통신 공학 분야에서 AI의 환각 현상을 탐지하기 위해 RF 및 레이더 계산 결과가 물리적 한계 내에 있는지 검증합니다.
다른 이름 JonesRobM/physbound
- 스타
- 2
- Tools
- —
- 다운로드, 30일
- 127
- 최신 버전
- 0.3.0
- 마지막 커밋
- 2026년 11월 9일
- 라이선스
- MIT
PhysBound MCP란?
통신 공학 분야에서 AI의 환각 현상을 탐지하기 위해 RF 및 레이더 계산 결과가 물리적 한계 내에 있는지 검증합니다.
프로젝트 README에서 생성된 요약
PhysBound MCP server 정보
PhysBound는 통신 공학 워크플로우에서 발생하는 수학적 오류와 환각 현상을 잡아내기 위해 설계된 물리 계층 린터(linter)입니다. AI 어시스턴트가 RF 및 레이더 계산을 물리 법칙에 따라 검증할 수 있도록 다양한 검증 도구를 제공합니다. 이 MCP server는 CODATA 상수와 차원 분석을 사용하여 처리량, 안테나 이득, 신호 강도 등에 대해 생성된 수치가 현실적인 범위 내에 있는지 확인합니다.
이 MCP server는 stdio를 통해 로컬에서 실행되며 Claude Desktop, Cursor 또는 Windsurf와 같은 클라이언트에 통합할 수 있습니다. 개발자는 uvx 명령어를 사용하여 수동 설치 없이 격리된 환경에서 최신 버전을 즉시 실행할 수 있습니다. 이 server에는 Friis 방정식을 검증하는 rf_link_budget, 채널 용량을 확인하는 shannon_hartley, 단일 정지 레이더 방정식을 검증하는 radar_range와 같은 특화된 도구가 포함되어 있습니다. 또한 일반적인 공학적 오류를 방지하기 위해 noise_floor 계산, antenna_gain 분석, radar_ambiguity 한계 처리 기능도 제공합니다.
활용 사례
- 1채널 용량 한계 검증특정 대역폭과 SNR이 주장하는 데이터 처리량을 지원할 수 있는지 확인합니다.
- 2레이더 탐지 거리 감사레이더 시스템이 특정 거리에서 특정 크기의 표적을 실제로 탐지할 수 있는지 확인합니다.
프로젝트의 README를 바탕으로 작성되었습니다. 모든 활용 사례는 해당 내용의 인용문과 대조하여 확인합니다.
PhysBound MCP 설치 방법
형식 확인: 2026년 11월 10일Cursor에 추가Goose에 추가Claude Code에 PhysBound MCP 설치
$ claude mcp add physbound -- uvx physboundCursor에 PhysBound MCP 설치
~/.cursor/mcp.json
{
"mcpServers": {
"physbound": {
"command": "uvx",
"args": [
"physbound"
]
}
}
}VS Code에 PhysBound MCP 설치
.vscode/mcp.json
{
"servers": {
"physbound": {
"type": "stdio",
"command": "uvx",
"args": [
"physbound"
]
}
}
}VS Code CLI에 PhysBound MCP 설치
$ code --add-mcp '{"name":"physbound","type":"stdio","command":"uvx","args":["physbound"]}'Claude Desktop에 PhysBound MCP 설치
claude_desktop_config.json
{
"mcpServers": {
"physbound": {
"command": "uvx",
"args": [
"physbound"
]
}
}
}Codex에 PhysBound MCP 설치
$ codex mcp add physbound -- uvx physboundCodex config.toml에 PhysBound MCP 설치
~/.codex/config.toml
[mcp_servers.physbound] command = "uvx" args = ["physbound"]
Gemini CLI에 PhysBound MCP 설치
$ gemini mcp add -s user physbound uvx -- physboundGemini settings.json에 PhysBound MCP 설치
~/.gemini/settings.json
{
"mcpServers": {
"physbound": {
"command": "uvx",
"args": [
"physbound"
]
}
}
}자동 검사
2026년 11월 14일에 실행됨전체 보고서 →자동 검사를 통과했습니다. 이는 보안 감사가 아닙니다. 코드를 직접 읽고 서버에 필요한 최소한의 권한만 부여하십시오. 검사 방식
- 소스 저장소 연결 및 접근 가능jonesrobm/physbound is public and reachablegithub.com/jonesrobm/physbound ↗통과
- 패키지가 이 저장소를 소스로 명시physbound names jonesrobm/physbound as sourcepypi.org/project/physbound/ ↗통과
- 빌드 출처(Sigstore 증명)No build attestation published for this release검사되지 않음
- 오픈 소스 라이선스 (OSI)MITspdx.org/licenses/MIT.html ↗통과
- 최근 1년 내 커밋 또는 릴리스Last commit or release 2026-11-09github.com/jonesrobm/physbound ↗통과
- OpenSSF Scorecard검사되지 않음
- 패키지 이름이 인기 패키지와 유사하지 않음No popular package one edit away from physbound통과
- 패키지 및 해당 종속성에 중요한 보안 권고 없음74 resolved dependencies checked; no critical advisoryosv.dev ↗통과
