Modelling Paper: From Model Results to a Full Draft

What it does

A modelling paper is judged on its mathematics: is the model written out in full, is every parameter sourced, do the derived results hold, and does the text claim more than the model shows? This skill makes your agent a co-author in two roles. As PI it owns the story, pins the one-sentence claim before anything is drafted, and asks "what will Reviewer 2 say?" before anyone else does. As postdoc it writes the full prose and the complete Methods mathematics, then checks it.

• Starts from what you have. It reads your existing files first: model code, results, outline, partial draft. The Methods describe what your code actually solves, and any difference between the code and your notes is reported.
• The full Methods mathematics. The complete model, every symbol defined, initial and boundary conditions, scaling, derived results (steady states, stability conditions, thresholds, distributions) with enough derivation to check, and the numerical settings. A notation table gives each symbol's meaning, units, value or range, and source.
• Checked, not asserted. Dimensions, limiting cases, the code against the equations, and a numerical check of every analytic result when Python is available. Every number in the prose comes from a computation. Claims in your notes that the model cannot support are flagged before they are written up.
• Any model class, any field. A per-class checklist covers deterministic ODEs, DDEs and PDEs, master equations and SDEs, agent-based models, statistical and inference models, networks, optimisation and control, numerical-methods papers and surrogates. Domain packs add the vocabulary, conventions and field-specific objections for biology, epidemiology, ecology, and physics and engineering.
• Reviewer objections, rated and planned. A running list in your notes: each objection specific to your paper, rated fatal, major or minor, with the planned answer, drawn from a general catalogue (model construction, parameters, identifiability, fitting, numerics, robustness, scope of claims, novelty) plus the domain pack.
• No invented references. A citation is either verified (seen in a search result, supplied by you, or read) or marked. A parameter with no known source gets a [CITE] marker, never a plausible-looking citation.

In the demo

Given a synthetic two-node thermal model of a battery cell, its code, four result files and a rough outline, the agent wrote the Methods with a closed-form runaway threshold that matched the user's bisection to 3e-6 K. It proved that the Hopf bifurcation the outline hoped for cannot occur (the system is cooperative, so its eigenvalues are always real). It narrowed "safe below the threshold" to a local result and caught that the outline said 2.5 A while the code used 3.0 A. It left the bibliography empty because no sources had been given, and listed eleven reviewer objections.

What's included

• SKILL.md: the two roles, session start, project layout, Methods rules, figure planning, citations, claims, reviewer objections, strategic questions, export
• references/methods-maths.md: what every model needs, what each model class adds, the verification pass, a Methods skeleton
• references/reviewer-objections.md: the general objection catalogue
• references/literature-and-citations.md: where to search, when a reference counts as verified, how to mark the rest
• references/domains/: biology, epidemiology, ecology, physics and engineering
• scripts/scaffold_paper.py (creates the paper skeleton, never overwrites) and scripts/check_refs.py (citation audit)

Part of the Aurora Lab research pipeline

Works entirely on its own. With a research/ folder it keeps the manuscript in research/paper/ and its working notes beside it. Pairs well with Aurora Lab's Claim Check, Scientific Voice and Journal House Style, but needs none of them.