Unified deck: Alexander's review rounds (opening graph, case 3 fleshed out, trims to 55 presented slides, The END)

[AlexanderFengler]

Four review rounds on apps/unified_slides.html, applied to both the canonical src and the built deck (this checkout has no book/build shards, so every edit was mirrored byte-identically into the built file — none touch tokens or DATA; the next shard-full make html-unified reproduces everything from src).

Opening

• S2 "PyMC Labs: what we do": Labs logo chip top-right (dark-mode safe) + a left-to-right ecosystem graph: PyTensor → PyMC → {PyMC-Marketing, CausalPy, Bespoke (dashed: customer-specific builds)}; "what we sell / how we work" bullets rewritten; the method note now reads "Our lectures will center around classical methods to stay didactically coherent…".

Part 1 (the cases)

• Colgate: "recovered" explained plainly (planted 50% → 49–59%, find a known answer before being trusted); "Bayesian" dropped from the method line; clearer sub.
• HelloFresh: the everyday-tool card now names the MMM (it was used undefined); the loop figure loses the synthetic-control node (not yet taught at that point) and closes on "that method is Part 2 of this session".
• Nürnberger fleshed out to a two-slide case: a new card slide defines last-touch attribution, the GDPR (spelled out), and the funnel-aware causal MMM before the poll; the poll is now concept-grounded ("which channels does last-touch systematically under-credit?") with per-option why-notes; the 27% CPL outcome moved into the reveal.

Part 2 (geo) trims

• Boardroom slide gains a schematic metros picture (29 grey donors, the treated metro with campaign arcs); the "unbiased" fold dropped.
• Simulate-the-world: legend fixed ("Black: the treated metro"), before/after means on by default (also on the data slide).
• "The estimator" deleted; "What the constraint buys" simplified to "The synthetic twin, graded against the truth" (OLS synthetic + n_eff removed).
• Act III reduced to the two falsification slides → straight into the money act; placebo-in-space, test inversion, the Act-III poll, and the statistics summary all moved to Backup.

Part 3 (IV) trims

• "The plan for Part 2" → "Going forward"; the machine-learning line dropped; "Why the division is forced", "The IV estimate", and "Weak instruments" moved to Backup. Reviewer note: worth one live click-through of the reduced-form → LATE handoff, since the on-screen division now lives in the appendix (later slides still cite €16.5 via tokens).

Closing

• Provenance slide logo row: PyTensor · PyMC · PyMC-Marketing · CausalPy (PyTensor vector logo vendored from pymc-devs/pytensor).
• "One breath" → a title-style "The END": both authors' contacts, the breadth message (Bayesian inference as a vast toolset beyond causal inference), and a get-in-touch callout.
• "The pattern in every engagement" moved to Backup.

Deck: 55 presented slides + Sources + 11 backup (67 total). Verified each round: labs claims + deck-wide sweeps green on both decks (90/90), node --check OK, headless DOM 0 JS / 0 MathJax errors, polls exercised end-to-end in a real browser, every changed slide screenshotted (light + dark spot checks). One shared-claims tweak: colgate_method relaxed to the fragment both decks contain ("interrupted time series").

Also carries the labs-deck scratchpad tooling commit (assembler + fragments) from the earlier branch.

🤖 Generated with Claude Code