. elegans sample our expected answer was wrong, not the agent. The sample's identity is unresolved, so we draw no biological conclusion from it R13.
The host clock stepped back about 104 s mid-Run. The Journey's ordering exposed it, so readiness now checks the clock first R7.
Runs that worked on a human sample struggled on non-human ones: different sources, names and formats, and rules that held only for human. One set of tools and "code copies, the model decides, code checks" fixed it.
We use one workflow, germline alignment and calling. RNA-seq alignment, somatic calling, fusion detection and DeepVariant-based calling would each add a kind of question.
Annotation and population frequency, structural variants, and a broad search of the whole sample instead of only named targets.
Several papers per question, and more literature sources.
Let a researcher add the tools their question needs, while the Journey still records every step.
Compare what several Runs found, across samples or questions, from their Journeys. Reports the researcher shapes.
Repeated Runs and held-out cases, scored by review of the recorded evidence: targets found, time, cost E1–E4.