Your dbt models are about to get Tickered - Sanity Check brief
The question
When LLMs make dbt code (and code-as-output generally) abundant, where does the value flow? Packy McCormick's restated Christensen claim says it flows to the adjacent layer - the judgment that decides what to model, against which outcomes, and how to verify. The Sanity Check angle this brief works out is: that adjacent layer is exactly what MAC (Model Acceptance Criteria) is selling. The conviction-check the founder needs to clear before drafting: Packy's Vertical Integrator framing assumes ONE VI played hard; RDCO is portfolio-shaped. The piece either has to defend the portfolio or re-shape the bet.
What we already know (from the vault)
- [[2026-04-30-not-boring-scarce-assets-abundance-driven-scarcity]] - Packy's Macro vs Micro scarcity taxonomy. The load-bearing move is the Christensen restatement: "when modularity / commoditization kills profits at one stage, attractive profits emerge at an adjacent stage" (Spolsky: "smart companies commoditize their complements"). His Federico da Montefeltro example - handwritten codices spiking in value the moment Gutenberg made print abundant - is the cleanest historical analogue for dbt-code-becoming-cheap. Already names "Your dbt models are about to get Tickered" as the candidate SC angle in section 4 of the vault note.
- [[2026-04-30-rdco-thesis-targeting-systems-feedback-loops]] - the canonical RDCO thesis (founder's articulation, 2026-04-30 16:01 ET). Four layers: targeting system + instrumentation + tools + recursive feedback loop. MAC is named explicitly as "a productized targeting system for data modeling." This is the doc the SC piece anchors to.
- [[2026-04-24-targeting-system]] - the concept page. RDCO-canonical term for "the thing that defines what good means for a niche." Christensen's "adjacent stage" maps 1:1 onto this.
- [[2026-04-24-three-decision-algorithms]] - already states the move out loud: "Not 'better data quality' - the implicit-to-agentic targeting bridge for data modeling, built on the premise that the old cost structure that let senior operators say 'trust me, it's fine' just collapsed."
- [[2026-04-29-every-compute-is-new-cash]] and [[2026-04-30-stratechery-amazon-earnings-trainium-commodity]] - the demand-side and supply-side companions to Packy's macro framework. Triad confirmation: per-seat dies, commodity-cost-structure beats moat, the adjacent (judgment) layer becomes scarce.
- [[2026-04-23-unhobbling]] - directly answers Kingsbury's dissent objection ("the verification layer is itself LLM-contaminated") by saying acceptance criteria must be human-authored. This is the load-bearing rebuttal the piece needs in its back pocket.
What the web says
- Steven Alber, Planet Pulsar, "The Verifier Layer: As AI Writes the Code, Human Judgment Becomes the Scarce Asset" (2026-04-19). The piece's subtitle nails the inversion: "In the verifier economy, output is abundant. Responsibility is scarce." Paywalled body, but the title + subtitle are the cleanest external articulation of the angle the SC piece would land. Worth citing as evidence that the framing is now ambient in the discourse, not RDCO-original [[https://www.planetpulsar.com/the-verifier-layer-as-ai-writes-the-code-human-judgment-becomes-the-scarce-asset/]].
- Reliable Data Engineering, Medium, "AI-Generated dbt Models Are Actually Good Now (I Tested 50 of Them)" (2026-01-26). Concrete numbers - 76% of Claude Sonnet 3.5-generated dbt models worked first try, 22% needed minor fixes, 1% failed entirely. Remaining failure categories cluster exactly where judgment matters: incremental logic edge cases (30% failure rate), NULL handling in aggregations (40%), missing partition / cluster optimization (70%), incomplete test coverage (60%). Author's own conclusion: "Human review still essential; domain expertise still required" [[https://medium.com/@reliabledataengineering/ai-generated-dbt-models-are-actually-good-now-i-tested-50-of-them-b87bd82bc7c2]].
- dbt Labs, "A new era of data engineering: dbt Copilot is GA." dbt Copilot is now generally available, pitched as automating documentation, tests, semantic modeling, and SQL formatting. Confirms the supply curve flattening - dbt Labs itself is commoditizing the code-writing layer of its own product [[https://www.getdbt.com/blog/dbt-copilot-is-ga]].
