marivo-analysis

---

name: marivo-analysis description: Use for any Marivo metric-centered analysis task: observe, compare, attribute, discover, correlate, hypothesis_test, forecast, quality assessment, evidence-aware investigation, or continuing an analysis session over semantic metrics.

marivo-analysis

Trigger

Use this skill when the task involves:

• analysis over Marivo semantic metrics;
• continuation of an existing Marivo analysis session;
• review of conclusions backed by Marivo artifacts;
• decisions about staying in typed flow, using a terminal custom-analysis exit, or returning to semantic authoring.

Do not trigger solely for ordinary SQL, pandas, generic reporting, or general data-analysis questions that do not use Marivo.

Mission and authority

This skill is a boundary protocol, not a manual, tutorial, or analysis planner. It protects Marivo-specific semantic, evidence, state, and routing boundaries. It does not describe how to perform an analysis, which operators to call, or how to interpret results.

An environment-verified live Marivo surface outranks any cached knowledge in this file. Unverified PATH output does not. If the installed Marivo package, its Python interpreter, or its help surface cannot be confirmed before analysis begins, the skill stops and requests environment repair rather than guessing.

Live-contract rule

Use the same project interpreter for discovery and execution. Start with <analysis-python> -m marivo help analysis or the corresponding <venv>/bin/marivo help analysis, verify the rendered Marivo version, resolved Python executable, and package path, then follow focused live help topics for every API contract.

A bare marivo resolved from PATH is not authoritative unless its rendered fingerprint matches the interpreter and package used by the analysis process. If the intended analysis interpreter cannot be identified, stop and request environment repair rather than selecting a likely executable.

After entry:

• live help owns APIs, operators, constraints, examples, and recovery;
• semantic object details own the business-object contract;
• artifact reads own current-state facts and mechanical compatibility;
• structured errors own repair guidance.

Focused help is responsible for being self-contained for one correct minimal invocation. Complex legal investigations may consult as many focused topics as their actual branches require; this skill neither counts nor forbids those calls. Help-call limits are interface evaluation thresholds, not runtime permissions enforced by the skill.

The skill does not enumerate any API details, signatures, or examples.

Script workspace

Use a rerunnable Python script as the default execution unit after analysis entry. Store it under the current session's project-local <project_root>/.marivo/analysis/sessions/<session.id>/scripts/ directory and run it with the verified <analysis-python>. On failure, repair and rerun the same script instead of replacing it with disposable snippets.

Before authoring equivalent code, the agent may read scripts from this project's current or prior sessions as reference material for imports, session recovery, and operator composition. A prior script is an eligible reference only when its originating session records a succeeded job and the output artifact remains recoverable. Reference-only means never executing it directly, copying it wholesale, or treating it as trusted current code.

For every new analysis, re-resolve semantic refs, time scopes, and parameters against live help and current state. The script is not evidence: material claims remain traceable to current semantic objects, artifacts/jobs, and analysis scope.

Historical session reference

Historical sessions are external reference memory, not prompt context that is loaded by default. Inspect them only when resuming work, when the current question clearly repeats earlier work, or when the same failure recurs. Use the bounded historical-session surface from live recovery help and inspect no more than three candidate sessions before returning to current live state.

Historical session metadata and scripts may suggest recovery or operator composition, but they do not support current material claims. Never inherit a prior conclusion, semantic assumption, or parameter without resolving it against the current semantic catalog, runtime fingerprint, and analysis scope.

Hard boundaries

Semantic authority

Business metrics, dimensions, time dimensions, relationships, and caliber come from the semantic catalog. Analysis code must not infer or redefine business objects inside the semantic layer. A missing or disputed semantic object stops the affected typed branch. During analysis the agent must not add, edit, or remove semantic definitions; durable authoring is deferred until the user approves the closeout proposal.

Live-state authority

The agent acts on the installed runtime, current semantic state, current artifacts, and current structured errors. Skill text and historical examples must never override live state.

Judgment separation

Artifacts, typed findings, bounded digests, candidates, scores, quality results, issues, and affordances are computed facts or mechanical compatibility information. They are not business conclusions, recommendations, priorities, or stop conditions. The agent owns cross-artifact synthesis and every next-step judgment.

Evidence integrity

An artifact digest is a bounded operator-local read model, not a replacement for the artifact or exact findings. When the question exceeds its retained items or inference boundaries, the agent must inspect the live fallback surface. It must not hide issues that affect validity or coverage, upgrade an epistemic kind, or sever the recoverable evidence chain during script, session, or agent transitions.

