sciforge

The unified scientific research toolkit — six modules, one coherent package that addresses the deepest pain points in reproducible, rigorous science.

pip install sciforge

---

Why sciforge?

Problem	Module
"My experiment isn't reproducible"	repro — scores your codebase, pinpoints the exact issues
"I can't track what I believed and why"	hypotest — version-controlled hypothesis ledger
"A unit mismatch crashed my analysis"	unitflow — physical units propagate through NumPy/Pandas
"My CV scores are inflated (same patient in train+test)"	crossvalbio — patient/batch/phylo/temporal-aware splits
"My figure is stale but I don't know which one"	papertrail — data→figure→claim lineage with staleness detection
"I can't see which papers my work builds on"	litmap — semantic citation dependency graphs
"I need a professional audit for submission"	audit — generates submission-ready Research Audit Trails (RAT)
"I need to share my work with reviewers"	bundle — packages everything into a reproducible ZIP archive

All modules share a common provenance store and work together seamlessly.

---

Quick start

import sciforge as sf

# 1 — Score reproducibility of a project
report = sf.repro.score("my_experiment/")
print(report.summary())
# Score: 78/100 (B — Good)
# Findings: 2 high, 1 medium
#   [HIGH] SEEDS  train.py:12 — stochastic ops but no seed set
#   [HIGH] HARDPATHS  preprocess.py:8 — /home/alice/data/train.csv

# 2 — Declare and test a hypothesis
h = sf.hypotest.declare("Drug X reduces IL-6 at 10mg/kg (in vivo)")
h.link_experiment("runs/exp_042")
h.test(p_value=0.028, effect_size=0.52, n=48,
       test_name="one-tailed t-test",
       notes="vs vehicle control")
print(h.status)   # SUPPORTED

# 3 — Physical units in pipelines
force = sf.unitflow.Quantity([9.8, 12.1], "N")
mass  = sf.unitflow.Quantity([1.0,  1.5], "kg")
accel = force / mass      # Quantity([9.8, 8.07], 'm / s²')

# Add unit contracts to functions
@sf.unitflow.requires_units(distance="m", time="s")
def speed(distance, time):
    return distance / time

# 4 — Biologically-aware cross-validation
from sciforge.crossvalbio import PatientSplit

for train, test in PatientSplit(n_splits=5, random_state=42).split(X, y, groups=patient_ids):
    model.fit(X[train], y[train])
    # Guaranteed: no patient in both train and test

# 5 — Data-to-claim lineage
with sf.papertrail.session("results/paper_v3") as trail:
    df    = trail.load("data/cleaned.csv")
    df2   = trail.transform(df, remove_outliers, "remove_outliers")
    fig   = trail.figure("fig2_roc", df2)
    trail.claim("AUC > 0.90 on held-out test set", fig)

# Later, after data changes:
report = sf.papertrail.check_stale("results/paper_v3")
print(report.summary())
# STALE: fig2_roc — data changed since last hash
# STALE: AUC > 0.90 on held-out test set — upstream data changed

# 6 — Literature dependency graph
graph = sf.litmap.build("papers.bib")
graph.show_dependencies("Attention Is All You Need", depth=2)
central = graph.most_central(n=5)
clusters = graph.cluster_by_keywords()

---

Module reference

sciforge.repro

report = repro.score("path/to/project/")    # or .py / .ipynb file
report.score           # int 0-100
report.findings        # list[Finding]
report.critical_findings()
report.summary()       # human-readable string
report.to_dict()       # JSON-serialisable

Checks: SEEDS, ENVLOCK, HARDPATHS, DATETIME, NOTEBOOK_ORDER, FLOATPREC, SECRETS, LEAKAGE

---

sciforge.audit

Generate a full Research Audit Trail (RAT) report for peer review:

from sciforge import audit

report = audit.generate_report()
print(report)

Or via the CLI:

sciforge audit

Includes reproducibility scores, hypothesis outcomes, data lineage, hardware snapshots, and carbon footprint estimation.

---

sciforge.hypotest

h = hypotest.declare("statement", alpha=0.05, tags=["bio"])
h.revise("new statement", reason="new data")
h.link_experiment("runs/exp_01")
h.test(p_value=0.02, effect_size=0.6, n=120, test_name="t-test")
h.retract("error found")
h.summary()

ledger = hypotest.ledger()
ledger.supported()        # list[Hypothesis]
ledger.pending()
ledger.by_tag("bio")

---

sciforge.unitflow

q = unitflow.Quantity(array, "N")
q + q2     # unit-checked addition
q / q2     # auto unit derivation (e.g. N/kg → m/s²)
q.to("kN") # unit conversion (requires pint)

uf = unitflow.UnitFrame(df, units={"col": "m"})
uf.derived("speed", lambda df: df["dist"] / df["time"], "m/s")
uf.to_quantity("speed")

@unitflow.requires_units(force="N", mass="kg")
def acceleration(force, mass): ...

---

sciforge.crossvalbio

Splitter	Use case
PatientSplit	Clinical / longitudinal data
BatchAwareSplit	Multi-batch experiments
PhyloSplit	Species / microbiome data
TemporalBioSplit	Time-series studies
ComboSplit	Multiple constraints combined

All are sklearn-compatible (implement split(X, y, groups)).

---

sciforge.papertrail

with papertrail.session("project/paper_v2") as trail:
    df    = trail.load("data.csv")
    df2   = trail.transform(df, fn, "step_name")
    fig   = trail.figure("fig1", df2)
    claim = trail.claim("Model X > baseline", fig)

report = papertrail.check_stale("project/paper_v2")
report.is_clean           # bool
report.stale              # list[TrailNode]
report.summary()

trail = papertrail.open_trail("project/paper_v2")
trail.lineage_of("fig1")  # all ancestor nodes
trail.claims()
trail.figures()

---

sciforge.litmap

graph = litmap.build("papers.bib")           # BibTeX file or dir
graph = litmap.build(["10.xxxx/doi"])        # DOI list (uses Crossref API)

graph.find("Attention")                      # title search
graph.dependencies("paper_id", depth=2)     # papers it cites
graph.dependents("paper_id")                # papers that cite it
graph.most_central(n=10)                    # by betweenness centrality
graph.roots()                               # uncited papers
graph.leaves()                              # papers not cited by others
graph.shortest_path("a", "b")              # citation path
graph.cluster_by_keywords()
graph.show_dependencies("Title fragment")

---

Requirements

• Python ≥ 3.10
• numpy, pandas, networkx, scipy, scikit-learn, rich, jinja2
• Optional: pint (for full unit conversion in unitflow)

---

Running tests

pip install -e ".[dev]"
pytest tests/ -v

---

Contributing

Pull requests welcome! See CONTRIBUTING.md.

---

License

MIT