- 모델에 대한 숨겨진 지침 없음0 tool descriptions and the README read; no override phrases, credential paths or hidden characters통과
- 파괴적인 도구에 라벨 표시됨검사되지 않음
- 원격 엔드포인트가 MCP 핸드셰이크에 응답검사되지 않음
README
GitHub에서 보기LLMs generate plausible-sounding RF numbers that violate fundamental physics — throughput above the Shannon limit, antenna gains no aperture can produce, radar ranges the range equation forbids. PhysBound gives any MCP-compatible AI assistant six validated calculation tools, backed by CODATA constants (via SciPy) and dimensional analysis (via Pint). Impossible claims return structured PhysicalViolationError responses with the violated law, the computed limit, and a LaTeX explanation — not silent failures.
| Tool | What it validates |
|---|---|
rf_link_budget |
Friis link budgets: FSPL, received power, antenna gains vs. aperture/Harrington limits |
shannon_hartley |
Throughput claims against channel capacity C = B log2(1 + SNR) |
noise_floor |
Thermal noise kTB, Friis noise-figure cascades, receiver sensitivity |
radar_range |
Detection-range claims against the monostatic radar range equation |
antenna_gain |
Gain limits, beamwidth, and far-field distance for a single antenna |
radar_ambiguity |
Pulse-Doppler unambiguous range/velocity, Doppler aliasing, range resolution |
Installation
PhysBound is a standard stdio MCP server published on PyPI. The recommended launch command is uvx physbound, which fetches and runs the latest release in an isolated environment — no manual install step.
First run:
uvxdownloads ~60 MB of dependencies (SciPy, NumPy) the first time. Runuvx physboundonce in a terminal to pre-cache them (Ctrl-C to exit); subsequent starts are instant.
Claude Code
claude mcp add physbound -- uvx physbound
Claude Desktop
Add to claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/, Windows: %APPDATA%\Claude\):
{
"mcpServers": {
"physbound": {
"command": "uvx",
"args": ["physbound"]
}
}
}
Cursor, Windsurf, and other MCP clients
Use the same JSON server entry as above in your client's MCP configuration file (Cursor: ~/.cursor/mcp.json; Windsurf: ~/.codeium/windsurf/mcp_config.json).
Without uv
If you prefer a plain Python install (requires Python 3.12+):
pip install physbound
then set "command": "physbound" (with no args) in the client configuration.
Once configured, ask your assistant an RF question — "Can a 20 MHz channel with 15 dB SNR support 500 Mbps?" — and it will answer with physics-validated numbers.
Command-Line Usage
The same validators are available directly from the terminal — useful in scripts and CI pipelines, since a physics violation exits with code 1:
$ physbound check shannon --bandwidth-hz 20e6 --snr-db 15 --claimed-throughput-bps 500e6
PHYSICS VIOLATION [Shannon-Hartley Theorem]
Claimed throughput 500000000.0 bps exceeds Shannon limit of 100556153.5 bps by 397.2%
Computed limit: 1.00556e+08 bps
Claimed value: 5e+08 bps
$ echo $?
