SciTeX

Every result should trace back to the data and code that produced it.

English · 日本語

A reviewer asks for the analysis behind Figure 3 of your 2023 paper. The figure is in the PDF. The script is in a directory called final_v2_fixed. The random seed is in nobody's memory — and when you run the script again, the number in the caption does not come back. Nothing in the paper is wrong; there is simply no path from the figure to the code and data that produced it.

That gap is what this organisation builds tooling against. Not a better notebook, not a better plotting library — a traceable path from raw data to a manuscript a reviewer can actually check.

SciTeX is one research project context with specialized interfaces around it:

                           SciTeX Clew
                    verifies provenance and drift
                                  │
  Scholar ── literature ──┐       │       ┌── data and files ── Storage
                          │       ▼       │
  Stats ───── analysis ───┼── SciTeX Hub ─┼── figures ───────── FigRecipe
                          │  one project  │
  Agents + Chat ─ commands┘    context    └── manuscript + PDF ─ Writer
                                  │
                             Custom Apps
                      extend the same project contract

The centre is not another file format. It is the shared project, identity, storage and command layer. Use the built-in apps from a phone or browser, work through SSH on a compute node, or add a custom app without making another copy of the research project.

These are four actual SciTeX Cloud development captures, not mockups. They show the Hub project surface, Scholar literature workflow, FigRecipe figure output, and Writer's source/PDF workspace. Source video hashes and capture details are in profile/assets/ASSET-MANIFEST.md.

Four crops of real SciTeX surfaces and outputs, in the order a project moves:

1. A project in the hub — data/, figures/ and references.bib sitting beside the manuscript. One project, not four tools with four copies.
2. Writer — LaTeX in the browser, and the compiler log ending in Preview compilation completed successfully. No export-and-reupload step.
3. Scholar — a populated library: real records, DOI and abstract in the detail pane.
4. The figures, as compiled into the manuscript PDF — Figure 1 of automated-research-demo, whose bar plots are the output of the synthetic sleep × age study (N=180) that one agent run carried from data to a manuscript in about 40 minutes.

All four are development builds, captured from a working dev instance with a throwaway demo account; every identifier — browser chrome, account and project names, avatar — is cropped out, and the recording's tutorial overlays are painted over. Sources, commits, dates, hashes and exactly what was redacted are in profile/assets/ASSET-MANIFEST.md. Rebuilding the composite is scripts/build_profile_proof.py. Nothing in the image is a mockup, and nothing in it is production. Watch the automated run · Read its outputs.

uv venv .venv && uv pip install scitex   # first run ≈ 45 s on a cold cache, ≈ 2 s warm

# quick.py — two groups, one test, one figure
import numpy as np
import scitex as stx

rng = np.random.default_rng(0)
good, poor = rng.normal(60, 8, 60), rng.normal(52, 8, 60)

r = stx.stats.test_ttest_ind(good, poor)
print(f"p = {r['pvalue']:.2e}  d = {r['effect_size']:.2f}  power = {r['power']:.2f}")

fig, ax = stx.plt.subplots()
ax.bar(["good", "poor"], [good.mean(), poor.mean()], yerr=[good.std(), poor.std()])
stx.io.save(fig, "quickstart.png")

$ .venv/bin/python quick.py
p = 1.17e-07  d = 1.03  power = 1.00
SUCC: Saved: ./quick_out/quickstart.{png,yaml} (Reproducibility Validation: PASSED)

The YAML is the part worth noticing. Saving a figure also writes its recipe — layout in millimetres, the matplotlib version, and the plotted values as CSV next to it — so the figure can be re-rendered and the numbers re-checked without the notebook that made it. scitex is on PyPI, requires-python >= 3.10; the transcript above is a real run on Python 3.12, seed 0.

Around 65 further packages — scitex-stats, scitex-io, scitex-hpc, scitex-clew, scitex-agent-container — each documented and licensed on its own, adoptable without the rest. Browse all repositories →

The Python packages above need neither: they run offline on a laptop. The hosted hub is a convenience, not a dependency — and it is the alpha component (its own README warns that data formats may change and to back up important work). If you need reliability today, self-host the library and treat the hub as a preview.

• Questions and ideas → Discussions on scitex-python. Open, and so far used once — the welcome thread. A second participant would make it a conversation.
• Bugs and feature requests → the issue tracker of the package concerned; a minimal reproduction gets answered much faster.
• Code → each repo ships its own CONTRIBUTING.md; PRs target develop, main is release-only. A one-time CLA is required before a first contribution can be merged.
• Fork PRs are approved by a maintainer before CI runs on them, and pull requests from forks are checked out into a branch for review rather than run on self-hosted infrastructure. That is deliberate; the reasoning is in this repo's README.

Measured 2026-09-17, because a profile page that only shows the good numbers is not worth reading:

Read honestly, that is a young, effectively single-maintainer project with a large surface area and no community yet. What it does have is unusually complete documentation per package and a pipeline that genuinely runs end to end — which is why the sensible way in is one package, not the organisation: take figrecipe if you fight with figures, scitex-scholar if you fight with references, and ignore the other 73 until one of them earns its place. Expect API churn, and treat the web hub as alpha.

_{info@scitex.ai · 日本語}