Scan Any AI Model.
Understand Every Neuron.
Probe56 captures neuron data. GI X-Ray maps it to the 56-element periodic table. GI Surgeon fixes what's broken — without retraining.
pip install probe56 && probe56 scan --model gpt2 --output my_model.p56 How It Works
Extract
Probe56 (open-source Python CLI) scans your model during inference and captures neuron state data from your dataset.
probe56 scan --model your/model --dataset truthfulqa Bundle
Neuron data is compressed into a P56 bundle — a signed, deterministic binary format with cryptographic hash of the model weights. Portable, verifiable.
→ scan_bundle.p56 (SHA-256 verified) Analyze
GI X-Ray maps your bundle against the Holosynthics 56-element periodic table. Produces a heatmap of which elements dominate which layers, and flags anomalies.
→ 56-element report + anomaly score Repair (optional)
GI Surgeon applies targeted weight edits to suppress noise neurons and amplify signal neurons. No retraining. Fully reversible. Validated against regression suite.
→ patched model weights See Your Model's Neuron DNA
Send us your Probe56 bundle. We run the full 56-element analysis and return a comprehensive report showing which neuron types dominate each layer, where anomalies concentrate, and what your hallucination risk profile looks like.
- Full periodic table heatmap per layer
- Anomaly cluster identification
- Entropy.Confidence trace (hallucination predictor)
- Cross-model comparison
- Signed PDF report + raw JSON data
Track Neuron Drift in Real Time
As models are updated, fine-tuned, or exposed to distribution shifts, their neuron composition changes. GI Monitor detects these changes before they manifest as output degradation.
- Continuous neuron composition tracking
- Drift alerts before output quality degrades
- Slack and PagerDuty integration
- Monthly composition summary
- Comparison across model versions
Fix Hallucinations Without Retraining
Targeted neuron-level edits in weight space. GI Surgeon identifies the specific neuron clusters responsible for a hallucination pattern and applies precision edits — amplification, suppression, or rerouting.
- Weight-space edits (permanent, reversible)
- Amplification · Suppression · Rerouting
- 50-prompt regression suite before every commit
- Minimal perplexity impact (validated)
- Air-gap deployment available
5 Dimensions of AI Model Health
Every Probe56 scan scores your model across five dimensions. A weak score in any one of them predicts where failures will appear in production.
Knowledge
The AI knows factual information correctly and does not invent facts it was never trained on.
Reasoning
The AI draws logically valid conclusions from premises — cause and effect, if-then, because.
Skills
The AI can perform tasks accurately — translation, coding, summarization, calculation.
Logical Generalization
The AI applies rules consistently across new cases it has not seen before during training.
Judgment
The AI makes sound decisions under uncertainty and knows the limits of what it knows.
GI Surgeon vs. the Field
Benchmarks run against CounterFact and RippleEdits. Every result is Ed25519 signed and independently reproducible.
Method protected by provisional patents filed with the USPTO.
Two-Layer Repair
One repair pass fixes the obvious problems. Two passes fix what's hiding underneath — without touching anything that already works.
Case Study: Medical AI Provider
A real product run. Numbers from the signed report. Nothing fabricated.
A medical AI provider's model was showing unexpected behavior on clinical knowledge questions. They ran a Probe56 scan, then a GI Surgeon repair cycle.
The scan found 78 anomalous neuron signatures across the model. Most were structural — they're how every model works. Two were fixable at the weight level without retraining.
After the Two-Layer Repair: both target behaviors corrected, all 20 preserved behaviors still passing, zero side effects detected in the regression suite.
Report fingerprint: a9c492a3… · Ed25519 signed
Get Your Free Scan Report
Send us your model identifier. We run a free Probe56 scan and deliver a signed health report by email within 24 hours.
No commitment. No credit card. Just the scan.
Get Started
Probe56 is free and open-source. GI X-Ray, Monitor, and Surgeon are managed services — contact us for access.
- Probe56 CLI
- Neuron data capture
- P56 bundle format
- MIT license
- Any HuggingFace model
- Full 56-element scan
- Layer-by-layer analysis
- Anomaly heatmap
- Hallucination risk report
- Signed PDF delivery
- Email in 24h
- Real-time drift detection
- Neuron composition alerts
- Dashboard access
- Monthly summary report
- Up to 3 models
- Slack/PagerDuty integration
- Targeted neuron editing
- Validated hallucination fix
- 50-prompt regression suite
- Minimal perplexity impact
- Air-gap deployment
- Enterprise contract
Start Scanning Your Model Today
Probe56 captures neuron data from any HuggingFace model. Free, open-source, and runs locally.
pip install probe56