| name | physics-research |
| description | Use when the work is a physical measurement, a fit, an error budget, a detector or instrument analysis, or a simulation that has to agree with known physics. Covers dimensional checks, CODATA constants, Type A/B uncertainty and coverage factors, chi-square and covariance, systematics and blind analysis, HDF5/ROOT/FITS data and the arXiv/Zenodo/HEPData archives.
|
| version | 1.0.0 |
| author | Remedy |
| tags | ["research","physics","metrology","uncertainty","fitting","simulation"] |
| requires | [] |
| tools | ["analysis_env","analysis_run","analysis_ledger","stats_assumptions","stats_effect_size","manuscript_build","lit_search","lit_fetch","cite_add","cite_check","skill_activate","file_read","file_write","repo_search"] |
| triggers | ["\\b(systematic uncertaint\\w+|statistical uncertaint\\w+|error budget|error propagation|dimensional analysis|units check)\\b","\\b(Monte[- ]Carlo (?:event|generator)|GEANT4?|ROOT (?:file|tree|macro)|detector (?:calibration|response|acceptance))\\b","\\b(blind(?:ed)? analysis|look[- ]elsewhere effect|five sigma|5\\s?sigma|goodness[- ]of[- ]fit|chi[- ]?squared per (?:dof|degree))\\b","\\b(hep-(?:ex|ph|th|lat)|cond-mat|astro-ph|gr-qc|quant-ph|nucl-(?:ex|th))\\b"] |
Physics research
Activate research-method first: skill_activate(skill="research-method").
That pack owns question framing, evidence standards, preregistration,
citation honesty and how to say "we do not know". Do not restate it. What
follows is only what physics does differently.
Order of operations
- Dimensions before numbers. Reduce every term of the expression to base
SI dimensions (M, L, T, I, temperature, N, J). Every summand must match,
and the result must match the quantity claimed. A dimensional mismatch is
findable in thirty seconds; a missing 2*pi is not, so also evaluate one
limiting case whose answer is already known.
- Constants from CODATA, never from memory. Take values and standard
uncertainties from the NIST Fundamental Physical Constants dataset and
record which CODATA adjustment year you used. Since the 2019 SI
redefinition c, h, e, k_B and N_A are exact by definition; G and particle
masses are measured and carry uncertainties. Never type a constant from
recall into a result the owner will publish.
- Uncertainty model before you fit. Per input, decide Type A (evaluated
from repeated observations) or Type B (calibration certificate, datasheet
tolerance, digitiser resolution, a published value). Write the list down
before touching the data — see
references/uncertainty-propagation.md.
- Fit, then interrogate the fit (below).
- Systematics last and hardest —
references/systematics-and-blinding.md.
- Reproduce. Every number that reaches the paper comes out of
analysis_run, so the ledger holds argv, input hashes and artifact
hashes. analysis_ledger(action="verify") is how you show a figure still
matches the data it came from.
Quoting a result
value ± u means nothing until the reader is told what the ± is. State the
standard uncertainty u (k=1) or the expanded U with its coverage factor k and
the approximate coverage probability; give statistical and systematic parts
separately; state correlations between quoted quantities. Round the
uncertainty to one or two significant figures first, then round the value to
the same decimal place — no more digits than the uncertainty supports, and no
fewer. Units on every axis, table column and abstract number.
Fitting
- Least squares is maximum likelihood only for Gaussian errors. Counting data
wants a Poisson likelihood; do not chi-square bins with a handful of
entries.
- Report chi-square per degree of freedom with the dof and the p-value.
chi2/dof near 1 says the residuals are consistent with the errors you
quoted — it does not say the model is right, and chi2/dof well below 1
usually means the errors are overstated.