1
$ physbound check shannon --bandwidth-hz 20e6 --snr-db 15 --claimed-throughput-bps 100e6
Shannon-Hartley Capacity: ... # exit code 0
Available subcommands: check {link-budget, shannon, noise, radar-range, antenna, radar-ambiguity} (flags mirror the MCP tool parameters; add --json for the full structured result) and serve --transport {stdio,http} for running the MCP server over HTTP. See physbound check --help for details.
What LLMs Get Wrong
Sixteen real hallucination patterns, each caught by PhysBound's test suite:
| # | Category | LLM Hallucination | PhysBound Truth | Verdict |
|---|---|---|---|---|
| 1 | Shannon-Hartley | A 20 MHz 802.11n channel with 15 dB SNR can achieve 500 Mbps | Shannon limit: 100.6 Mbps (not 500 Mbps) | CAUGHT |
| 2 | Shannon-Hartley | A 100 MHz 5G channel with 20 dB SNR delivers 2 Gbps | Shannon limit: 665.8 Mbps (not 2000 Mbps) | CAUGHT |
| 3 | Antenna Aperture | A 30 cm dish antenna at 1 GHz provides 45 dBi gain | Physical limit: 12.1 dBi (Harrington); aperture eta=1: 9.9 dBi; typical dish: 7.4 dBi (eta=0.55) (not 45 dBi) | CAUGHT |
| 4 | Thermal Noise | Receiver noise floor of -180 dBm/Hz at room temperature | Thermal noise floor: -174.0 dBm/Hz at 290K (not -180 dBm/Hz) | CAUGHT |
| 5 | Link Budget / FSPL | Wi-Fi at 2.4 GHz with 20 dBm TX reaches 10 km with -40 dBm RX power | Actual RX power at 10 km: -94.1 dBm (not -40 dBm) | CAUGHT |
| 6 | Link Budget / FSPL | A 1W transmitter at 12 GHz with 0 dBi antennas reaches GEO at -80 dBm | Actual RX power at GEO: -175.1 dBm (not -80 dBm) | CAUGHT |
| 7 | Link Budget / FSPL | Bluetooth at 2.4 GHz with 0 dBm TX and 0 dBi antennas reaches 1 km at -60 dBm | Actual RX power at 1 km: -100.1 dBm (not -60 dBm) | CAUGHT |
| 8 | Shannon-Hartley | A 10 MHz LTE channel at 10 dB SNR supports 1 Gbps | Shannon limit: 34.6 Mbps (not 1000 Mbps) | CAUGHT |
| 9 | Noise Cascade | Receiver NF is the same regardless of stage order: LNA(20dB/1.5dB) + Mixer(10dB/8dB) | LNA first: 1.66 dB vs mixer first: 8.03 dB (penalty: 6.4 dB) | CAUGHT |
| 10 | Antenna Aperture | A 10 cm patch antenna at 900 MHz provides 20 dBi gain | Physical limit: 4.4 dBi (Harrington, D < lambda); aperture eta=1: -0.5 dBi; typical: -3.1 dBi (eta=0.55) (not 20 dBi) | CAUGHT |
| 11 | Radar Range Equation | Doubling transmit power doubles radar detection range | Range increases by factor 1.189 (2^(1/4) = 1.189), not 2.0 | CAUGHT |
| 12 | Radar Range Equation | Small drone (0.01 m^2 RCS) detectable at 200 km by 1 kW X-band radar with 30 dBi gain | Max range: 2.7 km for 0.01 m^2 RCS at 1 kW X-band (not 200 km) | CAUGHT |
| 13 | Antenna Gain | A 0.5 m user-terminal dish at 12 GHz gives 50 dBi gain | Physical limit: 36.1 dBi (Harrington); typical dish: 33.4 dBi (eta=0.55) (not 50 dBi) | CAUGHT |
| 14 | Antenna Gain | A 3 m dish at 10 GHz is in its far field at 10 m, so gain can be measured there | Far-field distance 2D^2/lambda = 600 m; 10 m is in the near field | CAUGHT |
| 15 | Radar Ambiguity | A 10 GHz radar at 10 kHz PRF unambiguously measures 500 m/s | v_ua = lambda*PRF/4 = +/-74.9 m/s (blind speed 149.9 m/s); 500 m/s aliases (not 500 m/s) | CAUGHT |
| 16 | Radar Ambiguity | A 10 GHz pulse-Doppler radar can unambiguously cover 150 km and +/-300 m/s at once | R_uav_ua = clambda/8 = 1.123e+06 m^2/s for any PRF; claimed 4.5e+07 m^2/s (40x too large) | CAUGHT |
Generated automatically by pytest tests/test_marketing.py -s
Tools
Full derivations, sources, and worked examples for every formula are in docs/formulas.md.
