AETHER
Know when to automate vs. escalate
Every automation system asks: "Should I handle this, or escalate to a human?"
Most answer with static thresholds — if confidence < 90%, flag for review. This breaks in two ways:
- A well-calibrated model gets held back by thresholds tuned for a bad one
- Not all uncertainty is equal — sometimes more data would help, sometimes the process is just inherently random
AETHER solves this by decomposing uncertainty into what's reducible (more data helps) vs. what's irreducible (inherently random), then dynamically adjusting governance based on demonstrated model performance.
The result: fewer unnecessary escalations, fewer missed issues.
Your invoice processing bot flags 40% of invoices for human review because "confidence is below threshold."
But when you dig in:
- Half those flags are cases where the model just hasn't seen enough similar invoices yet — training on more examples would help
- The other half are vendors with legitimately unpredictable behavior — no amount of human review changes a coin flip
Static thresholds can't tell the difference. AETHER can.
| Capability | What It Means For You |
|---|---|
| Predicts next activities | Know what happens next in your order-to-cash, purchase-to-pay, or procurement process |
| Quantifies uncertainty | Not just "80% confident" but "60% of that uncertainty is reducible with more data" |
| Adapts governance thresholds | Thresholds tighten when the model is uncertain, relax when it's proven accurate |
| Provides prediction sets | Instead of one answer, get a calibrated set of likely outcomes with coverage guarantees |
| Tracks its own calibration | Know when the model is degrading before it causes problems |
effective_threshold = base × mode_factor × uncertainty_factor × calibration_factor
- Only reducible uncertainty tightens governance — the system won't waste human attention on inherently random outcomes
- Trust is earned slowly, lost quickly — 10 good windows to level up, 1 critical miss to demote
- Some constraints never relax — sensitive data patterns, high-conflict decisions, and circuit breakers always trigger review
Invoice arrives → AETHER predicts: approve/reject/needs-info
→ Uncertainty: 15% total (12% epistemic, 3% aleatoric)
→ Decision: HIGH epistemic ratio → route to human
(more training data on this vendor type would help)
Order in progress → AETHER predicts: on-time (73%), late (27%)
→ Uncertainty: 8% total (1% epistemic, 7% aleatoric)
→ Decision: LOW epistemic ratio → trust the prediction
(this route is just variable, human review won't help)
→ Action: Trigger proactive customer notification if >25% late probability
Application submitted → AETHER predicts: next step is "credit_check" (89%)
→ Conformal set: {credit_check, document_request} (90% coverage)
→ Governance: AUTO-APPROVE routing
(model well-calibrated, low uncertainty)
git clone https://github.com/chrbailey/aether.git
cd aether
# TypeScript (governance + MCP server)
npm install && npm run build
# Python (ML core)
pip install -e ".[dev]"# Terminal 1: Python inference server
python -m core.inference.server # Starts on localhost:8712
# Terminal 2: MCP server (connects to Claude, Cursor, etc.)
npm startnpm test # 99 TypeScript tests
python -m pytest core/tests/ -v # 303 Python testsAETHER exposes 7 tools via the Model Context Protocol:
| Tool | What You Get |
|---|---|
predict_next_event |
Top-K next activities with probabilities + uncertainty decomposition + conformal set |
predict_outcome |
Will this case be on-time, late, or need rework? With confidence intervals |
get_calibration |
Is the model trustworthy right now? ECE, MCE, Brier scores |
get_autonomy_level |
Current trust level: SUPERVISED → GUIDED → COLLABORATIVE → AUTONOMOUS |
get_effective_thresholds |
All 6 adaptive thresholds with full breakdown of why |
evaluate_gate |
Should this specific decision be auto-approved, held, or blocked? |
get_production_metrics |
Latency, prediction counts, calibration drift alerts |
{
"mcpServers": {
"aether": {
"command": "node",
"args": ["/path/to/aether/mcp-server/dist/index.js"]
}
}
}AETHER uses a JEPA-style architecture (Joint Embedding Predictive Architecture) adapted for business event sequences.
JEPA was developed by Yann LeCun's team for learning world models from images and video. AETHER asks: can the same approach model enterprise workflows, where the "world" is a structured sequence of business events?
MCP Tools (7)
predict_next_event
predict_outcome
get_calibration
get_autonomy_level
get_effective_thresholds
evaluate_gate
get_production_metrics
|
TypeScript MCP Server
├── Governance Modulation (adaptive thresholds)
├── Autonomy Controller (asymmetric trust)
├── Immutable Constraints (safety floor)
│ |
│ HTTP bridge (:8712)
│ |
Python FastAPI Server
├── EventEncoder (activity + time + context → 128D)
├── TransitionModel (JEPA predictor: z_t → z_{t+1})
├── EnergyScorer (energy-based anomaly scoring)
├── HierarchicalPredictor (activity / phase / outcome)
├── UncertaintyDecomposer (epistemic vs. aleatoric)
├── CalibrationTracker (ECE / MCE / Brier)
└── ConformalPredictor (distribution-free prediction sets)
| Concept | What It Means |
|---|---|
| Epistemic uncertainty | Model doesn't know — more data would help |
| Aleatoric uncertainty | Process is random — more data won't help |
| Conformal prediction | Calibrated prediction sets with coverage guarantees (e.g., "90% of the time, the true answer is in this set") |
| Adaptive thresholds | Governance limits that tighten/loosen based on demonstrated model performance |
Some constraints are immutable regardless of trust level:
- Forbidden mode → always block
- Sensitive data patterns (SSN, API keys) → always hold for human review
- Dempster-Shafer conflict > 0.7 → always review (the model is confused)
- Circuit breaker (3+ consecutive failures) → block
- Total uncertainty > 0.95 → hold
Evaluated on 11 process mining datasets:
| Dataset | Domain | Cases | Activity Accuracy | Notes |
|---|---|---|---|---|
| BPI 2017 | Finance (Loans) | 31,509 | 70.4% | Primary benchmark — simple rules only achieve 49.6% |
| Road Traffic Fine | Government | 150,370 | 81.6% | High volume, deterministic outcomes |
| SAP Workflow | Enterprise | 2,896 | 68.2% | Best enterprise result |
See docs/BENCHMARK_COMPARISON.md for full analysis.
| Approach | Limitation | AETHER Advantage |
|---|---|---|
| Celonis / UiPath | Static thresholds, no uncertainty decomposition | Adaptive governance that learns |
| Custom ML models | Raw confidence scores without calibration | Calibrated predictions with coverage guarantees |
| PM4Py | Discovery only, no prediction layer | Full prediction + governance stack |
| Rules engines | Brittle, manual tuning required | Self-adjusting based on performance |
AETHER builds on established research:
- JEPA — LeCun, 2022. A Path Towards Autonomous Machine Intelligence
- Conformal Prediction — Gibbs & Candes, NeurIPS 2021. Distribution-free prediction sets
- VICReg / SIGReg — Bardes, Balestriero & LeCun. Latent collapse prevention
- Energy-Based Models — LeCun et al., 2006. Anomaly scoring framework
MIT — Christopher Bailey, 2026