sim-paper-reproduction
Use this skill when the user wants to reproduce, validate, or diagnose a
simulation from a paper, report, thesis, benchmark, or reference study.
This is a task skill, not a solver driver. Pair it with the relevant solver
skill, such as HFSS, COMSOL, Fluent, MATLAB, or a plain Python library. The
execution tool is not fixed: use sim-cli when it adds discovery, session
control, inspection, logs, or artifact tracking; use solver-native scripts,
plain Python, vendor batch executables, or GUI workflows when they are the
right primitive.
Paper Evidence Map
Before turning a paper into model inputs, build a compact evidence map for the
parts that affect the task. For each fact, record the page or section, source
object (text, equation, table, figure, caption, or supplement), extracted value
and unit, and status: direct, derived, inferred, ambiguous, or missing.
- Keep figure axes, legends, series identity, normalization, and coordinate
conventions with any digitized values.
- Preserve conflicting text/table/figure values as separate evidence until the
conflict is resolved; do not silently choose one.
- Treat OCR and visual readings as candidates until checked against text,
captions, equations, or another source.
- Carry only
direct or explicitly justified derived facts into the
parameter manifest. Keep inferred values visible as assumptions.
Reproduction Loop
- Define the acceptance target before building.
- Capture the paper/result identifier, target figures/tables, units,
coordinate conventions, and numeric pass/fail metrics.
- A solver exit code, generated plot, or successful export is not acceptance.
- Build a parameter manifest.
- For each value, record
value, unit, source, confidence, and notes.
- Mark inferred or exploratory values clearly. Do not mix them with
paper/user-provided values.
- If a value is missing, leave it missing or run a labelled sensitivity
study; do not invent it as fact.
- Create the smallest credible runnable case.
- First prove the solver can build, solve, and export the leading metric.
- Use simplifications only when they are explicitly labelled and tracked.
- Prefer one bounded step at a time so each failure has a clear cause.
- When the full reference target is large, coupled, or under-specified, use
progressive fidelity: add geometry detail, secondary physics, coupling,
sweeps, optimization, or calibration only after the prior layer has
exported evidence and passed its acceptance check, unless a bypass reason
is recorded.
- Gate later results on earlier evidence.
- If the leading metric fails, stop and diagnose before producing secondary
figures.
- The gate depends on the task. For an antenna it may be S-parameters before
axial ratio and patterns; for another domain it may be residuals,
conservation, calibration error, or a benchmark norm.
- Record each iteration.
- Keep
hypothesis -> change -> evidence -> conclusion -> next hypothesis.
- Preserve logs, exported data, solver messages, screenshots/previews, and a
run manifest.
Tool Choice
Use the narrowest reliable execution primitive:
sim-cli persistent sessions when inspection, bounded snippets, or shared
session state are useful.
sim run when the plugin wrapper adds linting, parsing, profile checks, or
safer execution.
- Plain Python or solver SDK scripts when they are clearer and still produce
durable artifacts.
- Vendor executables or batch modes when they are the canonical way to solve or
export the reference case.
- GUI workflows only when the solver cannot expose the required state
programmatically, and then capture visual evidence after each GUI action.
Whatever path is chosen, keep the same evidence standard: version/install
probe, inputs, bounded execution, logs/messages, exported artifacts, and metric
evaluation.
Deliverables
For a first pass, produce:
- Parameter manifest with sources and confidence.
- Reproduction record describing the model, simplifications, and setup.
- Exported leading metric data and computed acceptance numbers.
- Solver artifacts needed to rerun or inspect the case.
- Failure diagnosis and next hypotheses when the result does not match.
Do not claim a reproduction is complete until the stated metrics pass. It is
valid and useful to deliver a failed first pass if it has clear evidence.
Publication Handoff
When the user wants a paper, report, or manuscript structure from the work,
build a claim ledger before drafting results or conclusions. For each proposed
claim, record its status (verified, derived, planned, or speculative),
the run and artifact that support it, and the strongest wording the evidence
allows. Planned runs and attractive hypotheses are not results.
For content-preserving manuscript edits, state the preservation contract before
changing the document: which scientific facts, numeric values, citations,
figure pixels, and wording must remain unchanged, and whether reordering,
headings, captions, cross-references, or formatting may change. Verify the
finished artifact against that contract as well as its rendered layout.
Boundaries
- This skill does not provide PDF/OCR extraction policy.
- Do not encode a paper's geometry, parameters, figure numbers, or expected
curves here. Keep those in the case artifact or dataset.
- Solver-specific launch, geometry, meshing, boundary, and export details belong
in the solver skill or the solver connector.