rf_link_budget
Computes a complete RF link budget using the Friis transmission equation and validates antenna gains against physical limits.
Example: "What's the received power for a 2.4 GHz link at 100 m with 20 dBm TX, 10 dBi TX gain, 3 dBi RX gain?"
Returns FSPL, received power, wavelength, and — when antenna diameters are supplied — gain limit checks. Gains above the hard bound G_max = max((pi D / lambda)^2, (ka)^2 + 2ka) (the eta = 1 aperture value, or Harrington's bound for electrically small antennas) are rejected; gains above the typical dish value (eta = 0.55) are accepted with a warning. Also warns inside the far-field distance 2D^2/lambda and rejects negative losses.
shannon_hartley
Computes Shannon-Hartley channel capacity C = B log2(1 + SNR) and validates throughput claims.
Example: "Can a 20 MHz channel with 15 dB SNR support 500 Mbps?"
Returns theoretical capacity, spectral efficiency, and whether the claim is physically possible, including the exact percentage by which a violating claim exceeds the Shannon limit.
noise_floor
Computes thermal noise power N = k_B T B, cascades noise figures through multi-stage receivers with the Friis noise formula, and calculates receiver sensitivity.
Example: "What's the noise floor for a 1 MHz receiver at 290 K with a two-stage LNA chain?"
Returns thermal noise in dBm and watts, cascaded noise figure, effective input noise temperature T_e = 290 K * (F - 1), and receiver sensitivity.
radar_range
Computes the monostatic radar range equation R_max = [P_t G^2 lambda^2 sigma / ((4pi)^3 S_min L)]^(1/4) and validates detection-range claims.
Example: "Can a 1 kW X-band radar with 30 dBi gain detect a 0.01 m^2 drone at 200 km?"
Returns maximum detection range, minimum detectable signal, wavelength, and intermediate values. Catches the common fourth-root fallacy that doubling power doubles range.
antenna_gain
Analyses a single antenna from its diameter or physical area: gain limits, beamwidth, and far-field distance, with optional validation of a claimed gain.
Example: "Can a 0.5 m dish at 12 GHz really give 50 dBi? What beamwidth should I expect?"
Returns the physical gain limit (aperture or Harrington bound, and which one governs), typical gain at the given efficiency (default 0.55), effective aperture, half-power beamwidth estimates, far-field distance 2D^2/lambda, and — for a claimed gain — the implied aperture efficiency and validity.
radar_ambiguity
Computes pulse-Doppler ambiguity limits for a given carrier frequency and PRF.
Example: "Can a 10 GHz radar at 10 kHz PRF unambiguously measure a 500 m/s target out to 150 km?"
Returns unambiguous range R_ua = c / (2 PRF), unambiguous velocity v_ua = lambda PRF / 4, blind speed, Doppler shift and apparent (aliased) velocity, range resolution (c tau / 2, or c / 2B with pulse compression), minimum range, duty cycle, and the range-Doppler dilemma invariant R_ua v_ua = c lambda / 8. Rejects range, velocity, or resolution claims the PRF or pulse cannot support.
