Overall Assessment

✅ Looks good — no unmitigated P0/P1 findings.

This PR is packaging/docs/API-surface work only. No estimator math, weighting, variance/SE, identification checks, or default behaviors changed. Execution-based validation was not possible in this environment because pytest and required runtime deps such as numpy are unavailable.

Executive Summary

• No methodology-sensitive code changed; the only methodology-adjacent edit is a path update in docs/methodology/REGISTRY.md:L2770.
• get_llm_guide() is a small whitelist-based resource accessor, so I did not find new correctness, security, or performance risks in the implementation.
• P3: the bundled practitioner guide still points to a repo-only docs/methodology/REGISTRY.md, which breaks the offline breadcrumb for wheel-only installs.
• P3-informational: build-surface validation is still thin; the added tests exercise source-tree resources, and related .txt guide CI-validation debt is already tracked in TODO.md:L95.

Methodology

No findings. Affected method(s): none. The REGISTRY change at docs/methodology/REGISTRY.md:L2770 is path-only and preserves the existing Baker et al. adaptation note.

Code Quality

No findings.

Performance

No findings. The new code path is an explicit on-demand text read, not estimator-path logic.

Maintainability

No findings.

Tech Debt

• P3-informational Impact: The feature depends on the distribution surfaces configured in pyproject.toml:L83-L97 and docs/conf.py:L74-L77, but the added tests in tests/test_guides.py:L35-L42 only validate source-tree resources. Related AI-guide CI debt is already tracked in TODO.md:L95. Concrete fix: when CI is next touched, add a wheel-install smoke test plus a docs-build assertion that the three llms*.txt files are published at the expected output paths.

Security

No findings. get_llm_guide() uses a fixed variant whitelist instead of interpolating user input into filesystem paths, so it does not introduce an arbitrary file-read/path-traversal surface.

Documentation/Tests

• P3 Impact: The shipped practitioner guide still tells readers to see docs/methodology/REGISTRY.md in diff_diff/guides/llms-practitioner.txt:L6. That path will not exist in a wheel-only/offline install, leaving a dead local breadcrumb inside the artifact this PR is trying to make self-contained. Concrete fix: replace that repo-only reference with an inline summary of the relevant deviation note, or point to another bundled/offline surface instead.

/ai-review

🔁 AI review rerun (requested by @igerber)

Head SHA: e65be1665c2d27c2f421619fe1cdcc1164a63cd2

Overall Assessment

✅ Looks good

No unmitigated P0/P1 findings. This re-review resolves the earlier dead offline-breadcrumb issue. The only remaining item I found is a tracked P3-informational validation gap around built artifacts and docs publication.

Executive Summary

• No estimator, weighting, variance/SE, identification-assumption, or default-behavior code paths changed.
• Methodology cross-check passed: the practitioner-guide deviations now inlined in diff_diff/guides/llms-practitioner.txt:L8-L20 match the documented **Note:** entries in docs/methodology/REGISTRY.md:L2774-L2785.
• The prior P3 about the bundled practitioner guide pointing wheel users at a repo-only path is resolved.
• Residual P3-informational: guide tests still stop at source-tree resource reads; wheel-install and docs-publication validation remain deferred, but that debt is already tracked in TODO.md:L95-L96.
• I could not execute pytest or import diff_diff in this environment because numpy is not installed.

Methodology

Affected method(s): none.

Severity: P3-informational
Impact: The only methodology-adjacent change is the bundled practitioner guide text. The inlined Baker et al. adaptation details in diff_diff/guides/llms-practitioner.txt:L8-L20 mirror the documented deviations in docs/methodology/REGISTRY.md:L2774-L2785, so under the review rubric this is a documented deviation, not an undocumented methodology change.
Concrete fix: None.

Code Quality

No findings.

Performance

No findings. The new guide read in diff_diff/_guides_api.py:L13-L48 is on-demand and outside estimator/inference hot paths.

Maintainability

No findings.

