Run a scientific experiment, inspect its assumptions, and leave a useful account for the next person.
Reproducible is the experiment workspace in a three-project workflow:
| Project | Responsibility |
|---|---|
| Prospector | Find and review papers; hand off a versioned candidate investigation. |
| Irreducible | Implement and test the physics, numerical calculations and scientific interfaces. |
| Reproducible | Design the selected experiment, run the compiled calculations, visualize outputs and keep a compact experimental history. |
New physics belongs in Irreducible. Questions, data lineage, parameter choices, fitting orchestration and plots for a particular investigation belong here. A reviewed idea may remain blocked if its model, data or likelihood is missing.
The roadmap is the sole active Reproducible programme, coordinated with Prospector and Irreducible. Current gaps records actual experiment, dataset, likelihood and inference readiness. Older campaign and checkpoint priorities are superseded; their findings and immutable pins remain.
An experiment normally has four files under experiments/<id>/: a Markdown
account, experiment.json, an Irreducible request.json, and an optional plot
script. A Prospector-origin experiment also retains the small candidate JSON.
The account explains the claim, simplifications, checks, findings and limitations;
Git records its changes. We do not publish a file for every rerun.
S3 holds persistent inputs, full engine receipts, tables, chains, simulations
and rendered figures. Ignored data/, results/ and simulations/ are temporary
scratch/cache restored by exact manifest pins. The public
packet identifies the evidence; a published scientific result should cite a
separate durable archive containing its exact inputs and full run records.
Use named shared storage for useful inputs, complete or
failed attempts and unfinished handoffs. Publish and pin before ending work;
push --evict-local removes unchanged scratch only after S3 byte verification.
Historical snapshots retain their original recovery route.
Python 3.11 or newer and uv are needed for this workspace. Build Irreducible separately using its build guide, at the commit named in the example packet. From this repository root:
uv sync --frozen
uv run python scripts/packet.py
uv run python scripts/run.py expansion-background --irred ../irreducible/target/debug/irred --name example
uv venv .work/plots
uv pip install --python .work/plots/bin/python --require-hashes --only-binary=:all: -r requirements/plots.txt
.work/plots/bin/python experiments/expansion-background/plot.py results/expansion-background/examplePlotting uses a separate optional environment. Its requirements retain the prior plot versions and artifact hashes; the core lock contains only runner dependencies. Keep the actual plotting environment identity with any published figure. Existing scientific runtimes and receipts are separate.
The background example evaluates
chosen flat ΛCDM parameters at eight redshifts. It uses no observed data and fits
nothing. The runner verifies the packet, input identities and clean engine source
revision, then calls the current irred run interface. It records the actual
build, executable, request and receipts. Each destination must be new; failed
attempts remain visible and must be published to their named S3 collection
before a cross-run handoff.
This first runner handles one compiled request per packet. Irreducible's proposed multi-step recipe runner remains proposed. More involved experiments can add a small, reviewed orchestration script using existing engine interfaces; keep shared physical equations and numerical kernels in Irreducible.
The observational cosmology experiment adds a fresh-checkout acquisition/build recipe and a bounded conditional H₀ fit to the exact DESI DR2 full13 Gaussian compression, for pinned CLASS ΛCDM and a declared constant-w alternative. It retains full covariance, predictions, native scores, plots and failure manifests. Optional official Planck primary scores remain separate from the DESI fit; the numerical search bracket is not a posterior uncertainty.
The LCDM baseline audit uses the installed Irreducible C++ SDK on exact ordered 13-row DESI DR2 Gaussian BAO compression. Its bounded experiment-specific controller compares distances, ratios and the normalized density with independent Decimal quadrature/LDLT references and retains failed attempts. It is a chosen supplied-drag massless-radiation control; full Planck 2018 massive-neutrino/thermal-drag/CMB reproduction remains blocked. Numerical success establishes neither a fit nor a posterior result. Follow the packet's native command and exact source/SDK/data pins.
The current reproduction campaign retains all nineteen source and native contracts as a historical diagnostic snapshot informing the current roadmap. Its separate thermal DESI consumer tests chosen massless and massive FD models with supplied drag against the same full 13-row compression. Full Planck reproduction and observational qualification stay blocked; older packets and their SDK identities remain intact.
Subsequent engine work has passed a released-ladder matrix comparison and added an explicit thermal-relic E/H provider and conditional estimator variance. The baseline account separates that progress from this packet's original massless model and build. A newer engine does not automatically replace a pinned SDK or change an earlier run's scientific identity.
The conditional SN working profile compares the native common-M profile with an independent full-mode spectral reference for a separately declared symmetric matrix and all 1657 selected SN+SH0ES occurrences. The original nonsymmetric covariance remains refused; numerical agreement does not qualify the released likelihood or an H0 posterior.
The released ladder experiment now executes the unchanged 3492-row, 47-column compact SH0ES design through the installed native library. It verifies all 37 host columns against the pinned primary Cepheid table and compares the full fit, H0-related contrast variance and declared synthetic constraint-row responses with independent QR/SVD. The source MCMC's fixed coordinate, unresolved nuisance identities and paper/release selection discrepancy remain explicit. Its narrower controller runs while the full paper-reproduction packet stays blocked; it supplies no observational posterior or reconstructed systematic H0 uncertainty.
The released finite-box control retains all 3492 rows and fixes only original 44 at zero outside active 46-coordinate product measure. Its separate request compares three log normalizations and the original46 median through a pinned native SDK and independent original-input reference. Numerical comparison, conditional enclosures and observational qualification remain separate; the original ladder packets and SDKs remain intact.
The released SDSS lineage reader retains 321 original source joins and one 115-record photometry slice, including coordinate discrepancies and signed fluxes. Its dedicated route verifies source serialization only. Physical units, event/frame/reduction, calibration and dependence remain unresolved; the experiment's scientific execution stays blocked.
Keep visualization source next to its experiment and render into the run folder.
Start with a plain Python script. Use a Jupytext percent-format .py source when
cells improve exploration; keep executed .ipynb copies and their outputs in
ignored notebooks/. Experiment conventions explain the
handoff and notebook workflow. No notebook service or automatic paper importer
is installed.
The old supernova/cosmology code and generated results were removed from the active tree during the October 2026 reset. The historical packet retains a short account, qualification limits and the exact previous Git snapshot. It guides future experiments; the new example does not reproduce that campaign.
Storage and restoration describe the retained input identities and ignored local archives. A fresh clone has source manifests, rather than the bulk input bundle. Deleting current files has not reduced the remote Git history; use a shallow clone when you only need the new workspace.
Use a branch and a reviewable PR, including for agent work. See CONTRIBUTING.md, AGENTS.md and the development workflow. Lightweight CI checks packet integrity, storage rules and runner behavior; it does not qualify a scientific result. BSD 3-Clause covers original repository code and documentation; third-party inputs retain their own terms.
The SN observer and historical passband control checks 321 supplied released redshift pairs under chosen radiation-free LambdaCDM and analytic coasting backgrounds, plus a pinned optical filter with an explicitly synthetic source. Numerical agreement is conditional; unresolved event/frame and joint-calibration information remains explicit.
Shared datasets and private evidence use the common storage layout.
New local or cloud chats use storage startup discovery to find the configured bucket and verified catalog without relying on chat history.