summary-accuracy-verification

name	summary-accuracy-verification
description	Verify the generated corpus summary against the prepared NotebookLM notebook, source-evidence.md, optional targets.md, .workflow_state.json, and prompt2.txt.

Summary Accuracy Verification

Overview

Verify the generated summary by querying the prepared corpus NotebookLM notebook, saving refreshed source-grounded evidence to output/source-evidence.md, and refining output/summary.md with Gemini CLI using prompt2.txt. Treat NotebookLM responses as the only substantive source evidence, including citation details.

Prefer the scripted entrypoint when running the full project workflow:

./run_workflow.sh --verify

Workflow

1. Verify NotebookLM readiness before source queries.

   • Run nlm login --check.
   • Read the notebook ID from .workflow_state.json using the knowledge:notebook key.
   • Use the single project corpus notebook.
   • Do not delete notebooks or sources.
   • If the notebook is missing from state, stop and report that ./run_workflow.sh --execute must prepare it first.

2. Query the corpus notebook for verification evidence.

   • Include the processed-source manifest.
   • Include targets.md when present and treat it as the hard scope.
   • Ask NotebookLM for detailed source-grounded evidence, coverage requirements, citation details, and accuracy warnings.
   • Do not send the entire current summary to NotebookLM; long generated summaries can exceed practical query size and fail.

3. Save the evidence sidecar.

   • Write combined NotebookLM verification notes to output/source-evidence.md.
   • The evidence file should contain NotebookLM citations, page references, source labels, or other available citation details.

4. Run the final Gemini verification pass with prompt2.txt.

   • Assemble stdin with:
     • corpus context
     • optional hard-scope targets
     • current markdown summary when revising
     • NotebookLM verification evidence
   • Invoke Gemini with project defaults unless the user requests a model:

gemini --model pro --prompt="$(cat prompt2.txt)" < verification-input.txt

1. Handle the result.
   • The expected Gemini output is only final markdown summary text.
   • The summary must use numbered inline citations and end with ## References.
   • Replace output/summary.md only when the task asks to update or revise files.
   • If the task asks only for an audit, report the accuracy findings and do not write output files.
   • Preserve markdownlint-compatible formatting and the rigorous academic tone from AGENTS.md.

Validation

After writing revised summaries or skill changes, run the repo checks that apply:

jq . .gemini/settings.json
shellcheck run_workflow.sh
shfmt -d run_workflow.sh
gemini skills list
markdownlint -c .markdownlint.json README.md AGENTS.md .agents/skills/**/*.md .gemini/agents/*.md

If markdownlint is not installed, state that it was skipped.