
Noether Physics
io.github.RanshivPublished by GitHub account Ranshivv0.1.2Updated Oct 24, 2026
Physics assistant: checked equations, live simulations, claims verified against arXiv papers
Overview
A physics research assistant that explains concepts, runs simulations, checks equations and verifies citations against real arXiv papers.
- What you can do
- Explain physics concepts with analogies, dimension-checked equations and recomputed worked examples
- Run six physics simulations (pendulum, oscillators, projectile, Rabi, decoherence, Jaynes-Cummings) with drift and closed-form comparisons
- Check equations for dimensional consistency and flag undeclared symbols as unknown
- Ingest arXiv LaTeX sources, list their equations and search across the paper corpus
- Audit a draft's bibliography, resolving references against arXiv, Crossref, OpenAlex and INSPIRE
- When to use it
- Useful when an assistant needs physics answers with verifiable grounding: explaining concepts, running quick simulations, checking equations, or catching fabricated or unresolvable references in a draft before submission.
- Requirements
- Runs locally over stdio as a PyPI package (noether-physics) with the mcp extra; needs Python 3.12+ and network access to arXiv and citation indexes. The quantum simulations need the quantum extra (QuTiP). No accounts, API keys or environment variables are declared.
Installation
In SourceWeft
- Open Noether Physics in the dashboard and add it to a workspace.
- Enable the server for the chats that should use its tools.
Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.
Other MCP clients
Follow the launch instructions in the repository.
README
NOETHER
A physics research assistant that shows its work.
Explains physics concepts with everyday analogies, dimension-checked equations, worked examples it recomputes, and simulations it actually runs. Answers questions from real arXiv papers with every sentence anchored to a verbatim span. Audits your own draft's bibliography for references that don't exist.
On 0.1.0 or 0.1.1? Upgrade. Those releases confirm fabricated references — the specific failure this tool exists to prevent. See CHANGELOG.md.
Read this before you trust it
This is a research-integrity tool, so its limits matter more than its features.
The verification gate checks that a quote is real. It does not check that the quote supports the claim. A claim can contradict its own citation and still pass. Demonstrated during development: a claim asserting "the experiment detected a nonzero neutron EDM at five sigma" was accepted while citing a passage reporting d_n = 0.0 ± 1.1 × 10⁻²⁶ e·cm — a null result. The quote was genuine, so the gate let it through. Closing that gap needs entailment checking, which needs a language model, and keeping the model outside the verifier is what makes the other guarantees meaningful. So every answer this tool emits ends by saying so, and you should read the quotes.
The reference resolver is now measured in both directions. It resolves 86.5% of real references and was fooled by 0 of 8 deliberately fabricated ones. That second number did not exist until a labelled benchmark was built for it — and the first run of that benchmark caught the resolver confirming 80% of fabrications, some at 0.95 confidence. Three defects were found and fixed. Eight fabrications is still a small sample: 0% means "fooled by none of eight", not "cannot be fooled".
It is small. 58 papers in the test corpus, 8 concepts in the explanation library, quantum coverage limited to two-level systems and one cavity mode. No PDF fallback, so roughly 9% of quant-ph papers (measured: 5 of 55) cannot be ingested at all.
docs/DEFERRED.md is the full account — every gap, with the measurement behind
it. It is the honest document; this README is the summary.
Why it exists
2026 audits found a 17% phantom-citation rate in AI-assisted survey papers — references resolving to no digital object — and roughly 146,900 AI-generated fake citations across 2.5M papers in 2025. arXiv now bans authors who submit hallucinated references.
Physics has a structural advantage nobody exploits: arXiv distributes the actual LaTeX source.
Equations arrive with their \labels, citations with their \cite keys, sections with their real
structure — read, not guessed. Tools in other fields OCR PDFs because their domains have no source.
Physics does.
What it does
Explains. Curated prose, not generated — then machine-checked. Every equation is parsed and dimension-checked; every worked example is recomputed and compared against its stored answer, so a typo in the library fails CI rather than teaching you a wrong number.
Simulates. Six models — pendulum, damped/driven oscillator, projectile with drag (SciPy); Rabi, decoherence, Jaynes-Cummings (QuTiP). Every run halves the step size and reports the drift, and where a closed form exists the numerics are compared against it automatically.
Checks equations. Undeclared symbols return unknown, never a guess — h is Planck's constant
or a height, T a period or a temperature. unknown is not a pass.
Reads papers. Ingests LaTeX source, so "equation 14" means the equation the authors numbered 14.
Audits your draft. Resolves every \bibitem against arXiv, Crossref, OpenAlex and INSPIRE.
Exits non-zero on unresolvable references or cite keys with no bibliography entry, so it works as a
pre-commit hook.
Measured results
Run noether bench status yourself. As of the last run:
Read these with their scope attached. "Right paper in the top 3" over 58 papers is a real task; over
8 it was nearly free, and an earlier run reported a meaningless 100% because top_k exceeded the
corpus size. The citation threshold is 85% rather than 100% because books, theses and pre-digital
articles are genuinely not in any index — the residual failures on the nEDM letter are a 1954 Danish
journal, a 1955 Pauli chapter and a 1982 Oxford book.
Use it from Claude
14 tools. Claude reasons and writes; NOETHER retrieves, checks and records. There is deliberately no
write_answer tool — the component that can hallucinate is not the component that decides what
ships.
Install
Python 3.12+. From source: uv venv && uv pip install -e ".[dev,quantum,mcp]", then uv run pytest
(288 tests, no network — cassettes default to replay and a miss is an error, not a live fallback).
Terms
arXiv forbids bulk crawling. This fetches one paper at a time behind a 3-second interval and caches
permanently, so a given paper is downloaded exactly once. Every source, its rate limit and its terms
are declared in research/sources/source_registry.yaml; a source not in that file is not called.
No AI-generated imagery. Every figure is extracted from a real paper or rendered deterministically from code and data, and is tagged with which.
Licence
MIT — see LICENSE. Release history: CHANGELOG.md.
The name honours Emmy Noether. The PyPI distribution is noether-physics because noether belongs
to an unrelated physical-units library; the command is
noether either way.
Source: README.md at commit 35faf21
Tools
0Version history
1- v0.1.2LatestSep 16, 2026
Score
50% of the applicable checks met
How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.
- Reached in the last 30 days25 pts
Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.
- Auth declared10 pts
Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.
- Tools listed15 pts
Answer tools/list without a credential, or ship a package that starts in a sandbox without one.
- DescriptionRequired15 pts
- Logo10 pts
- Documentation15 pts
- License10 pts