Tech Debt

Severity: P3-informational
Impact: The new “wheel content” check in tests/test_guides.py:L35-L42 still compares source-tree package resources rather than a built wheel, and the new public docs surface depends on docs/conf.py:L74-L78 and pyproject.toml:L83-L97. This validation gap is already tracked in TODO.md:L95-L96, so it is not blocking.
Concrete fix: Keep the tracked follow-up to add wheel-install and docs-build smoke checks.

Security

No findings. The new accessor uses a fixed variant whitelist instead of interpolating user input into filesystem paths in diff_diff/_guides_api.py:L6-L48.

Documentation/Tests

No new untracked findings. The added coverage in tests/test_guides.py:L10-L67 is reasonable for variant loading, invalid inputs, namespace export, and docstring surfacing, and the earlier offline-breadcrumb issue is fixed by diff_diff/guides/llms-practitioner.txt:L3-L20. Verification note: I could not run the tests here because importing diff_diff fails without numpy in this review environment.

/ai-review

🔁 AI review rerun (requested by @igerber)

Head SHA: b5f063a6d8a141b43a930100701d6f1d209f8ab8

Overall Assessment

✅ Looks good

No unmitigated P0/P1 findings. This re-review stays clear of estimator, weighting, variance/SE, identification-assumption, and default-behavior changes. The only issues I found are non-blocking P3 items around docs/dependency tracking and release validation.

Executive Summary

• No methodology-bearing estimator code changed; the PR is limited to packaging, docs, and a public guide accessor.
• Methodology cross-check passed: the inlined Baker et al. adaptation notes in diff_diff/guides/llms-practitioner.txt:L3 match the documented **Note:** entries in docs/methodology/REGISTRY.md:L2770.
• The prior offline-breadcrumb issue remains resolved: bundled guides now point users to diff_diff.get_llm_guide("practitioner") in diff_diff/guides/llms.txt:L30 and diff_diff/guides/llms-full.txt:L36.
• get_llm_guide() is correctly surfaced through the top-level namespace and module docstring in diff_diff/init.py:L12, diff_diff/init.py:L205, and diff_diff/init.py:L409, and the accessor uses a fixed variant whitelist in diff_diff/_guides_api.py:L6.
• P3: docs/doc-deps.yaml updates moved guide paths but still does not map the new diff_diff/_guides_api.py source file near docs/doc-deps.yaml:L628; that is a non-blocking maintainability gap and falls under the already-tracked doc-deps integrity debt in TODO.md:L96.
• P3: the new guide tests are still source-tree oriented rather than built-artifact oriented; tests/test_guides.py:L35 does not validate a built wheel or the published docs path configured in docs/conf.py:L74 and pyproject.toml:L83.

Methodology

• Severity: P3-informational
• Impact: Affected method(s): none. The only methodology-adjacent change is the practitioner-guide framing text, and its deviations from Baker et al. are explicitly documented in both diff_diff/guides/llms-practitioner.txt:L3 and docs/methodology/REGISTRY.md:L2770. Under the review rubric, that is a documented deviation, not a defect.
• Concrete fix: None.

Code Quality

No findings.

Performance

No findings. get_llm_guide() is an on-demand text read outside estimator/inference hot paths in diff_diff/_guides_api.py:L13.

Maintainability

• Severity: P3
• Impact: The PR updates many moved-guide references in docs/doc-deps.yaml:L612 and docs/doc-deps.yaml:L628, but it does not add an entry for the new public helper module diff_diff/_guides_api.py. That means docs-impact tooling can miss future documentation drift when the accessor changes.
• Concrete fix: Add a diff_diff/_guides_api.py entry to docs/doc-deps.yaml mapping it to the bundled guide docs and/or the public API surface, or explicitly document why it is intentionally excluded.

