A good question is
just the beginning.
Watch the real application, then follow the steps to design, inspect, and preserve your research.
Watch the live app walkthrough.
Open on YouTube ↗YouTube loads only after you choose to play. This actual browser recording shows a previously completed investigation; it does not launch a new research run.
Read the full text transcript →From checkout to workspace.
This static site works without an AI server. To run real investigations, use Python 3.12+, the repository, and an installed model on a reachable Ollama-compatible server.
Full installation instructions →git clone https://github.com/mso-docs/LazyLab.git
cd LazyLab
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e '.[dashboard]'
# Replace this with an installed model.
export OLLAMA_MODEL='your-installed-model'
slothlab init
slothlab doctor
slothlab dashboardOpen http://127.0.0.1:8501. Install and start Ollama separately; the app never downloads model weights automatically.
Design → inspect → confirm → preserve.
01 · Design the investigation
Use New investigation to specify the scientific question and permitted conditions. The registered experiment studies oracle prior-probability correction under synthetic class prevalence shift.
Choose controls, finite budgets, and reserved confirmation seeds. Define a competing prediction that can explain contrary results.
02 · Inspect the evidence
Research shows measurements, proposals, decisions, and provisional interpretations. Exact-result reuse saves an execution; a new independent seed tests variability.
Stop run pauses a worker. Resume investigation continues a nonterminal checkpoint. A terminal investigation is not a resumable pause.
03 · Keep confirmation independent
Reserve at least three independent seeds per confirmation group. Freeze the selected condition before using them; do not tune against confirmation outcomes.
Read uncertainty and negative results. A budget stop or a model's explanation alone does not establish a scientific claim.
04 · Preserve your work
Browser profiles remember separate local histories on the same browser and origin. Guest uses temporary JSON files and artifacts without SQL.
In Provenance & export, download a portable bundle. In Guest, use Open research bundle to restore it. Import starts no inference; review imported settings before resuming.
Questions you might have.
Can GitHub Pages run the AI agent?
No. This is a static evidence explorer and planning tool. Python experiments and AI requests run through the separately installed Streamlit app.
Can I connect a remote AI server?
Yes, through a reachable Ollama-compatible HTTP(S) base URL. Model discovery checks installed models. Research context goes to that server; experiments and files remain on the app host. Generic OpenAI/Anthropic APIs and authenticated gateways are not implemented.
Why does my investigation say insufficient evidence?
A planning limit, exhausted budget, mixed measurements, or missing independent confirmation can prevent a justified conclusion. Review the termination reason and recorded evidence. Completion is scoped to the finite protocol.
Why is my browser profile missing?
Local profiles depend on browser storage and the same origin. A different browser, port, hostname, or clearing storage loses the remembered locator. Clearing that locator does not delete server-side research files. Export bundles before switching environments.
Is this ready for a public multi-user server?
The full application is intended for a trusted single-user loopback installation. Optional identity hooks do not make it a reviewed multi-tenant service.
What does this website store or send?
The static site stores only your theme preference locally. Research-plan text stays in page memory unless you download it. No analytics or AI requests are made. YouTube receives a request only after you click a playback or chapter button. Downloaded notes and research bundles can contain your own research and should be reviewed before sharing.