- dbt Labs blog, "How AI helps with data modeling." Same direction: AI generates first drafts from natural language descriptions, suggests naming conventions and folder structures. dbt's own pitch is now "we generate the code; you decide what's right" [[https://www.getdbt.com/blog/ai-data-modeling]].
- Industry baseline, multiple sources. AI accounts for ~42% of committed code, but 96% of developers do not fully trust it functionally correct, only 48% always check it before committing. The verification gap is the bottleneck, not the generator [[https://www.qodo.ai/blog/building-the-verification-layer-how-implementing-code-standards-unlock-ai-code-at-scale/]].
- O'Reilly, "Comprehension Debt: The Hidden Cost of AI-Generated Code." "When you write code yourself, comprehension comes with the act of creation. When the machine writes it, you'll have to rebuild that comprehension during review." This is the data-team-specific failure mode the SC piece can name out loud - dbt projects accumulate untested model graphs nobody fully understands [[https://www.oreilly.com/radar/comprehension-debt-the-hidden-cost-of-ai-generated-code/]].
- Packy McCormick, "Vertical Integrators" / "Vertical Integrators: Part II". Quoted thesis: "There will be a SpaceX and Anduril in every big category led by sclerotic incumbents. These Vertical Integrators will be much bigger than the incumbents they replace." Implication for the portfolio question is direct: Packy's bet IS single-VI-played-hard [[https://www.notboring.co/p/vertical-integrators-part-ii]].
Convergences and contradictions
Strong convergence (4 sources pointing the same way):
- Packy (macro framework) + Alber (verifier economy) + Reliable Data Engineering (concrete dbt numbers) + dbt Labs (own product confirming commoditization) all describe the same phenomenon - code generation moves down the value stack, judgment / verification moves up. The dbt-specific landing is now empirically documented, not speculative.
- The 70% failure rate on partition / cluster optimization and 60% failure rate on test coverage map directly onto MAC's two strongest claims - that what's missing in AI-generated dbt isn't syntax, it's the targeting system that says "what passes for production in our warehouse."
One soft contradiction the piece must handle:
- Packy's single-VI framing vs RDCO's portfolio shape. Already named explicitly in the [[2026-04-30-not-boring-scarce-assets-abundance-driven-scarcity]] note's section 2 caveat. Worth the founder's explicit conviction-check before draft, see the dedicated section below.
One subtle reinforcement:
- The Three Decision Algorithms vault note ([[2026-04-24-three-decision-algorithms]]) already articulates the move - "the old cost structure that let senior operators say 'trust me, it's fine' just collapsed." That's the load-bearing sentence the SC piece can build on. RDCO got there independently of Packy. The piece is not re-stating Packy; Packy is the evidence that the macro pattern (which RDCO had already landed in its own vocabulary) is real.
The Sanity Check landing (the load-bearing section)
The piece is not "code generation is getting cheap." That's a Packy / Alber / Every / Stratechery restate and fails the no-derivative bar [[feedback_no_derivative_sanity_check_pieces]]. The piece is:
"Your dbt models are about to get Tickered."
Packy's "Tickered" framing - that the same asset feels scarce as a startup and boring once it's screener-sortable next to 248 competitors - is the load-bearing analogy. When an AI agent can spin up a dbt model that's 76% production-ready first try, every dbt model in your repo just got Tickered. The model itself - the SQL, the schema, the materialization choice - is screener-sortable. What is NOT Tickered: the explicit criteria for what makes your model production-ready in your warehouse against your downstream consumers' decisions. That criteria layer is the Federico da Montefeltro handwritten codex of data engineering. It's what MAC names, formalizes, and turns into a product.
