marketing: MMM calibration launch carousel (Signal palette) + claims-sync guard

[igerber]

Summary

• New LinkedIn launch carousel for the diff_diff.mmm interop: carousel/generate_mmm_carousel.py renders carousel/diff-diff-mmm-carousel.pdf (10 slides, 4:5), following the generate_cic_carousel.py architecture with a new "Signal" palette (violet = calibrated, coral = uncalibrated, amber reserved for ground truth in every chart).
• Ease-first narrative with Tutorial 30 (Meridian geo launch) as the spine and Tutorial 29 (PyMC-Marketing spend boost) as the secondary beat: three-line handshake, the verbatim to_code() snippet, the default-vs-calibrated payoff interval chart (3.29 [3.02, 3.57] vs truth 2.5 -> 2.54 [2.47, 2.62]), and a capabilities slide.
• tests/test_mmm_carousel_claims.py (18 tests, normal suite): every deck number is located on the committed executed notebook surfaces; slide-4 snippet lines must be byte-verbatim from the committed to_code() output; the confidently-wrong interval structure and 6x-narrower ratio are recomputed from the deck's own constants; slide-8 call arguments re-derive from tutorial 29's DGP; the four named aggregate('total') adopters are pinned against each results class's _AGGREGATE_SUPPORTED; the "Zero new dependencies" strip is backed by an mmm.py import scan plus a pyproject.toml requirement scan; the version strip derives from the tutorials' own requirement pins; competitive claims and unqualified guardrail absolutes are ban-listed.
• Tracking: TODO.md gains a deck-agnostic PDF-vs-generator drift-guard row; DEFERRED.md's decision record documents the marketing-surface claim posture (scoped guardrail copy; slide-8 linearity scoping lives in tutorial 29, which the CTA points to).

Methodology references (required if estimator / math changes)

• Method name(s): N/A - no methodology changes. Deck copy about to_meridian_roi_prior / meridian_calibration_mask / to_pymc_marketing_lift_test / aggregate('total') states committed-tutorial facts only.
• Paper / source link(s): N/A
• Any intentional deviations from the source (and why): None. The marketing-surface claim posture (scoped, never absolute) is recorded in DEFERRED.md's decision record.

Validation

• Tests added/updated: tests/test_mmm_carousel_claims.py (new, 18 tests; skips cleanly when carousel/ or docs/ is absent, matching test_cic_carousel_claims.py).
• Backtest / simulation / notebook evidence (if applicable): every number on the deck is a committed, seed-locked value from the executed tutorials 29/30; the committed PDF is regenerated from the final generator source in this commit.

Security / privacy

• Confirm no secrets/PII in this PR: Yes

[github-actions]

Overall assessment

⚠️ Needs changes — one unmitigated P1 API/methodology-contract mismatch.

Executive summary

• No estimator, weighting, variance, or inference implementation changes.
• Tutorial-derived numerical claims and MMM formulas are consistent with the registry.
• All 18 claim-sync assertions passed via a lightweight runner.
• The “Paste and Run” slide shows an invocation that necessarily raises ValueError.
• PDF drift is already tracked in TODO.md.

Methodology

P1 — [Newly identified] Required Meridian scope arguments are omitted

• Location: carousel/generate_mmm_carousel.py:L831-L845
• Impact: The slide presents prior.to_code() under “Paste and Run,” but channel and time scope are mandatory. Calling it exactly as shown raises ValueError; the registry and exporter explicitly require channel/media_channels plus roi_calibration_period or full_model_window=True (docs/methodology/REGISTRY.md:L6577-L6580, diff_diff/mmm.py:L815-L903).
• Concrete fix: Display the tutorial’s actual invocation:

prior.to_code(
    channel="search",
    media_channels=["search", "tv"],
    roi_calibration_period=mask,
)

The tracked omission of slide-8’s linearity qualifier is P3-informational under DEFERRED.md:L137; the shown tutorial calculation is valid for its linear, homogeneous DGP.

Code Quality

No additional findings.

Performance

No findings.

Maintainability

No findings.

Tech Debt

P3 — PDF/source drift guard is tracked

• Location: TODO.md:L88
• Impact: Future generator edits could leave the committed PDF stale.
• Concrete fix: None required for this PR; the work is properly tracked.

Security

No findings. No secrets or suspicious PDF metadata were identified.

Documentation/Tests

P2 — Claim guard does not validate the displayed invocation

• Location: tests/test_mmm_carousel_claims.py:L156-L160
• Impact: The suite validates generated output lines but allows the invalid prior.to_code() call to ship.
• Concrete fix: Pin the complete displayed invocation to the tutorial source or validate it against MeridianROIPrior.to_code’s required scoping contract.

Validation: both Python files compiled successfully; all 18 claim assertions passed; the committed PDF reports 10 pages. Full pytest was unavailable in the review environment.

Path to Approval

1. Replace the bare prior.to_code() display with the scoped tutorial invocation.
2. Add a regression assertion that the displayed call includes channel and time scope.

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: c2660fc875891f2d61d3b1610a9d106bd24c8752

---

Overall assessment

✅ Looks good — the prior P1 is resolved; no unmitigated P0/P1 findings remain.

Executive summary

• Slide 4 now shows the required channel and time scope, matching the tutorial and exporter contract.
• The committed PDF contains the corrected invocation.
• Regression coverage rejects a bare prior.to_code() call.
• No estimator, weighting, variance, inference, or default behavior changed.
• All 19 claim assertions passed via a lightweight runner; both Python files compile and the PDF has 10 pages.

Methodology

The prior P1 is fixed at carousel/generate_mmm_carousel.py:L835-L845 and guarded at tests/test_mmm_carousel_claims.py:L162-L179, consistent with docs/methodology/REGISTRY.md:L6577-L6580.

P3 — Documented marketing-surface scoping

• Impact: Slide 8 omits an on-slide linearity qualifier, but this posture is explicitly recorded at DEFERRED.md:L137 and the CTA points to the scoped tutorial.
• Concrete fix: None required.

Code Quality

No findings.

Performance

No findings.

Maintainability

No findings.

Tech Debt

P3 — PDF/source drift guard is tracked

• Impact: Future generator edits could leave committed PDFs stale.
• Concrete fix: No PR action required; the deterministic PDF/manifest guard is tracked at TODO.md:L88.

Security

No findings. No secrets or unsafe externally controlled operations identified.

Documentation/Tests

No findings. The new regression directly covers the previously invalid displayed invocation. Full pytest was unavailable in the review environment; the claim tests were executed through a lightweight direct runner.