Research Analysis

research-analysis

Purpose

/research-analysis manages the records behind a package's existing Analysis page. It owns two editorial decisions:

1. whether an experimental result supports a mechanism-level Insight;
2. whether that Insight warrants a reusable package Rule.

The records live in unified research state. The browser page under .research/interface/ is a derived view and must never become a store.

Authority and storage

Authority runs in this order:

1. the user's request and explicit payload;
2. the management schema and EventStore policy;
3. the trust and interface rules shipped by /research-dashboard;
4. this skill's editorial contract.

The relevant data is:

All mutations use skills/research-op/scripts/research_op.py, except init, whose small helper commits the package page-selection event through the same management layer. Do not edit .research/state/events.jsonl, .research/state/current.json, or anything below .research/interface/.

Boundary

• results.html answers what happened. Its rows come from verified run results.
• Analysis answers why it happened and what should be remembered.
• Insights and Rules require an explicit editorial decision. Fact propagation does not create them.
• An Insight is a non-binding Learning. A Rule is binding package memory and needs stronger evidence.
• A package Rule is not a methodsTried verdict row. The two records answer different questions.
• Terminal packages are frozen. If research-op rejects a mutation, surface the rejection instead of patching the rendered page.

Preflight

Confirm that the versioned research root and target Package exist:

test -f .research/VERSION
python skills/research-op/scripts/research_op.py \
  show package <package-id> --workspace .

If the Package is missing, use /research-package. If the workspace is legacy or unversioned, stop; automatic migration is unsupported. The generated interface is not an execution prerequisite.

Operations

init <package-id>

New Packages include Analysis in their standard page set. Use init only to repair an older Package whose state-backed pages[] omitted it:

python skills/research-analysis/scripts/init_analysis_page.py \
  --workspace . \
  --package-id <package-id>

The command appends "analysis" to Package.pages[] through a PackageMutationApplied event. It is idempotent and marks the interface stale; the Dashboard or next static-page request rebuilds the projection.

add-insight <package-id> <insight-id>

Record one evidence-backed Insight through the analysis-insight facade:

python skills/research-op/scripts/research_op.py \
  --workspace . \
  --pkg <package-id> \
  --op insert \
  --target analysis-insight \
  --payload '{
    "id":"<kebab-case-id>",
    "title":"<short title>",
    "lead":"<observed pattern>",
    "reading":"<what the evidence shows>",
    "mechanism":"<why it happened>",
    "provenance":".research/experiments/<pkg>/<experiment>/<run>/result.json"
  }'

Required fields:

• id or slug;
• title;
• at least one of lead, reading, or mechanism;
• provenance that identifies the evidence.

The gateway records a Learning and updates the package's renderer-facing analysisInsights[] fields in the same operation. Use --op update with the same stable id to revise it. Use --op delete with {"id":"<id>","reason":"<why>"} for an explicit correction.

The interface renders each Insight as a closed <details> block. Text fields are HTML-escaped. Raw HTML is not part of the current Insight payload contract.

add-rule <package-id> <slug>

Distill an existing Insight into a package Rule only when finalized result evidence supports a general lesson:

python skills/research-op/scripts/research_op.py \
  --workspace . \
  --pkg <package-id> \
  --op insert \
  --target rule \
  --payload '{
    "level":"package",
    "kind":"lesson",
    "slug":"<kebab-case-slug>",
    "title":"<short title>",
    "text":"<standalone rule prose>",
    "rationale":"insight-<insight-id>",
    "addedAt":"<YYYY-MM-DD>"
  }'

The Rule text must stand on its own and remain plain text. Put thresholds, named subjects, and other binding details in text; put the supporting Insight id in rationale. research-op rejects a lesson when the package has no finalized result evidence.

lint <package-id-or-all>

Lint the authoritative state:

python skills/research-analysis/scripts/lint_analysis.py \
  --workspace . \
  --package-id <package-id>

python skills/research-analysis/scripts/lint_analysis.py \
  --workspace . \
  --all

The linter checks:

• Analysis is enabled in Package.pages[];
• Insight ids are unique kebab-case ids;
• every Insight has a title, content, and provenance;
• active lesson Rules contain plain text;
• each lesson Rule cites an Insight that exists for the package.

It does not parse or mutate rendered HTML.

Human layout contract

The current Analysis page keeps the existing human layout:

1. Rules appear first as a numbered list.
2. Insights follow as closed, collapsible cards.
3. The interface renderer supplies the empty placeholders.
4. The package navigation and shared CSS remain owned by /research-dashboard and lib.interface.

skills/research-package/templates/analysis.html is the single layout template. This skill does not carry or write a second page template.

Visualization boundary

references/viz-templates.md records the approved visual language for a future typed visualization field. The current analysis-insight facade accepts text fields and does not render arbitrary HTML. Do not paste a visualization into .research/interface/ or smuggle raw HTML into state. Until the schema and renderer gain a typed visualization payload, place the finding in reading and mechanism, with the result artifact in provenance.

Inspection

Inspect the stored records through bounded queries:

python skills/research-op/scripts/research_op.py \
  show learning '<package-id>::learning::<insight-id>' --workspace .

python skills/research-op/scripts/research_op.py \
  show rule '<package-id>#<rule-slug>' --workspace .

After any accepted mutation, confirm that the automatic interface rebuild succeeded, then run the state linter. A clean lint is silent and exits with status 0.

Final response

Report:

• the operation and package id;
• the committed event ids;
• the Learning or Rule aggregate id;
• the lint result;
• whether the interface rebuild succeeded;
• the next decision, if one remains.

Bundled resources

• scripts/init_analysis_page.py: enables Analysis in Package state.
• scripts/lint_analysis.py: validates the state contract.
• templates/rule-bullet.html: package Rule payload reminder.
• references/viz-templates.md: frozen visualization layout reference, not a mutation path.