Local models by default
Ollama runs the reasoning on your own machine. No prompt, file, or graph is sent to a cloud model to answer a question.
Your research never
leaves your computer.
This is what sets it apart: reasoning runs on local models (Ollama) by default, we never use your data to train any model, and the whole record is files on your own disk. Want an external model? Your choice — never the default.
Research prototype · local by default; external models are opt-in, and your data is never used to train.
Local by default.
Yours, always.
Reasoning runs on local models by default, we never train on your data, and the whole record is files on your own disk. An external model is a choice you make — never the default.
Never trained on your data
Nothing leaves your devices to train any model, ever. Your research is not someone else's training set.
External models are opt-in
Want a hosted model for a specific task? Choose from many providers — only if and when you decide. Off by default, always your call.
The record is yours
Graph, ledger, notebook, and files live on your disk. Pull the cable and the lab still opens, reads, and reasons over what you have.
Pure C++ on the metal.
A WetHands bench beside it.
One half of this lab writes pure C++ with no runtime between the code and the machine — native binaries built, installed, and driven at the console, on modern targets and on decades-old systems nothing else will talk to. The other half runs WetHands laboratory setups: detached experiments, workspace files, a ledger that cannot be quietly rewritten. As far as we know, nobody else ships both — and nobody puts them on the same console.
Native code, no intermediary
Runtime-free C++ builds aimed at the machine itself — including reconstructed legacy systems the modern world abandoned. The console drives the compiler, the install, and the running binary.
Experiments that outlive the session
Background runs up to 24 hours with disk-backed state, workspace-contained files, and the hash-chained scientist ledger. The same gate that admits a compiler's output admits a variant's classification: evidence, or nothing.
The claim, plainly
Plenty of platforms orchestrate cloud APIs. This lab connects the instruments that were never supposed to connect — bare-metal native code, retired operating systems, nineteen public science services, and a WetHands bench — through one symbolic, source-pinned runtime. That interconnect is the moat.
The instruments
no catalog sells.
The public services answer questions. These four decide what the lab is allowed to believe, compute, and remember.
Lois, the gate
Typing lois at the console arms a fail-closed research gate: a research pass must ground in PEEL evidence or it does not pass. Typing peel reads the mode back — gate, lane, ledger — without ever changing it.
The scientist ledger
A hash-chained hypothesis ledger: questions pre-registered before results arrive, confidence computed mechanically, blame routed to the generator that earned it. An experiment cannot quietly rewrite its own history.
The compute bench
Python beside the data lanes — biopython, pandas, scipy, numpy, RDKit — for sequence work and cheminformatics on what the instruments fetched. Labeled honestly: process isolation, not a security sandbox.
run_pythonThe WetHands lane
Files and long jobs: detached background runs up to 24 hours with disk-backed state that survives restarts, and workspace-contained file operations for everything the lab writes down.
wetware_job_start · wetware_job_status · wetware_job_output · wetware_job_cancel · wetware_jobs · wetware_write_file · wetware_read_file · wetware_move · wetware_remove · list_workspaceType peel.
The lab answers.
$ kist › lois research gate armed — fail-closed › peel ◫ S C I E N C E peel gate: ACTIVE — research requires grounded PEEL evidence · "lois off" to release research lane: wired — "lois <question>" runs a PEEL-gated pass scientist ledger: 1 run(s), 31 chained event(s) under scientist/dogfood
Representative session. The SCIENCE banner and status lines are the verbatim output format of the shipped console's read-only peel status word; run and event counts vary by lab.
The papers behind
the practice.
Every design decision on this wall is written up. The architecture is public; the lab is the proof it runs. Research papers are preprints and have not been peer-reviewed.
Peel: the inversion
The symbolic store as the only author of facts — structure at ingestion, structural hallucination prevention.
Read ↗Retrieval Is Not Memory
Memory as governance over experience — nine decisions vector stores skip.
Read ↗Verified Before Acting
The pre-action adversarial loop — judging separated from committing, with five proved theorems.
Read ↗Traversing Data in Symbolic Systems
Typed-relation traversal as retrieval — the path is the justification.
Read ↗The Orchestration Gap
Why chain-level invariants need a symbolic layer, not trained guardrails.
Read ↗TinkyVision: Let There Be Sight
Light vision for air-gapped and bandwidth-starved benches.
Read ↗What researchers ask
before they visit the lab.
How can I search PubMed, UniProt, AlphaFold, and ClinVar from one tool?
The Perslis science runtime puts 44 bioinformatics instruments across 19 public services — literature search, protein structure lookup, genome and variant annotation, drug-discovery chemistry, and pathway analysis — on one console. Every answer is source-pinned to the database, accession, and URL it came from.
What is WetHands lab automation?
WetHands is the runtime's actuation lane: computational experiments run as managed background jobs for up to 24 hours with disk-backed state, monitoring, and cancellation. The same job contract is the integration surface for lab information systems (LIMS), electronic lab notebooks (ELN), and instrument automation APIs. Computational experiments run today; no autonomous laboratory is claimed.
Do these research tools work offline or air-gapped?
Live lanes query public interfaces when the network exists. Everything the lab keeps — topic packs, the hash-chained scientist ledger, the Peel knowledge graph — is local files that keep working with the cable pulled, so a lab record can stand on its own in a disconnected, air-gapped review.
How does the runtime prevent AI hallucinations in research?
Structurally, not statistically. A result that cannot name its source database and accession is refused by construction, clinical significance is reported verbatim and never re-graded, and the PEEL symbolic gate decides what enters the lab record. Unknown is a valid answer; an eloquent guess is not.
Which AI assistants can use these bioinformatics tools?
The runtime registers as a Model Context Protocol (MCP) research server, so any MCP-capable AI assistant or agent wakes with all 44 instruments present under the same fail-closed source contract.
How do I set up my lab with Perslis?
We set up pilot labs with research partners around a bounded scientific question and a clear definition of success — one research server on your machines, the same gate, ledger, and source contract in every session. Start by telling us the question.
Bring the question.
The bench is ready.
The runtime registers as a single research server on your machine, and every console session wakes with all 44 instruments present — the same gate, the same ledger, the same source contract. We set up pilot labs with research partners around a bounded scientific question and a clear definition of success.
Research prototype. Not intended for diagnosis, treatment decisions, or replacing experimental validation.
Continue exploringBioscience