---

Why ACIES?

Every perception system wastes resources processing information that doesn't change the decision. ACIES solves this by treating perception as a cost-risk optimization problem.

	Without ACIES	With ACIES
Resolution	Fixed (max)	Adaptive (optimal)
Cost	385.8	93.6
Accuracy	98.8%	90.8%
Savings	—	76%

Trade 8% accuracy for 76% cost reduction. Or tune the threshold to find your own sweet spot.

---

Quick Start

Python (stdlib only, zero dependencies)

from acies import APCController, APCConfig, HardwareProfile

apc = APCController(APCConfig(
    confidence_threshold=0.92,
    hardware=HardwareProfile.jetson_orin(),
))

def clarity_fn(action):
    clarities = {
        "64p": 0.55, "128p": 0.65, "224p": 0.75,
        "320p": 0.82, "512p": 0.88, "1024p": 0.93,
        "crop_224": 0.85, "crop_320": 0.90, "crop_512": 0.92,
    }
    return clarities.get(action.name, 0.5)

result = apc.run(true_class=1, clarity_fn=clarity_fn)
print(f"Decision: {result.decision} | Cost: {result.total_cost:.1f} | Steps: {result.n_steps}")

Go CLI (32,800 runs/sec)

go build -o acies-cli .

./acies-cli run --hardware jetson --verbose
./acies-cli bench --iterations 5000 --hardware rpi

Docker

docker build -t acies .
docker run acies bench --iterations 1000

---

Installation

git clone https://github.com/NICE-DEV226/ACIES.git
cd ACIES

Component	Command	Notes
Python	(none)	stdlib only, zero deps
C++ accelerator	cd cpp && make	Optional, 3-5x speedup
Go CLI	go build -o acies-cli .	Single binary
Docker	docker build -t acies .	Multi-stage, all-in-one

Verify:

python3 test_apc.py          # 8/8 tests
./acies-cli version          # ACIES v0.1.0

---

Architecture

Control loop: Sample → Score (ΔR/C) → Adjust → Filter → Execute → Observe → Update → Repeat

---

How It Works

Step	Component	What it does
1	Belief Tracking	Maintains P(Y=1 | observations) via Bayesian filtering
2	Clarity Estimation	Thompson Sampling learns P(correct | action) online
3	Action Scoring	Computes ΔR/C (risk reduction per cost) for each action
4	Conviction	Anti-oscillation boosts high-clarity actions near threshold
5	Safety Filtering	Rejects actions that could exceed risk threshold
6	Change-Point Detection	Resets posteriors on distribution shifts
7	Decision	Stops when confidence ≥ threshold or max steps reached

---

Benchmarks

MNIST (10,000 real images)

python3 examples/real_benchmark.py

Method	Accuracy	Cost	Savings
Fixed 1024p	98.8%	385.8	—
ACIES	90.8%	93.6	76%
Fixed 224p	82.1%	76.5	80%
Random	95.5%	149.6	61%

Go CLI Performance

./acies-cli bench --iterations 5000

32,800 runs/sec — 70× faster than Python (476 images/sec)

Hardware Profiles

Profile	APC Cost	Fixed 1024p	Savings
Default	260.9	448.4	42%
Jetson Orin	204.4	347.1	41%
Raspberry Pi 5	287.5	525.2	45%
Desktop GPU	290.6	479.0	39%
Edge TPU	53.4	92.0	42%

---

Testing

python3 test_apc.py

#	Test	Description	Result
1	Base functionality	Accuracy, cost, exploration	✓
2	Hard tasks	High difficulty scenarios	✓
3	Distribution shift	Thompson adaptation	✓
4	Sensor failure	Degraded reliability	✓
5	Hardware profiles	All 5 profiles	✓
6	Stress test	5,000 iterations	✓
7	Belief math	Bayesian verification	✓
8	Thompson convergence	Posterior accuracy	✓

---

Project Structure

ACIES/
├── acies/                  # Python package (8 modules)
│   ├── controller.py       # Main APC loop
│   ├── belief.py           # Bayesian belief tracker
│   ├── clarity_learner.py  # Thompson Sampling
│   ├── safety.py           # Risk guarantees
│   ├── conviction.py       # Anti-oscillation
│   ├── change_point.py     # BOCPD shift detection
│   ├── actions.py          # Action space & HW profiles
│   └── accelerator.py      # C++ ctypes wrapper
│
├── cpp/                    # C++ core library
│   ├── belief.h/.cpp       # Belief state
│   ├── clarity_learner.*   # Thompson Sampling
│   ├── acies.h/.cpp        # C API
│   └── Makefile
│
├── core.go                 # Go implementation
├── main.go                 # Go CLI (run/bench/config)
│
├── test_apc.py             # 8 robustness tests
├── examples/               # Benchmarks & demos
├── docs/                   # 9 documentation files
├── assets/                 # Logo & visual identity
├── Dockerfile              # Multi-stage build
└── README.md

---

Documentation

Doc	Description
Installation	Setup guide for Python, Go, C++, Docker
Architecture	Deep dive into control loop and algorithms
Configuration	All configuration parameters
CLI Reference	Go CLI commands and flags
Python API	Complete Python API reference
C++ API	C API for FFI (Python/Go/Rust)
Examples	10 usage examples
Benchmarks	Performance results on MNIST
Contributing	Development guide

---

Tech Stack

Layer	Tech	Performance
CLI	Go	32,800 runs/sec
Core	Python (stdlib)	476 images/sec
Accelerator	C++ via ctypes	3-5× Python speed
Tests	8 robustness tests	all passing
Benchmark	5 methods × 5 HW profiles	+ MNIST 10k
Deploy	Multi-stage Dockerfile	Python + Go + C++

---

Contributing

See docs/contributing.md.

License

MIT License — see LICENSE.

Citation

@software{acies2026,
  title = {ACIES: Adaptive Perception Control},
  year = {2026},
  url = {https://github.com/NICE-DEV226/ACIES}
}

---