AMS-02 Detector and Physics Analysis

Guide realistic AMS-02 detector and charged-cosmic-ray analysis work while separating documented facts (public AMS sources), general methods (textbook/PDG/statistics literature), proposals (choices you recommend for this analysis), and unknowns (internal AMS detail). Scope: public-domain reasoning only. AMS Collaboration internal data, notes, calibration constants, production software, conditions databases, pass names, trigger bits, good-run rules and cuts are not available here; never invent them. This skill does not replace collaboration review, blinding, or peer review. For generic ROOT/statistics/detector questions with no AMS tie, do not use this skill; use hep-analysis (whose reference 38 is a brief AMS overview that this skill supersedes for depth).

Routing: read the minimum references

Multi-mode requests load the union. "Species file" means antimatter-and-leptons for e±/antiprotons/antinuclei, nuclei-and-isotopes for elements/isotopes, charged-cosmic-rays for p/He fluxes and ratios; add a second only if the question crosses species. A simple factual question gets a short answer and loads only the one concept reference, plus source-index whenever an AMS number, date, or status is quoted (that is the only exception to "one reference"); do not impose the full design outline on it: answer in a few sentences, without response-chain, failure-mode or validation sections unless asked. For a review request, answer verdict-first, then defects in priority order, then the inputs that would change the verdict. source-index is the claim-to-source ledger (generated from data/*.json); consult it before quoting any AMS number.

Common workflow (analytical dependency order, not prose order)

estimand → observable space → data/MC scope → reconstruction → selection → signal/background model → efficiency/acceptance/response → inference → validation → reporting

Substantial design/review answers normally cover: measurement target; relevant subsystems; reconstruction and selection; signal, backgrounds, controls; efficiency, livetime, acceptance, response; inference and systematic propagation; closure and cross-checks; sources, unknowns and required inputs. Detect and flag a proposal that starts from a preferred algorithm (BDT, Bayesian unfolding, template fit) before the estimand and response are defined. Schemas for measurement, selection, background and systematic ledgers are in efficiency-acceptance-backgrounds and calibration-mc-systematics; use them internally or show them when they clarify. Load analysis-artifacts only when a specification, formal artifact or formal review is the deliverable.

Non-negotiable invariants

1. Rigidity R = pc/(Ze) (GV) is not momentum; p = |Z|R/c. Energy, kinetic energy, kinetic energy per nucleon, and rigidity are different variables; every conversion states Z and A (or mass). Sign of R follows a stated convention; |Z| and charge sign are separate observables from separate subsystems.
2. Charge sign comes from the Tracker/magnet only. TOF, TRD, ECAL and RICH do not measure charge sign; the TRD is never a standalone charge-sign detector.
3. Acceptance, efficiency, exposure and livetime are distinct. Do not multiply conditional efficiencies as if independent, and do not correct one effect in two categories (veto, calibration, efficiency, response).
4. Charge confusion and rare misidentification are tail phenomena. Reject any estimate based on the Gaussian core of the rigidity resolution; require tail validation.
5. Ratios and fractions do not cancel acceptance/efficiency/systematics automatically; classify each nuisance as correlated, partial, or independent and verify.
6. Low counts: use Poisson/exact or toy-validated intervals; no automatic S/sqrt(B), Wilks, or Gaussian errors near boundaries or with weak constraints.
7. Never present another experiment's method as AMS practice. Label it a proposal or an analogy.
8. Never turn an unpublished candidate report (e.g. antihelium) into a discovery claim, and never supply candidate counts, cuts, bins, livetimes, acceptances or systematic sizes that a source does not give.
9. Every AMS-specific number quoted must appear in source-index with its context (species, range, period, selection); otherwise answer symbolically or qualitatively. General-method numbers (e.g. the zero-count Poisson bound -ln α, kinematics from PDG masses) are exempt but must be labeled [General method]. User-supplied numbers are labeled as such.
10. "Reasonable agreement" is not validation: state a metric and a trigger for changing the method or adding an uncertainty.

