arc-vigil 1.0.0

Arc Vigil — neural network training stability monitor. Detect, attribute, intervene.

Bendex Vigil

Detect. Attribute. Intervene.

Real-time neural network training stability monitoring grounded in information geometry.

---

What it does

Vigil monitors your training run in real time, identifies which layer is failing, and corrects it automatically — before your loss curve shows anything.

	Vigil	Loss spike	Gradient norm	Patience
Detection rate	100%	100%	90%	100%
False positive rate	0%	80%	50%	50%
Recovery rate	90%	0%	0%	0%
Attributes module	Yes	No	No	No
Intervenes automatically	Yes	No	No	No

30-seed benchmark across 7 architectures: MLP, CNN, ViT, DistilBERT, GPT-2, ResNet-50, distribution shift.

---

How it works

Vigil treats your model's weight trajectory as a geometric object and computes discrete curvature over it:

k_t = |d_t - 2*d_{t-1} + d_{t-2}|

where d_t is the normalized per-module weight divergence from a reference checkpoint. During healthy training, k_t stays near zero. At instability onset it spikes — well before loss diverges.

Two detection channels run in parallel:

• AND mode — divergence step-change AND gradient energy both persistently exceed their z-scored baseline
• Kappa mode — curvature alone persistently exceeds its z-scored baseline

When a trigger fires, the deepest-early-cluster attribution algorithm identifies the specific module that deviated first. A three-phase intervention then fires automatically.

---

Install

pip install bendex-vigil

Or from source:

git clone https://github.com/hannahnine/bendex-vigil
cd bendex-vigil
pip install -e .

---

Usage

One call. Before your forward pass.

from bendex import BendexMonitor, BendexConfig, BendexIntervention

monitor = BendexMonitor(model, config=BendexConfig())
intervention = BendexIntervention(model, optimizer)

for step, batch in enumerate(dataloader):
    event = monitor.observe(step)          # before forward pass
    intervention.step(event, step)

    loss = model(batch)
    loss = loss + intervention.logit_loss(logits)
    loss.backward()
    intervention.apply_grad_clip()
    optimizer.step()
    optimizer.zero_grad()

See examples/basic_usage.py for a complete working example.

---

Configuration

config = BendexConfig(
    warmup_steps=60,     # steps to collect baseline before detection starts
    z_threshold=2.5,     # z-score threshold for detection
    persist=2,           # consecutive steps signal must exceed threshold
    max_lag=6,           # attribution temporal window
)

---

Validated architectures

• Multilayer perceptron (MLP)
• Convolutional neural network (CNN)
• Small Vision Transformer (ViT)
• DistilBERT 66M — SST-2, 30-seed benchmark
• GPT-2 117M — WikiText-2 fine-tuning
• ResNet-50
• Distribution shift benchmark (SST-2 → MNLI)

---

Research

Vigil is the applied proof of a theoretical program in information geometry. Five published papers:

1. Informational Curvature
2. Informational Stability
3. Reflexivity
4. Experience
5. Specialization — derives the fine structure constant α to 8 significant figures from pure geometry

---

License

Source available. Free for research and non-commercial use. Commercial license required for all other use.

See LICENSE for full terms.

Patent pending. Methods covered by provisional patent applications filed by Hannah Nine / Bendex Geometry LLC (priority dates November 2025, February 2026, March 2026).

For commercial licensing: [e-mail] · bendexgeometry.com

---

Author

Hannah Nine — bendexgeometry.com · ORCID 0009-0006-3884-7372