FUSE

Try it without installing: huggingface.co/spaces/ForgottenForge/fuse

Know your breaking point.

You have a parameter you turn and a result you measure. FUSE tells you where the system breaks, how confident it is, and how far you are from the edge — in one function call, with a publication-ready report.

from fusepoint import analyze

card = analyze(x, y, current_x=0.01)
print(card.score)   # 87
card.save("fuse_report.png")

What FUSE answers

• Where is the tipping point? Bootstrap confidence interval included.
• Is it real or noise? Permutation test against your own data, not an arbitrary threshold.
• How sharp is it? k = peak / mean (cliff vs gentle slope).
• How safe is the current operating point? Distance to the edge as a fraction of the parameter range.
• One number to act on. A 0-100 Stability Score combining the above.

All of this in a single StabilityResult you can print, .show(), or .save("report.png").

Install

pip install fusepoint              # core library
pip install "fusepoint[ui]"        # + Streamlit web UI (fuse-ui)

Core dependencies: numpy, scipy, pandas, matplotlib, sigma-c-framework (theorem-anchored layer).

Two layers

FUSE ships two layers over the same data:

Statistical layer (default)

Bootstrap + permutation + 0-100 stability score. Works on any tabular data — no physics assumed.

from fusepoint import analyze
result = analyze(df, x="load", y="latency", current_x=5000)
print(result.score, result.grade)        # 87 STABLE
print(result.critical_x, result.ci)      # 7245.0 (6890, 7580)

Theorem-anchored layer (deep=True)

Adds regime classification, gamma_O stability indicator, and theorem citations — each output traceable to a proof in the foundation paper.

result = analyze(df, x="load", y="latency", deep=True)
print(result.regime)        # "I_geom"  (single mode, paper Thm 8.5)
print(result.gamma_O)       # 24.3      (strict-SOC indicator)
print(result.citations)     # ["def:sigmac", "thm:trichotomy-geometric", ...]
print(result.paper_doi)     # "10.5281/zenodo.20548818"

The statistical layer scores how stable the system is. The theorem-anchored layer says which regime the system is in and which paper theorem licenses that reading.

What deep=True surfaces on top of regime + gamma_O

With sigma-c-framework >= 5.0, the theorem-anchored layer reports three additional pieces of provenance on the diagnosis string:

• Branch-(ii) standing caveat for regime I. A single visible mode in the chi-profile of a single probe does not exclude a slower mode that this probe suppressed below the visibility threshold. The framework's published two-probe verification protocol is the standard cross-check; FUSE surfaces this caveat by default whenever a regime-I reading is returned from a single probe.
• Operational-floor flag. When the spectral-attribution gate sits at the threshold, the regime call is reported as a soft reading rather than an asymptotic one.
• Spectral type (gapped / III-pp / III-anom) when the input is an autocorrelation profile rather than a parameter sweep. Parameter sweeps (FUSE's default mode) leave this field empty.

These appear in result.diagnosis and result.regime_detail only when deep=True; the default statistical layer is unchanged.

Quick start

Array mode

import numpy as np
from fusepoint import analyze

lr = np.linspace(1e-5, 0.1, 80)
loss = your_training_function(lr)

card = analyze(lr, loss, current_x=0.01,
               x_name="Learning Rate", y_name="Loss",
               label="Training Stability")
print(card.score)        # 87 -- you're safe
print(card.critical_x)   # 0.035 -- this is where it blows up
card.save("lr_report.png")

DataFrame mode

import pandas as pd
from fusepoint import analyze

df = pd.read_csv("server_metrics.csv")
card = analyze(df, x="concurrent_requests", y="response_time_ms",
               current_x=5000, label="Production Server")
card.save("server_report.png")

Column names become axis labels automatically.

Scan mode

from fusepoint import scan

results = scan("data.csv")                # auto-detect x, analyze every y column
results = scan(df, x="time", top_n=5)     # explicit x, top 5

for r in results:
    print(f"{r.y_name}: {r.score} ({r.grade})")
    r.save(f"{r.y_name}_report.png")

Accepts CSV, TSV, JSON (Plotly, Elasticsearch, pandas formats), Excel, Parquet.

Compare mode

from fusepoint import compare

delta = compare(x, y_before, x, y_after,
                current_x=0.2,
                label_before="Before Fix",
                label_after="After Fix")
print(delta.delta_score)  # +18 points
delta.save("improvement.png")

Web UI

pip install "fusepoint[ui]"
fuse-ui

Drag-and-drop CSV upload, demo datasets included, scans every numeric column and renders the cards side by side.

The Stability Score

The Stability Score (0-100) combines four independently validated statistical components:

Component	Weight	What it measures
Detection	40%	Is the tipping point real? (permutation p-value)
Clarity	20%	How sharp is it? (k = peak / mean ratio)
Precision	15%	How precisely located? (CI width / range)
Safety	25%	How far from the edge? (margin / range)

The score is self-calibrating: detection is measured against your own data's null distribution, not against arbitrary thresholds.

What FUSE is not

• Not a curve fitter (use scipy for that).
• Not an anomaly detector (use isolation forests).
• Not a time-series tool (use ruptures for changepoint detection).

FUSE finds parameter-space tipping points — the critical value of a knob where your system's behaviour qualitatively changes — and tells you how confident it is.

Version history

Version	What changed
1.1.0 (current)	deep=True layer aligned to sigma-c-framework v5: surfaces the Branch-(ii) standing caveat on regime I, operational-floor flag, and spectral-type axis. Streamlit web UI forces dark theme. Explanation block visible on landing.
1.0.x	Production release line. Streamlit UI bundled (fuse-ui). Theorem-anchored deep=True layer powered by sigma-c-framework v4. Dual licensure (AGPL + Commercial).
0.1.0	Initial PyPI release (renamed from fusekit). CLI / library only.

Citation

When deep=True results inform a publication, please cite the foundation paper:

Wurm, M. C. (2026). Operational scale selection: axioms, spectral concentration, and a regime trichotomy. Zenodo, doi:10.5281/zenodo.20548818.

License

Copyright (c) 2026 Forgotten Forge — forgottenforge.xyz

FUSE is dual-licensed: see LICENSE.

• AGPL-3.0-or-later for open-source and academic use (license_AGPL.txt).
• Commercial licence available for use cases where AGPL obligations are inappropriate (license_COMMERCIAL.txt). Contact nfo@forgottenforge.xyz.