Skip to content
OpenAgentSkill
Browsing is open · Catalog captured 2026-11-21 · Automatic updates and account services are paused.
Design & UI

deep-analyzer

Optional modelling stage: regression, mixed effects, clustering, dimensionality reduction, forecasting. Only runs when the design supports it; every model carries diagnostics.

MAGA2010 · Static checks on file

Installation

Read the source and dependencies before running a command. Copying does not install or execute anything.

npx skills add MAGA2010/data-master --skill deep-analyzer

About this skill


name: deep-analyzer description: "Optional modelling stage: regression, mixed effects, clustering, dimensionality reduction, forecasting. Only runs when the design supports it; every model carries diagnostics." parent: data-master stage: 2d

deep-analyzer

Stage 2d of the data-master pipeline. The router in ../../SKILL.md owns the end-to-end run; this skill owns one stage and one gate. Read the parent contract before changing anything here: the invariants in ../../SKILL.md (additive-only, types decided once, default-FAIL gates, three-part provenance) are not restated, they are inherited.

When to invoke

Only when the user asks for a model, or when a stated question cannot be answered by a group comparison. Never the default: a model adds assumptions that must be defended.

Inputs

  • cleaned data
  • a declared model specification: outcome, predictors, random effects, family, link

Outputs

  • analysis/model.json: coefficients with CIs, fit statistics, diagnostics, n, unit_of_analysis, rule_source
  • analysis/model_residuals.png when diagnostics are visual

Commands

# v1 ships no model-fitting script: a deep-analysis run is an explicit,
# supervised step that must leave analysis/model.json in place.
python -c "import json; d=json.load(open('outputs/analysis/model.json',encoding='utf-8')); assert d.get('rule_source'); print('model artifact OK')"

Gate (default FAIL)

Every coefficient has an estimate, a CI and the unit it applies to. Every model reports at least one diagnostic appropriate to its family (residual vs fitted, overdispersion, VIF, convergence). A model that did not converge is reported as such, never as a fit.

Failure modes seen in real runs

  • Pseudo-replication. Rows that share a subject are modelled as independent, so every standard error is too small. Random effects are required when the design is nested.
  • A converged-looking fit with separation. Logistic regression on a rare outcome can produce enormous coefficients; the diagnostic must be printed, not skipped.
  • Comparing models on different n. AIC across different complete-case sets is not a comparison.

Do not

  • do not run a model because it is available;
  • do not report a coefficient without its scale and CI;
  • do not drop the diagnostics to keep the table short.

Upstream / downstream

  • previous: ../statistical-analyzer/SKILL.md
  • next: ../visualizer/SKILL.md
  • contract: ../../references/workflow-contract.md