Source-verification rule

Label each AMS-specific statement (pure arithmetic and general statistics answers are labeled [General method], with the assumptions stated): [Documented] (every such tag carries the claim ID from source-index, e.g. [Documented, C35]; describe its evidence only at the claim's own verification level, never as "verified" or "full text" beyond it), [General method], [Proposal], or [Unknown/needs input]. For "latest", "current status", "has AMS published X", or any post-2025 claim: check current primary sources (INSPIRE-HEP, journal, ams02.space) when browsing is available, report publication date versus data-taking period, and check for supplements and superseding papers; when browsing is unavailable say so and do not claim currency (this skill's index was last verified 2026-09-20). Cite the claim ID from the claim's own row in source-index (not from memory or a neighbouring row) and check that its species, range, period and paper cover the statement; do not extend a claim to sibling papers or other species. Never cite search snippets, and never validate a paper you cannot locate: say it is unverified. Details: source-policy.

Response policy for missing information

Identify only the missing inputs that materially change the answer: data vs MC; species; rigidity/energy range; charge sign; measurement level (detector, object, flux); public vs internal work; target type (flux, ratio, fraction, limit, discovery). Then proceed with explicit assumptions and symbolic quantities (e.g. N_i, ε_i, A_i) instead of asking a long questionnaire. State what evidence would resolve each assumption. If the user supplies internal numbers, use them as user-supplied and label them so.

Scripts (deterministic checks)

Run from the skill directory with python3; each has --help, prints JSON, and exits 0 (ok), 1 (defects found), 2 (input unreadable or rejected). None modifies user data. Details: analysis-artifacts.

• scripts/ams_kinematics.py: [General method] conversions among rigidity, momentum, energies and T/A with explicit |Z|, A, mass; never AMS performance. State every converted value from the script's output for that exact input (do not scale by hand; p = |Z|R for the top edge as well as the bottom).
• scripts/validate_covariance.py, scripts/validate_response.py: self-consistency of supplied matrices under their declared metadata; a pass is not physical validity.
• scripts/audit_analysis_spec.py: audits a JSON or YAML analysis specification (.yaml/.yml, strict subset) into errors, warnings, proposals, unresolved inputs. When reviewing a correction chain or suspected double counting without a specification, use its fields (corrections with one effect_id per effect, response.includes, estimator.applies; exposure already contains acceptance and livetime) as the checklist and propose them in the answer.
• scripts/poisson_diagnostics.py: exact Poisson upper limit, Garwood interval, Feldman-Cousins interval, CLs limit with expected band, seeded coverage; known or marginalized background.
• scripts/statistical_toys.py: seeded toys for the boundary (q0) p-value, finite-template effect (plain and Barlow-Beeston-lite), D'Agostini iteration scan, and ratio spread (simple, across-bin correlated, or from a measured covariance); approximations, labeled.
• scripts/likelihood_limits.py: profile-likelihood limit, interval (profile Feldman-Cousins), CLs and significance, single bin, multi-bin and with shape and several nuisances (Gaussian, log-normal or gamma priors, correlations); Berger-Boos guaranteed-coverage limit and coverage check; toy calibration. scripts/unfolding_diagnostics.py: Tikhonov/SVD scans, L-curve/cross-validation choice, closure and pulls, parametric and non-parametric (several priors, Poisson-exact CV) forward fold versus unfolding, response statistics (toys, analytic multinomial covariance, efficiency systematic, measured covariance or replicas). scripts/template_fit.py: multi-template fit with the full Barlow-Beeston likelihood, weighted MC and nuisances.
• scripts/validate_evidence_ledger.py, scripts/render_source_index.py: check data/sources.json and data/claims.json and keep source-index in sync.
• scripts/fetch_papers.py: maintainer tool for data/papers_manifest.json (metadata for every AMS Collaboration article in INSPIRE-HEP, linked to ledger source IDs) and a local PDF cache outside the repository; it records blocked papers instead of bypassing them. A cached file is not a read paper: claims still need ledger entries at the level actually read. See source-policy.

Reference index