[Farzad Sunavala (farzad528)]

Description

Add Don't answer yet: an evidence gate with Foundry IQ, a runnable tutorial that separates evidence retrieval, sufficiency assessment, and application-owned answer policy.

Foundry IQ retrieves Microsoft Learn documentation and three indexed sample files about Atlas, a fictional IT-ticket portal. Microsoft-Decision-1 returns a typed sufficiency score; application code either generates a cited answer or preserves the unanswered question and evidence for review.

The recipe includes a frozen-evidence gated/ungated comparison, real saved outputs, an optional bounded retrieval retry, reproducible setup helpers, and 34 regression tests. Historical Relay evaluation is an explicitly separate companion, not evidence for Atlas accuracy. Hamel Husain's decision-model evaluation article is credited for evaluation guidance, without transferring its benchmark results to this workload.

Related issue: none.

Scope and provenance

• Include the execution helpers, three sample team files, current and historical snapshots, claim-review ledger, and historical evaluation dependencies.
• Preserve the observed unsupported-role-equivalence failure, its prompt mitigation, and the possible unnecessary platform handoff.
• Exclude obsolete unreferenced diagrams, unused setup tooling, local environment files, and the Windows media junction mirror.
• Integrate current main and resolve the registry append conflict by retaining both recipes.
• No recipe code, threshold, rubric, or recorded model results were changed during PR preparation.

Validation

Command	Result
python -m unittest discover -s notebooks\data\foundry-iq-decision-1-gate -p test_gate.py	Passed: 34 tests, repeated after integrating main.
python .agents\skills\cookbook-review\scripts\recipe-health-check.py notebooks\foundry-iq-decision-1-gate.ipynb --repo-root .	All primary mechanical checks passed; function-definition/setup cells produce output warnings only.
python .agents\skills\cookbook-review\scripts\recipe-health-check.py notebooks\foundry-iq-decision-1-evaluation.ipynb --repo-root .	All checks except filename-derived asset-slug inference passed. The companion intentionally reuses the primary recipe's historical dataset; the reviewer accepted this shared-data exception.
From scripts: npx tsx validate-registry.ts	Passed: 20 recipes. Existing upstream title-length advisory for acu-dynamic-hitl remains unchanged.
From site: npm run build	Passed after integrating main and restoring its dependencies; includes astro check with 0 errors, 0 warnings, and 12 hints, plus static build and Pagefind.
git diff --check origin/main...HEAD	Passed.

The recorded live recipe execution completed on 2026-10-24. Its code, helper, team-data, and ledger provenance matches the submitted recipe. PR preparation did not repeat paid live model calls or provision Azure resources. The four-answer claim ledger is an authoring-assistant check, not independent human approval.

Review focus

Independent cookbook-review agent: 89/100 — Publish after fixes. It found no confirmed critical implementation blocker and considered the complete package suitable for a normal ready-for-review PR, not approval to merge or publish.

Axis	Score
Thesis	13/15
Opinionated defaults	9/10
When / What / How	9/10
Code cell discipline	9/10
Before/after evidence	12/15
Runnability	13/15
Scope	9/10
Dev-to-dev voice	5/5
Failure modes	5/5
Takeaway artifact	5/5
Total	89/100

Merge/publication holds

• Author personally checks the four accepted Atlas answers against their retained evidence; the assistant-authored ledger does not establish that review.
• Decision-1 owner clarifies the observed score variation for identical recorded inputs, expected repeatability, relevant request settings, and deployment-alias revision behavior. Both historical platform scores fail this recipe's 0.95 cutoff, but their cause remains unknown.
• Relevant Foundry PM, engineering, or DX reviewer checks technical accuracy and positioning.
• A maintainer explicitly approves retaining notebook outputs. They intentionally show the experiment's results and failures, so the output-clearing checkbox below remains unchecked rather than claiming compliance.

The gate measures evidence sufficiency, not generated-answer faithfulness. The 0.95 operating cutoff is not an accuracy guarantee. Please review those boundaries and the disclosed possible false negative. Advisory improvements are a shorter opening and lighter optional reading; no broader classifier or benchmark is proposed.

---

Checklist

Notebook contributions

Skip if this PR only touches site code.

• Recipe author is an approved Microsoft Forgebook author, or this contribution has maintainer approval
• Recipe was drafted or validated through the internal authoring workflow
• Foundry PM, engineering, or DX reviewer has checked product accuracy and positioning
• Notebook added to notebooks/ directory
• Entry added to registry.yaml with all required fields (slug, path, title, authors)
• Approved author entry exists in authors.yaml (with at least name and title)
• Tags are from the allowed list in tags.yaml
• Registry validation passes: cd scripts && npx tsx validate-registry.ts
• Notebook runs cleanly end-to-end
• Cell outputs cleared before committing
• No site-specific metadata inside the notebook
• No secrets, tenant-specific values, or private endpoints are committed
• Images stored in notebooks/media/<slug>/ with descriptive alt text
• Data files stored in notebooks/data/<slug>/ when applicable

Site changes

Skip if this PR only adds a notebook.

• Site builds without errors: cd site && npm run build
• No TypeScript errors: cd site && npx astro check

[github-actions]

🔍 Preview

Build and checks passed, but a live preview wasn't auto-deployed because Farzad Sunavala (@farzad528) isn't a member of forgebook-contributors yet. A maintainer can add you to the team and re-run this workflow to deploy the preview.