docs: CiC launch carousel - "The effect your DiD can't see" (Tide deck)

[igerber]

Summary

• New LinkedIn launch carousel for ChangesInChanges: carousel/generate_cic_carousel.py -> carousel/diff-diff-cic-carousel.pdf (11 slides, regenerated in the same commit per carousel/README.md's contract). "Tide" visual identity (ocean navy / teal / seafoam on a light aqua gradient, coral kickers, amber significant-band highlight, distribution-as-water cover motif) - distinct from every existing deck palette.
• Narrative (villain-first, grounded in Athey & Imbens 2006): the effect mean DiD can't see (a simulated program with truth known by construction: $0.22 p=0.90 verdict vs a $3.01 true effect) -> dark twist (same data in logs: +14.0% p=0.001; the scale choice IS the assumption) -> the paper's answer (compare ranks, not scales; verbatim abstract pull quote; equivariance receipt: dollars fit == exp(logs fit)) -> the math (annotated Theorem 3.1 composition + QTE definition) -> the QTE-profile payoff -> sup-t uniform bands ("the bottom half moved" as one simultaneous claim) -> code slide (sklearn-style fit()) -> covariates (confidently-wrong 9.38 [8.32, 10.45] -> truth-covering 6.53 [5.35, 7.70]) -> production grid + validation strip -> CTA (pip install; the single Tutorial 27 mention).
• Claim discipline in the module docstring: every number is Tutorial 27's seed-locked value; parity wording scoped per the parity-suite header (unconditional point estimates match the R qte 1.3.1 golden files to 1e-10; "bit-exact" reserved for the type-1 quantile arithmetic at atol=0; covariate branch = layered parity per the REGISTRY Deviation note); equivariance qualified to unconditional fits on-slide; power-law (never "multiplicative") trend framing; bootstrap-only inference and fixed-95% sup-t bands framed as qte-parity conventions; no speed claims; QDiD not mentioned.
• tests/test_cic_carousel_claims.py (10 tests) closes the synchronization chain: deck constants are parsed via ast (no carousel-only dependencies in CI) and compared against values parsed from the committed Tutorial 27 notebook surface - which the T27 drift test locks against the library - plus visible-slide-surface phrase locks on every load-bearing qualifier (docstrings excluded; negative lock on "multiplicative").

Methodology references (required if estimator / math changes)

• Method name(s): ChangesInChanges (CiC) - marketing/documentation artifact only; no estimator implementation changes.
• Paper / source link(s): Athey & Imbens (2006), Econometrica 74(2), 431-497 (Theorem 3.1 composition; abstract pull quote verbatim from the published PDF, p. 431); R qte 1.3.1 (parity conventions).
• Any intentional deviations from the source (and why): None - the deck restates documented behavior with the REGISTRY's own qualifiers; the claim-discipline rules are written into the generator's module docstring.

Validation

• Tests added/updated: tests/test_cic_carousel_claims.py - 10 tests syncing every visible deck number to the committed Tutorial 27 notebook surface (QTE table, SEs, sup-t crit, ATT row, covariate CIs, receipt pairs, DGP constants) plus qualifier phrase locks. All green alongside the T27 drift suite.
• Backtest / simulation / notebook evidence (if applicable): all deck numbers originate from docs/tutorials/27_cic_distributional_effects.ipynb (seed-locked, drift-tested in CI).

Security / privacy

• Confirm no secrets/PII in this PR: Yes (PDF checked for local-path/username leakage).

---

Generated with Claude Code

[github-actions]

Overall Assessment

✅ Looks good. No unmitigated P0/P1 findings.

Executive Summary

• No estimator, weighting, SE/variance, or default behavior changes.
• The affected method is ChangesInChanges / CiC, but this PR adds a carousel artifact and claim-sync tests only.
• The visible methodology claims match docs/methodology/REGISTRY.md and the CiC implementation docstrings: qte parity, bootstrap-only inference, fixed-95% sup-t bands, and covariate-branch caveats are qualified.
• The Athey-Imbens abstract quote used on slide 4 matches the published abstract wording. (onlinelibrary.wiley.com)
• I could not run pytest or extract PDF text because pytest, pdfinfo, pdftotext, and PDF Python libraries are absent in this environment. I did reproduce the new claim-sync checks with plain Python.

