Caustryn 0.2.1: a clearer identity, compatible experiments

Phaseglass is now Caustryn, a studio for exploring the locations that explain
a simulated radio signal and designing measurements that challenge the answer.

This is an identity and compatibility release. It updates the README, local
Studio and recorded browser replay branding, source links and wave mark, and
adds caustryn as the preferred command. phaseglass remains available and
calls the same implementation. Python imports/distribution, data schemas and
portable replay formats keep their original names.

The solver and simulation models are unchanged. Public recorded calculations
remain the original 0.2.0 outputs. Existing tags, historical assets and original
Phaseglass browser replay are preserved. No new hardware or sensing capability,
scientific accuracy, security advisory or outside adoption is claimed.

Fresh install checks exercise both console commands from the new wheel outside
the source checkout and verify the packaged branded Studio resources.

Phaseglass 0.2 · Fit here, predict there

Fit here, predict there

Phaseglass 0.2 adds a coherent receiver-sweep experiment. Fit known receiver positions while profiling an unknown complex response at every antenna pair and frequency. Then predict entire held configurations without refitting gains, locations or selecting modes from their IQ.

The studio has nine original simulation cases, editable strict training/held packets, orbitable geometry, retained mirrors, actual complex IQ and noise-scaled residual plots, cancellation, stale-result guards and optional separate context. The portable ZIP contains exact solver source, a tested dependency lock, separate packets, fresh results, checksums and an outside-folder runner.

Try the static recorded demonstration at https://nazeeh111.github.io/Phaseglass/demo/. This replay uses frozen calculations; the local studio performs new inference.

A useful result and its boundary

Known coherent motion corrects the stable phase-slope example that fits a point over three metres wrong in the single-snapshot baseline. Coplanar motion retains both reflected explanations. Response changes can fail training or held prediction.

An intentional unrestricted changing-response example also passes both tests while over three metres wrong. Stable response is a trusted assumption, not something the residuals can certify. All evidence is synthetic under an idealized single-point model with exactly supplied geometry. There is no measured CSI, room/body scanner, hardware accuracy, calibrated confidence or global uniqueness claim.

Installation and replay

Python 3.12+; NumPy and SciPy are the only runtime dependencies. Install the wheel or repository, run phaseglass studio, then open http://127.0.0.1:6365/sweep.html. Existing snapshot commands remain available at the root studio. Version 0.1 remains downloadable for Python 3.10.

Extract a sweep ZIP into a fresh folder, install requirements.lock, then run python rerun.py --output NEW-result.json from any working directory. Existing output files are refused. Checksums detect changes against their manifest; they are not authenticity signatures. Context is not inference input.

Source, original fixtures, licenses and methods are included. Developed with Codex assistance.

Phaseglass 0.1.0 — One signal. More than one place.

Phaseglass makes radio ambiguity visible, then lets you design a measurement to distinguish competing explanations.

• Original 3D Studio with editable antennas/point/frequency/noise, actual complex IQ plots, finite search field and separate optional generator truth.
• Blind inverse retains discovered modes, underconstrained families, weak/noisy holds and model-mismatch/search-failure refusal.
• Receiver-placement experiment profiles unknown constant link gains, changes one receiver and regenerates new IQ before solving.
• Twelve reproducible simulations, including the mirror trap and a good-fit/wrong-location phase-slope counterexample.
• CLI and portable source-bearing experiment ZIP. Export recomputes results from measurements; reruns do not require generator truth.

Try it

Install the wheel with Python 3.10+ (python -m pip install phaseglass-0.1.0-py3-none-any.whl), then run phaseglass studio. Open http://127.0.0.1:6365/.

Or extract phaseglass-mirror-experiment.zip into a fresh directory, install its requirements.txt, and run python rerun.py --output independently-rerun.json.

Verification

17 Python tests and editor input-event controls pass on all three hosted gates: Linux Python3.10, Linux Python3.14 and macOS Python3.14. Each also replays twelve blind experiments and verifies a fresh installed wheel outside the checkout. Exact-source CI.

Independent scalar math and a fresh review check the forward model, profiled residuals, coarse field, pair intervention and deliberate model counterexamples. The actual browser journey verifies two modes, receiver intervention, new IQ/one found mode, separate truth evaluation, invalid input preservation, imported measurement without truth, portable download/rerun and responsive layout.

Scope

Simulation only: one ideal isotropic point scatterer with baseline-isolated channel. No WiFi hardware capture, human pose, room reconstruction or empirical accuracy claim. Finite search and heuristic gates do not prove global uniqueness or statistical coverage. Unknown hardware phase slopes may fit well at a wrong position.

Original application MIT, NumPy/SciPy installed separately under their own licenses. Codex assistance disclosed. SHA256SUMS covers the attached assets; it is an integrity manifest, not an authenticity signature.