Tech Debt

• Severity: P3-informational
• Impact: test_wheel_content_matches_package_resource() in tests/test_guides.py:L35 only compares the accessor against source-tree package resources; it does not prove that the wheel actually contains the files or that the new ReadTheDocs root-path publication configured in docs/conf.py:L74 and linked from pyproject.toml:L83 is present after build/deploy. This is non-blocking and compatible with ✅, especially given the already-tracked docs/CI validation debt in TODO.md:L95 and TODO.md:L96.
• Concrete fix: Add one wheel-install smoke test that imports diff_diff from the built artifact and calls get_llm_guide(), plus one docs-build smoke test that asserts llms-practitioner.txt is emitted at the expected site-root path.

Security

No findings. The accessor dispatches through a fixed whitelist rather than interpolating user input into filesystem paths in diff_diff/_guides_api.py:L6.

Documentation/Tests

No new untracked findings. Coverage in tests/test_guides.py:L10 is reasonable for variant loading, invalid inputs, namespace export, and docstring surfacing.

Verification note: I could not run pytest tests/test_guides.py or import diff_diff in this environment because numpy, pandas, scipy, and pytest are not installed.

/ai-review

🔁 AI review rerun (requested by @igerber)

Head SHA: 70b547bd8e3f905c9deb0b296778edd14dd1930a

Overall Assessment

✅ Looks good

Executive Summary

• No estimator, weighting, variance/SE, identification-assumption, or methodology-bearing default behavior changed. The only methodology-adjacent edits are guide-text updates, and the Baker et al. deviations remain explicitly documented and aligned between diff_diff/guides/llms-practitioner.txt:L3-L20 and docs/methodology/REGISTRY.md:L2782-L2797.
• Re-review status: there are still no unmitigated P0/P1 issues in scope, so the assessment stays ✅.
• The new get_llm_guide() accessor is wired correctly into the public surface and implemented with a fixed variant whitelist in diff_diff/init.py:L1-L15, diff_diff/init.py:L204-L205, diff_diff/init.py:L406-L409, and diff_diff/_guides_api.py:L1-L48.
• P3-informational: the prior doc-deps finding remains. diff_diff/_guides_api.py:L1-L48 is still unmapped in docs/doc-deps.yaml:L616-L637, but that falls under the already-tracked unmapped-source debt in TODO.md:L95-L96.
• P3-informational: the prior release-validation gap also remains. tests/test_guides.py:L35-L42 checks source-tree/package-resource equivalence, not a built wheel or docs-build artifact.
• P3: the new llms-full.txt wording for get_loo_effects_df() is slightly inaccurate versus the actual API.

Methodology

• Severity: P3-informational. Impact: Affected methods: none in estimator/inference code. The changed practitioner-guide text preserves the already-documented deviations from Baker et al., and those deviations remain mirrored in the Methodology Registry at docs/methodology/REGISTRY.md:L2782-L2797. Concrete fix: None.

Code Quality

No findings.

Performance

No findings.

Maintainability

• Severity: P3-informational. Impact: diff_diff/_guides_api.py:L1-L48 introduces a new public, doc-bearing source file, but docs/doc-deps.yaml:L616-L637 still has no entry for it. Future accessor/docstring changes can drift without doc-deps tooling noticing. This is already tracked generically in TODO.md:L96-L96.
  Concrete fix: Add a diff_diff/_guides_api.py block to docs/doc-deps.yaml mapping it to the bundled guide docs and/or public API documentation.

Tech Debt

• Severity: P3-informational. Impact: tests/test_guides.py:L35-L42 proves get_llm_guide() matches package resources in the source tree, but it does not verify the built wheel or the Sphinx extra-path publication configured in pyproject.toml:L85-L98 and docs/conf.py:L74-L78. That leaves the README claim in README.md:L72-L82 validated only indirectly. This is compatible with the tracked .txt-guide CI gap in TODO.md:L95-L96.
  Concrete fix: Add one CI smoke test that installs the built wheel and calls get_llm_guide(), plus one docs-build assertion that llms-practitioner.txt is emitted at site root.

Security

No findings.