1---2name: sim-paper-reproduction3description: Cross-solver discipline for extracting simulation evidence from papers, reproducing reference results, and turning verified solver evidence into manuscript or report claims. Use with the relevant solver skill; this skill covers task framing, provenance, staged acceptance, and publication handoff rather than solver-specific commands.4---56# sim-paper-reproduction78Use this skill when the user wants to reproduce, validate, or diagnose a9simulation from a paper, report, thesis, benchmark, or reference study.1011This is a **task skill**, not a solver driver. Pair it with the relevant solver12skill, such as HFSS, COMSOL, Fluent, MATLAB, or a plain Python library. The13execution tool is not fixed: use `sim-cli` when it adds discovery, session14control, inspection, logs, or artifact tracking; use solver-native scripts,15plain Python, vendor batch executables, or GUI workflows when they are the16right primitive.1718## Paper Evidence Map1920Before turning a paper into model inputs, build a compact evidence map for the21parts that affect the task. For each fact, record the page or section, source22object (text, equation, table, figure, caption, or supplement), extracted value23and unit, and status: `direct`, `derived`, `inferred`, `ambiguous`, or `missing`.2425- Keep figure axes, legends, series identity, normalization, and coordinate26 conventions with any digitized values.27- Preserve conflicting text/table/figure values as separate evidence until the28 conflict is resolved; do not silently choose one.29- Treat OCR and visual readings as candidates until checked against text,30 captions, equations, or another source.31- Carry only `direct` or explicitly justified `derived` facts into the32 parameter manifest. Keep inferred values visible as assumptions.3334## Reproduction Loop35361. Define the acceptance target before building.37 - Capture the paper/result identifier, target figures/tables, units,38 coordinate conventions, and numeric pass/fail metrics.39 - A solver exit code, generated plot, or successful export is not acceptance.402. Build a parameter manifest.41 - For each value, record `value`, `unit`, `source`, `confidence`, and notes.42 - Mark inferred or exploratory values clearly. Do not mix them with43 paper/user-provided values.44 - If a value is missing, leave it missing or run a labelled sensitivity45 study; do not invent it as fact.463. Create the smallest credible runnable case.47 - First prove the solver can build, solve, and export the leading metric.48 - Use simplifications only when they are explicitly labelled and tracked.49 - Prefer one bounded step at a time so each failure has a clear cause.50 - When the full reference target is large, coupled, or under-specified, use51 progressive fidelity: add geometry detail, secondary physics, coupling,52 sweeps, optimization, or calibration only after the prior layer has53 exported evidence and passed its acceptance check, unless a bypass reason54 is recorded.554. Gate later results on earlier evidence.56 - If the leading metric fails, stop and diagnose before producing secondary57 figures.58 - The gate depends on the task. For an antenna it may be S-parameters before59 axial ratio and patterns; for another domain it may be residuals,60 conservation, calibration error, or a benchmark norm.615. Record each iteration.62 - Keep `hypothesis -> change -> evidence -> conclusion -> next hypothesis`.63 - Preserve logs, exported data, solver messages, screenshots/previews, and a64 run manifest.6566## Tool Choice6768Use the narrowest reliable execution primitive:6970- `sim-cli` persistent sessions when inspection, bounded snippets, or shared71 session state are useful.72- `sim run` when the plugin wrapper adds linting, parsing, profile checks, or73 safer execution.74- Plain Python or solver SDK scripts when they are clearer and still produce75 durable artifacts.76- Vendor executables or batch modes when they are the canonical way to solve or77 export the reference case.78- GUI workflows only when the solver cannot expose the required state79 programmatically, and then capture visual evidence after each GUI action.8081Whatever path is chosen, keep the same evidence standard: version/install82probe, inputs, bounded execution, logs/messages, exported artifacts, and metric83evaluation.8485## Deliverables8687For a first pass, produce:8889- Parameter manifest with sources and confidence.90- Reproduction record describing the model, simplifications, and setup.91- Exported leading metric data and computed acceptance numbers.92- Solver artifacts needed to rerun or inspect the case.93- Failure diagnosis and next hypotheses when the result does not match.9495Do not claim a reproduction is complete until the stated metrics pass. It is96valid and useful to deliver a failed first pass if it has clear evidence.9798## Publication Handoff99100When the user wants a paper, report, or manuscript structure from the work,101build a claim ledger before drafting results or conclusions. For each proposed102claim, record its status (`verified`, `derived`, `planned`, or `speculative`),103the run and artifact that support it, and the strongest wording the evidence104allows. Planned runs and attractive hypotheses are not results.105106For content-preserving manuscript edits, state the preservation contract before107changing the document: which scientific facts, numeric values, citations,108figure pixels, and wording must remain unchanged, and whether reordering,109headings, captions, cross-references, or formatting may change. Verify the110finished artifact against that contract as well as its rendered layout.111112## Boundaries113114- This skill does not provide PDF/OCR extraction policy.115- Do not encode a paper's geometry, parameters, figure numbers, or expected116 curves here. Keep those in the case artifact or dataset.117- Solver-specific launch, geometry, meshing, boundary, and export details belong118 in the solver skill or the solver connector.