Methodology

Finding: None.

Affected method: ChangesInChanges / CiC. The deck’s claims align with docs/methodology/REGISTRY.md:L4109-L4168, including the Theorem 3.1 composition, type-1/qte conventions, documented qte covariate deviations, bootstrap-only inference, and fixed-95% sup-t uniform bands. The slide copy preserves important qualifiers at carousel/generate_cic_carousel.py:L904-L912, carousel/generate_cic_carousel.py:L993-L998, and carousel/generate_cic_carousel.py:L1224-L1227.

Code Quality

Finding: None.

The generator is self-contained and avoids importing carousel-only dependencies from tests by using AST parsing in tests/test_cic_carousel_claims.py:L33-L58.

Performance

Finding: None.

The added generator is an offline artifact builder. It does not affect runtime library performance.

Maintainability

Finding: None.

The constants are centralized in carousel/generate_cic_carousel.py:L153-L226, and the new tests sync those constants against Tutorial 27 output and code at tests/test_cic_carousel_claims.py:L88-L197.

Tech Debt

Finding: None.

Existing CiC deferred scope is already tracked in TODO.md:L115-L119 and documented in the registry. This PR does not add new deferred methodology work.

Security

Finding: None.

I found no obvious secret or local-path strings in the added Python/test files or via strings over the PDF. No networked or credential-handling code is added.

Documentation/Tests

Finding: None.

The new test file covers headline numeric claims, QTE table values, sup-t split, covariate claims, receipt rows, DGP constants, and load-bearing qualifier phrases at tests/test_cic_carousel_claims.py:L88-L243.

Verification note: pytest was unavailable, so I could not run the test suite directly. Plain-Python replication of the new test logic passed, and file identifies carousel/diff-diff-cic-carousel.pdf as an 11-page PDF.

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: 5beb922d3ce6d0edf3a1f6418dd6a18af99ecd25

---

Overall Assessment

✅ Looks good. No unmitigated P0/P1 findings.

Executive Summary

• Re-review confirms the prior assessment; no previous P1+ findings required remediation.
• Changes are documentation, generation, and claim-sync tests only—no estimator or inference logic changed.
• CiC equations and claims match Athey–Imbens and the Methodology Registry.
• The committed PDF is 11 pages and contains the required visible qualifiers.
• Manual replication of all numeric claim-sync checks passed.

Methodology

• Severity: P3 — informational, mitigated
  Impact: Bootstrap inference, fixed-95% sup-t bands, and covariate solver differences are documented deviations, not defects. The carousel accurately labels them at carousel/generate_cic_carousel.py:L929-L931, carousel/generate_cic_carousel.py:L1012-L1015, and carousel/generate_cic_carousel.py:L1243-L1246, consistent with docs/methodology/REGISTRY.md:L4136-L4148.
  Concrete fix: None required.

The Theorem 3.1 composition and scale-equivariance scope also match docs/methodology/REGISTRY.md:L4115-L4134.

Code Quality

No findings.

Performance

No findings. Generation costs are confined to the offline carousel builder.

Maintainability

No findings. Numeric claims are centralized and synchronized with Tutorial 27 through tests/test_cic_carousel_claims.py:L88-L197.

Tech Debt

No findings. Existing CiC limitations remain documented in TODO.md:L115-L119 and are not expanded by this PR.

Security

No findings. No credentials, private paths, or sensitive metadata were found.

Documentation/Tests

No findings. Tests cover the headline effects, QTE table, uniform-band split, covariate example, equivariance receipt, DGP constants, and visible qualifiers.

pytest and carousel dependencies were unavailable in the review environment; equivalent claim-sync checks were reproduced directly and passed.