Add Gemma curation agents case study

[cmungall]

Adds a new case study documenting the Gemma curation agents pipeline, a batch-based agentic curation system for transcriptomic study metadata.

Summary

This case study documents a production agentic curation project that differs structurally from the Git-native knowledge bases covered elsewhere in the guide. Rather than pull-request-based workflows, Gemma uses a batch CLI pipeline with database output and web-based review. It is included for its rigorous measurement practices and architectural patterns worth copying, not for its workflow shape.

Changes

• New case study document (docs/case-studies/gemma-curation.md): Comprehensive guide covering:

  • Project overview and key repositories (curation agents, benchmark data, review UI)
  • Detailed "What to look at" table pointing to specific implementation patterns
  • Four key architectural wins: curator manual as rule corpus, visible escalation ladder (mechanical → LLM stages), grounding reuse, and gold-blind pipeline design
  • Practical guidance on what to copy first (distilling curation manuals into rule files)
  • Honest assessment of gaps (non-diffable review, unversioned rule sources, radical human-review minimization stance)

• Updated case studies index (docs/case-studies/index.md): Added new "Block C: pipelines" section to categorize non-Git-native projects

• Updated navigation (mkdocs.yml): Added Gemma curation agents under new Pipelines subsection

Notable details

The case study emphasizes reproducibility and honesty:

• Documents the benchmark as a separate versioned repository with checksums and difficulty stratification
• Explains design rules (gold-blind construction, never silent information loss) as explicit constraints
• Flags limitations: inter-curator agreement only 60–90%, cost figures subsidized by Anthropic, omission errors not predictable
• Provides concrete paths to specific implementation files rather than abstract descriptions

https://claude.ai/code/session_01SUoEy2WWzPRSGoopbot6YX

[claude]

Claude encountered an error —— View job

---

I'll analyze this and get back to you.

[cmungall]

claude-review is failing on an expired credential, not on this PR

claude-review fails on this PR's head (4fc4cca), and it is not this diff's failure. The job's own log gives the cause on its first API call:

API Error: 401 {"type":"error","error":{"type":"authentication_error",
"message":"OAuth access token has expired. Re-authenticate to continue."}}

The run reports num_turns: 1, input_tokens: 0, total_cost_usd: 0 — the reviewer never read a file or reached the diff. The repository's CLAUDE_CODE_OAUTH_TOKEN secret has expired, which affects every pull request in this repository equally, not just this one.

No fix exists that I can port into this PR. Rotating a repository secret needs a maintainer with repository settings access: re-authenticate and update CLAUDE_CODE_OAUTH_TOKEN under Settings → Secrets and variables → Actions. Once that is done, the check will pass on a re-run with no change to this branch.

I have not re-run the job. A 401 on a stored credential is deterministic — the same expired secret would produce the same error — so a re-run would cost a cycle and tell us nothing. Happy to re-run it as soon as the secret is rotated.

The rest of CI is green on this commit:

Check	Result
Build site	success
Check links	success
check-mention	success
claude-review	failure (expired credential, above)

Check links passing is the one worth noting, since this PR adds a batch of new external links to the PavlidisLab repositories and the paper DOI. There is no merge conflict against main.

---

Generated by Claude Code