report

Assemble the session's analysis into a report a colleague can trust and rerun. Works by slash command or conversationally ("write this up"). The report renders what was actually computed this session; it never invents numbers, methods, or citations.

Knowledge coupling (standing rule)

This skill is a procedure, not a store of dataset facts. Every dataset detail the report states (product version, verification date, expected range, uncertainty framing, release caveat, convention, or gotcha) is read from a concept, per the consult-knowledge skill, restated in the report's own words, and cited by bundle path with its status. The report carries none of these facts inline; when a stated fact has no owning concept, that gap is itself a Quality or Provenance note.

The confirmation gate (hard gate, before anything is written)

Before writing any file, present:

• the proposed filename (and format),
• the section list with one line on what each will contain,

and wait for explicit confirmation. The gate lives here in the skill body so it fires on every surface, conversational included. On surfaces where file writing is unavailable, the same gate confirms scope, then the report renders in-session with a note saying so.

Format: markdown by default; docx only when asked.

The six sections

1. Data Description. Each dataset used: product and version (with verification date when it came from a dataset concept), period, region, resolution, access path.
2. Methods. What was computed and how: weighting, baselines, trend method and its autocorrelation handling, calendar choices, block lengths and seeds for resampling methods. Specific enough to reimplement.
3. Results. The findings, and the house rule is enforced here (hard refusal): every headline quantity carries an uncertainty statement (interval, spread, or native product uncertainty) with method and level, or an explicit one-line reason why none is available. A result that arrives without uncertainty goes back for one (or gets its waiver) before the report is assembled.
4. Quality Notes. QC findings on the inputs: what was flagged, masked, or gap-ridden, and what that means for the conclusions; "QC not run" is itself a note when true.
5. Provenance. The knowledge concepts consulted, cited by bundle path (for example core/knowledge/conventions/calendars.md, or a domain bundle's dataset and gotcha concepts), with concept status noted when it is anything other than verified. Connector and search paths used (Earthdata MCP vs knowledge fallback) are recorded here too.
6. Reproducibility. Package versions that produced the results, the environment file when one exists, random seeds, input file identity (names, versions, DOIs), and the provenance history line written into any saved outputs (per the reproducibility skill).

Rules

• Content comes from the session's actual computations and their outputs; a section with nothing real to say says so briefly rather than padding.
• Figures included in the report follow cartography's rules (defined captions, uncertainty visualization where results carry intervals).
• Saved report files carry CF-style provenance where applicable: at minimum a generation timestamp, tool identity, and input list in the document header.

Must NOT (hard refusals: invariant, universal, gate-shaped)

• Never write a file before the gate is confirmed.
• Never assemble a Results section containing a headline quantity with neither an uncertainty statement nor a waiver.
• Never cite a concept that was not actually consulted, and never consult one without citing it.
• Never fabricate a Methods detail that cannot be traced to what ran; a gap in the record is reported as a gap.