Physics Guarantees
Every calculation is validated against hard physical limits:
- Speed of light:
c = 299,792,458 m/s— no exceptions - Thermal noise floor:
N = -174 dBm/Hzat 290 K — the IEEE standard reference - Shannon limit:
C = B log2(1 + SNR)— no throughput claim exceeds this - Antenna gain limit:
G_max = max((pi D / lambda)^2, (ka)^2 + 2ka)— the eta = 1 aperture value for planar apertures and Harrington's bound for any antenna in a sphere of diameter D; eta = 0.55 is a warning threshold, not a limit - Radar range equation:
R_max = [P_t G^2 lambda^2 sigma / ((4pi)^3 S_min)]^(1/4)— range obeys the fourth-root law - Receiver sensitivity:
S_min = k (T_A + T_0 (F - 1)) B * SNRwithT_e = T_0 (F - 1)referenced to 290 K - Unambiguous range:
R_ua = c / (2 PRF)— echoes beyond it fold into a later pulse interval - Unambiguous velocity:
v_ua = lambda PRF / 4— Doppler is sampled at the PRF, so|f_d| <= PRF/2 - Range-Doppler dilemma:
R_ua * v_ua = c lambda / 8— no PRF choice beats it for a given carrier
Violations return structured PhysicalViolationError responses with LaTeX explanations, not silent failures.
Examples and Documentation
- Catching Hallucinations — walkthrough of five real LLM failure modes with full JSON responses
- Interactive Demo Notebook — Jupyter notebook calling the physics engines directly
- Formula Reference — every formula, constant, and validation guard with sources
- Changelog — release history
Development
Requires uv and Python 3.12+ (CI covers 3.12–3.14):
# Clone and install
git clone https://github.com/JonesRobM/physbound.git
cd physbound
uv sync --all-extras
# Run checks
uv run pytest tests/ -v
uv run ruff check src/ tests/
uv run mypy src/physbound/
# Regenerate the hallucination table above
uv run pytest tests/test_marketing.py -s
# Start the MCP server locally
uv run physbound
Contributions that expand the set of validated physics domains are welcome — see CONTRIBUTING.md for the architecture guide and the step-by-step recipe for adding a new validator.
Use Cases
- RF system design review — validate link budgets, receiver sensitivity, and noise cascades
- Telecom proposal vetting — catch impossible throughput claims before they reach a customer
- Radar system sizing — sanity-check detection range, PRF selection, and ambiguity trade-offs
- Educational tools — teach Shannon-Hartley, Friis transmission, and thermal noise with verified calculations
- CI/CD for physics — integrate as a validation step in engineering pipelines
Support
If PhysBound is useful in your work, consider buying me a coffee.
License
MIT License. See LICENSE.
Related
- Model Context Protocol — the open standard for AI tool integration
- MCP Server Registry — official directory of MCP servers
- FastMCP — Python framework for building MCP servers
PhysBound MCP 질문
PhysBound MCP를 Claude Code에 추가하려면 어떻게 하나요?
터미널에서 claude mcp add physbound -- uvx physbound를 실행합니다. 모든 프로젝트에서 사용할 수 있게 하려면 --scope user를 추가합니다.
PhysBound MCP를 Cursor에 추가하려면 어떻게 하나요?
~/.cursor/mcp.json를 열고 mcpServers 아래에 서버를 추가합니다. 위 설치 섹션에 표시된 Cursor 구성을 사용합니다.
PhysBound를 어떻게 설치하나요?
'pip install physbound'를 통해 PyPI에서 설치하거나, 격리된 환경에서 'uvx physbound'를 사용하여 실행할 수 있습니다.
계산 결과가 물리 법칙을 위반하면 어떻게 되나요?
오류를 무시하는 대신, 위반된 법칙, 계산된 한계값, LaTeX 형식의 설명이 포함된 구조화된 PhysicalViolationError 응답을 반환합니다.