Documentation/Tests

• Severity: P3. Impact: diff_diff/guides/llms-full.txt:L1032-L1034 describes get_loo_effects_df() as leave-one-out influence “per treated unit” and implies a treated-side support condition, but the method actually returns both control and treated LOO rows and also requires more than one control with nonzero/effective weight to be available, per diff_diff/results.py:L1007-L1049 and diff_diff/synthetic_did.py:L1251-L1281. This is a documentation-accuracy issue only, not a methodology defect.
  Concrete fix: Reword the bullet to “per-unit leave-one-out influence” and either list the full availability conditions or point readers to the method docstring/error path.
• Verification note: I could not rerun pytest tests/test_guides.py or import diff_diff end-to-end in this environment because pytest and numpy are not installed.

/ai-review

🔁 AI review rerun (requested by @igerber)

Head SHA: 89ee3378a467c6450b840b662a6154d4fd32e7b2

Overall Assessment

✅ Looks good

Executive Summary

• Re-review status: still no unmitigated P0/P1 issues. This PR does not change estimator logic, weighting, variance/SE computation, identification assumptions, or methodology-bearing defaults.
• The methodology-adjacent practitioner-guide edits in diff_diff/guides/llms-practitioner.txt:L3-L20 match the documented Notes in docs/methodology/REGISTRY.md:L2782-L2797, so there is no undocumented deviation from Baker et al.
• The prior P3 documentation issue around get_loo_effects_df() appears resolved: diff_diff/guides/llms-full.txt:L1032-L1035 now matches diff_diff/results.py:L1007-L1038.
• Remaining items are P3 only: _guides_api.py is still unmapped in docs/doc-deps.yaml, artifact-level wheel/docs publication is still not exercised by tests, and the SyntheticDiDResults table in llms-full.txt still omits "jackknife" from variance_method.
• I could not rerun pytest or import diff_diff end-to-end here because this environment does not have pytest or numpy.

Methodology

No findings. Affected methods: none. The only methodology-adjacent change is guide prose, and its explicit deviations from Baker et al. remain documented in both diff_diff/guides/llms-practitioner.txt:L3-L20 and docs/methodology/REGISTRY.md:L2782-L2797.

Code Quality

No findings.

Performance

No findings.

Maintainability

• Severity: P3-informational. Impact: diff_diff/_guides_api.py:L1-L48 introduces a new public, doc-bearing source file, but the infrastructure block in docs/doc-deps.yaml:L616-L637 still has no entry for it. Changes to the accessor or its docstring can drift without the doc-deps map noticing. This limitation is already tracked generically in TODO.md:L96-L96. Concrete fix: add a diff_diff/_guides_api.py block mapping the accessor to the bundled guides and any public API docs that surface get_llm_guide().

Tech Debt

• Severity: P3-informational. Impact: pyproject.toml:L83-L98 and docs/conf.py:L74-L77 now advertise wheel-bundled guides and RTD-published text artifacts, but tests/test_guides.py:L35-L67 only exercises source-tree package resources; it does not validate a built wheel or built docs output. Packaging/publishing regressions would therefore slip past CI. This is compatible with the existing tracked guide-validation debt in TODO.md:L95-L95. Concrete fix: add one wheel-install smoke test that imports diff_diff from the built artifact and one docs-build assertion that llms-practitioner.txt is emitted at the site root.

Security

No findings.

Documentation/Tests

• Severity: P3. Impact: In the edited SyntheticDiDResults section, diff_diff/guides/llms-full.txt:L1024-L1024 still describes variance_method as "bootstrap" or "placebo" only, even though the same section now documents a jackknife-only diagnostic at diff_diff/guides/llms-full.txt:L1034-L1034 and the implementation supports "bootstrap", "jackknife", and "placebo" in diff_diff/synthetic_did.py:L50-L64 and diff_diff/results.py:L696-L724. This is a minor documentation inconsistency only. Concrete fix: update the attribute row to include "jackknife" or point readers to the estimator/results docstring.
• Verification note: I could not execute pytest tests/test_guides.py or import diff_diff locally in this environment because pytest and numpy are not installed.