| name | scientific-analysis |
| description | Run a manuscript-supporting scientific investigation that ends in a figure + an ana_log — pose a question, pull ensemble/run data, compute the statistic/mechanism, make a figure, cite evidence, and write an ana_log. Use when the user asks to "investigate whether X", "is X correlated with Y", "analyze the mechanism of X", "make a manuscript figure for X", "P-pool / cross-regime / attribution analysis". For a standardized single-round report use summarize-calibration-round; for cross-round figures use compare-calibration-rounds. |
| modes | {"requires_fates":false,"nutrient_pathway":"any","scope":["analysis"],"summary":"Investigation -> figure -> ana_log workflow; model-agnostic (examples are FATES/Kougarok)."} |
Scientific Analysis → Figure → ana_log
The interactive agent does the open-ended, manuscript-supporting science that the fixed
calibration loop can't: "is PFT10 P uptake growth-limited or stoichiometric?", "attribute
the R2→R4 change", cross-regime / P-pool investigations. This skill codifies that
workflow so the result is reproducible and properly recorded.
Use this for free-form investigation. For the canned single-round report use
summarize-calibration-round; for cross-round comparison figures use
compare-calibration-rounds. Those are the standardized deliverables; this is the
exploratory one.
Step 1 — pose the question precisely + scope the data
State the question as something falsifiable ("P uptake tracks NPP, not leaf stoich"), and
identify exactly what data answers it: which round/ensemble, which variables, which
validation targets, which cases (top-N? all? a contrast pair?). If the question touches a
FATES mechanism, check RAG/GraphRAG / docs/fates-knowledge-base/ first — don't assert mechanism
from names.
Step 2 — pull + analyze
Reuse the analysis tooling rather than re-deriving:
use_cases/<site>/analysis/ scripts (attribution, contrast, forcing comparison) and
tools/plot_ensemble_cases.py / tools/extract_ADSP_RGSP_slim.py for ensemble data.- Compute the actual statistic (correlation r, attribution, regression) — quote the
number, don't hand-wave. If a tiny/odd subset is involved, run
diagnose-forensics
triage first (is it real or an artifact?).
Step 3 — make the figure (filename convention)
Save figures under use_cases/<site>/analysis/ with the round + axis-mode + case
count embedded in the name (memory: plot filename convention) — e.g.
R4_combined_519yr_top50_<topic>.png — to avoid the ensemble_biomass_all_cases.png
ambiguity from past sessions. Figures are gitignored, so the filename is the only durable
pointer.
Step 4 — write the ana_log (cite explicit evidence)
Write an ana_log via the /log skill (/log ana <topic> → memory/ana_logs/). The
load-bearing rule: every quantitative claim names its figure / statistic / data file
inline, and a section drawing on several artifacts gets an "Artifacts this section is
based on" table. A claim with no cited source is a red flag — find the source or soften
it. Render to PDF with the markdown-to-pdf skill if it's a shareable note.
Step 5 — land the lesson (optional)
If the analysis yields a vetted, generalizable discovery, add it to the curated KB — but
Tier-3 is interactive-only, so author it deliberately (an interactive-mode
MemoryManager / by hand) or stage + promote via curate-knowledge. Do NOT inject an
unverified hunch.
Notes
- ana_logs are manuscript working notes — tracked on this branch, excluded from public
sync.
- Pairs with:
diagnose-forensics (triage odd results), compare-calibration-rounds /
summarize-calibration-round (standardized figures), /log (the ana_log + supersede
protocol), markdown-to-pdf (render), curate-knowledge (land a lesson).
Changelog
- 2026-06-17:
## Changelog convention adopted (see .claude/skills/README.md). Earlier history: git log + memory/dev_logs/.