Terminal boundary

Leaving typed Marivo analysis, adding semantic objects, and producing or publishing deliverables must use the corresponding public boundary. session.observe(...) is the sole producer of an initial canonical MetricFrame. frame.to_pandas() and md.raw_sql(...) are the sole terminal exits; results from either cannot re-enter typed analysis. Missing business semantics remain unresolved until approved semantic authoring; runtime capability gaps remain custom terminal work until modeled explicitly. One-off analysis code must not absorb another layer's responsibility.

Choose a terminal exit deliberately from current artifact state and mechanical compatibility. Familiarity with local pandas, SQL, or prior scripts is not a reason to leave typed flow; cross the boundary only when the remaining work is intentionally custom and terminal.

When current governed refs are sufficient but no catalog metric expresses the required session caliber, follow live help for the closed runtime metric expression surface and materialize it through session.observe(...). This remains typed analysis and is preferred to a terminal escape; it does not create catalog authority and triggers the conditional closeout reference below.

When a semantic gap blocks typed analysis, md.raw_sql(...) is an allowed terminal escape without prior approval. The agent may use explicit temporary inferred semantics, but must record every assumption and keep the result separate from canonical Marivo evidence. A raw-SQL result cannot re-enter typed analysis, become a semantic object, or erase the underlying semantic gap.

Routing

Condition	Route
A required business object is missing or must change	Stop the affected typed branch; optionally use terminal md.raw_sql(...); defer marivo-semantic until closeout approval
Semantic authoring returns ready refs	Read the current ReadinessReport; after blockers are cleared and warnings are disclosed, consume only analysis_ready_refs through the ordinary analysis APIs
The task needs terminal custom analysis	md.raw_sql(...) or frame.to_pandas() (terminal; cannot re-enter typed analysis)
The user requests a durable report, notebook, slides, HTML, or publishing	The corresponding independent delivery capability
The work is Marivo repository maintenance or dogfooding	Follow repository-local maintainer instructions; do not use the public skill as maintainer guidance

A missing or changed semantic object produces an AnalysisRepair(kind="semantic_authoring"). This identifies why typed analysis cannot continue; it is not permission to mutate the semantic layer during the analysis. The agent records the exact gap, may use terminal raw SQL, and requests approval for the smallest semantic change at closeout. Only after approval does it follow the semantic help target through marivo-semantic. After that change passes explicit readiness, analysis resumes with the report's current analysis_ready_refs.

Boundary-violation behavior

The skill does not provide fallback implementations.

Situation	Required behavior
Missing or ambiguous semantic object	Distinguish invalid lookup from genuine absence; stop the affected typed branch, use terminal raw SQL only with disclosed temporary assumptions, and defer authoring until closeout approval
Invalid API, shape, parameter, or operator	Follow the current structured error and live help; do not use a skill-cached workaround
Result-impacting blocker	Repair it, weaken the conclusion explicitly, or stop; never silently emit a stronger claim
Session or artifact cannot be recovered	Use the live recovery surface and disclose evidence-chain loss if recovery still fails
Correct Marivo version/help is unavailable or its fingerprint differs from the execution environment	Stop guessing and diagnose the environment; do not continue analysis
Task crosses a typed-analysis boundary	Hand off explicitly to the appropriate semantic, terminal-analysis, or delivery capability
Skill conflicts with live Marivo	Treat the live surface as authoritative and the skill as drifted

Closeout obligations

The skill does not prescribe a report structure. Require only that:

• material claims remain traceable to the relevant semantic object, artifact/job, and analysis scope;
• Marivo facts, agent interpretations, and unverified hypotheses remain distinguishable;
• result-impacting issues, quality limitations, omissions, and inference boundaries are disclosed;
• semantic gaps that weakened or blocked the task are named and handed back as a smallest-change proposal that requires explicit user approval before semantic authoring;
• every raw-SQL escape names the missing or incorrect semantic object, why the typed branch stopped, the datasource and analysis purpose, all temporary inferred semantics, and the terminal/bounded/no-lineage/no-evidence-continuity limitations; canonical artifact claims and raw-SQL-supported claims remain visibly separate;
• absolute interpreter and package paths from the environment fingerprint do not enter user-facing reports or deliverables. Internal diagnostic logs and evaluator transcripts may retain them.

When a material claim uses a frame whose persisted metric identity is runtime_expression, read and apply references/runtime-metric-closeout.md. Runtime expressions are controlled session-caliber computations, not catalog semantic authority; their labels must never be presented as governed business definitions.