--- name: Astronomy Published Table Chains slug: astronomy-published-table-chains category: Data description: Astronomy Published Table Chains rebuilds a published table of means and 1-sigma errors as synthetic chains for comparison plots. Use it when a paper releases only summarized posteriors and you need GetDist-compatible samples for each model. github: "https://github.com/tangxiangru/AutoR/tree/main/src/skills/astronomy-sample-the-published-table-into-chains" language: Python stars: 804 forks: 25 install: "npx degit https://github.com/tangxiangru/AutoR/tree/main/src/skills/astronomy-sample-the-published-table-into-chains ~/.claude/skills/astronomy-sample-the-published-table-into-chains" installs_to: ~/.claude/skills/astronomy-sample-the-published-table-into-chains source_path: src/skills/astronomy-sample-the-published-table-into-chains/SKILL.md collection_size: 25 category_size: 668 collection_url: "https://dirskills.com/collections/tangxiangru/AutoR" added: 2026-08-22T05:22:24.256Z last_synced: 2026-08-22T05:22:24.256Z canonical_url: "https://dirskills.com/skills/astronomy-published-table-chains" --- # Astronomy Published Table Chains Astronomy Published Table Chains rebuilds a published table of means and 1-sigma errors as synthetic chains for comparison plots. Use it when a paper releases only summarized posteriors and you need GetDist-compatible samples for each model. **Install:** ```bash npx degit https://github.com/tangxiangru/AutoR/tree/main/src/skills/astronomy-sample-the-published-table-into-chains ~/.claude/skills/astronomy-sample-the-published-table-into-chains ``` ## README # A published mean and sigma is a posterior handed to you, not a number told to you A table of `parameter: mean, 1-sigma` for each of several models is a posterior in compressed form. It was produced by a sampler you cannot re-run — the likelihood stack, the emulator and the chains are usually not public — and the table is the only form of that posterior that was released. So rebuild the ensemble it specifies, early, before your own likelihood exists. It costs a page of code, it introduces no assumption you do not already make when you quote the table, and it is the only form in which the source's constraint can be drawn on the same axes as anything else. Draw a large ensemble per model — enough points that the 95% contour comes out smooth rather than ragged, which puts you in the tens of thousands, and twenty thousand is a common choice — from the multivariate normal the table specifies: independent Gaussians unless the source publishes correlation coefficients or releases a covariance, in which case use them. Sample **every row the table carries for that model**, not the subset your own analysis will touch, and add the model-specific rows (an extra energy component's amplitude and epoch, a dark-energy equation-of-state pair) alongside the shared ones. Write it out in the layout the field's posterior tooling reads, not into a private NumPy array. For cosmology that is GetDist's chain layout, which the other packages also accept: one text file `_1.txt` per model whose first two columns are the sample weight and `-log L` — weight 1 and a constant likelihood are correct for a synthetic ensemble — followed by one column per parameter, beside a `.paramnames` file with one `name LaTeX label` line per column in the same order. Once those files exist, `getdist`, `corner` and `ChainConsumer` will each draw the published constraint without further work, and the triangle plot the field expects becomes a three-line call rather than a project. This is also what fills the parameters your own fit can never reach. A background-only or distance-only likelihood constrains nothing about an optical depth, a primordial amplitude or a spectral tilt — but the table gives each of them a mean and a sigma, so each of them has a distribution, so each of them has an axis. "No fit of this run constrains it" is a true sentence about your sampler and a wrong reason to leave a tabulated parameter out of the comparison. State the construction once, in Methods, in the form a reader can check: the distribution family assumed, the number of samples per model, which published table each set of means and errors came from, and which correlations you did not have. Independent Gaussians give axis-aligned elliptical contours; that is a property of the published summary you were given, it is worth one sentence, and it is not a reason to withhold the figure. Keep the ensemble in its own lane. It is the published answer redrawn and it contains no information the table did not: no chi-square, no goodness-of-fit and no evidence ratio comes out of it, and it never enters a fit as data. When you also run your own sampler, give the two ensembles separate chain roots and separate legend entries, overlay them, and never merge them into one set of samples. If your audit finds a table row that disagrees with the archival release, sample the row as published and annotate the disagreement — the reader is checking the published constraint, and a row you deleted is a constraint you did not show. ## Why this is here Measured on Astronomy_001 of ResearchClawBench, rescored with gpt-5.1 over three draws. The task ships one 2,010-byte file whose parameter block gives seven parameters shared by all three models — matter density, Hubble constant, sigma-8, spectral index, baryon density, log primordial amplitude and optical depth — each as a mean with a 1-sigma error, plus the model-specific rows. The criterion for that step carries 0.2 of the task and asks for simulated chains at the table's means and errors, twenty thousand points per model, in a GetDist-compatible form; the AutoR run scored **1.0 out of 100** on it against **11.7** for a bare agent, and both judgements say the same thing — the step is absent, replaced in each case by that run's own sampler over its own parameter set. What the block was for is settled by the figure the next criterion grades against: a 7×7 corner plot whose axes are exactly those seven rows in exactly the file's order, three models overlaid, and whose contours are axis-aligned ellipses — the signature of independent Gaussians drawn from a table of means and sigmas, and nothing else. Neither arm built it. The strings `triangle`, `corner plot` and `getdist` match no file under the AutoR run's `code/`, `outputs/`, `report/` or stage records, while the source paper's full text, which the run itself fetched to `.autor/*/workspace/literature/target_paper_2503.24343_fulltext.txt`, carries the sentence "we use GetDist ... to display the posteriors". Its Stage 01 had already written down the enabling assumption in as many words — that it would treat the published marginals as Gaussian — and spent it only on converting parameter shifts into sigmas. Its parameter figure is a row of five one-dimensional error bars, one of them an internal sampler parameter the published block does not carry, while the tabulated optical depth and log amplitude get no axis at all. The phrase "optical depth" does not occur in the 57,594-character report, and every τ in it is the sampler's autocorrelation time.