The non-derivative move is the inversion: every data leader is being told "AI will write your code." The actual problem this introduces is that the judgment about what's correct now has to be written down before generation, not implicitly after. You can't verify what you can't articulate. The piece's promise to the reader is to make this concrete - here's the failure mode (untested model graphs proliferating, every model 76% right but nobody can say what the other 24% is), here's the fix (MAC's targeting-system-first discipline), here's the bigger frame (Packy's Christensen restatement explains why this is happening now, across every code-as-output discipline at once).
The structural move the piece makes: take Packy's macro framework, plant the flag specifically in dbt / analytics engineering territory, and use MAC as the worked example of what the "judgment layer" looks like when it ships. The reader leaves with three things: (1) a frame for the abundance-of-code-generation moment that is theirs, not borrowed from VC discourse; (2) a vocabulary - targeting system, acceptance criteria, Tickered - they can use in their own engineering reviews; (3) a credible RDCO-shaped pointer (MAC) without the piece being a MAC ad.
What makes it non-derivative: Packy never names dbt. Alber never names MAC. dbt Labs never names the judgment-layer-as-scarcity move. The synthesis - macro framework + concrete data-engineering territory + MAC as the worked product expression - is RDCO-originated and lands in RDCO's actual expertise (data discipline), not borrowed from anyone else's territory.
Portfolio-vs-VI tension (founder conviction-check)
Packy's Vertical Integrator thesis is single-bet: "There will be a SpaceX and Anduril in every big category." The bet is one VI per founder, played hard, owning the whole stack in one vertical. The reward is the macro-scarcity premium - Anduril stock is scarce because Anduril is the only credible defense bet.
RDCO is structurally a portfolio of small vertical integrators: Squarely owns the indie-puzzle-shop loop, MAC owns the mid-market data-modeling loop, Sanity Check owns the operating-discipline-content loop. Each is the four-layer pattern (targeting + instrumentation + tools + feedback) applied to a different niche. The portfolio gives up macro-scarcity (no single bet is the category-defining play) for niche-scarcity in many places at once.
The founder needs to choose one of two framings before drafting:
Option A - Defend the portfolio explicitly. The argument: the pattern (four-layer targeting system) is the asset, not any individual bet. The portfolio is a hedge on which niche compounds first AND a way to keep the pattern sharp by stress-testing it against multiple niches. The macro-scarcity Packy describes accrues to whoever proves the pattern is repeatable. SpaceX is one bet, but Musk's pattern (vertical integration + first-principles engineering + cadence) is itself an asset that compounds across SpaceX / Tesla / xAI / Neuralink. RDCO is making the same shape of bet at a smaller scale.
- Risk: this is half a step from "Musk-comparison hubris." Has to be argued from the four-layer-pattern abstraction, not from "RDCO is the Berkshire of targeting systems."
- Strength: it's an honest answer the founder can stand behind in iMessage if challenged on it.
Option B - Reframe RDCO's bet as a different shape than VI. The argument: RDCO is not a Vertical Integrator. It's an operating-system vendor - the targeting-system-pattern is the OS, each bet is an app written against it. The macro-scarcity isn't in any one bet; it's in being one of the few teams articulating the pattern publicly while the rest of the industry is still talking about "AI agents" generically. Sanity Check is the platform-evangelism layer; MAC and Squarely are the reference applications. The portfolio shape is the bet's defensibility.
- Risk: harder to fundraise against, but RDCO isn't fundraising.
- Strength: it's more honest to what RDCO actually is, less stretched.
Founder's call. Either lands cleanly. Option B is closer to what RDCO actually does today; Option A is closer to the language Packy's readers will pattern-match on. Recommend the brief surface both and let the founder pick before draft - because the piece's middle section reads materially differently depending on which framing wins.
Compute-economics triad - the 3-source convergence
Three sources, all Apr 29-30, describe the same phenomenon from three angles:
- Demand side - [[2026-04-29-every-compute-is-new-cash]] (Every, Laura Entis). Per-seat / flat-subscription pricing structurally incompatible with agent workloads. GitHub moved to token-based billing. Anthropic shifted enterprise to usage-based. The "millennial lifestyle subsidy" of cheap subscriptions is ending. Pricing is reorganizing around metered compute.
- Supply side - [[2026-04-30-stratechery-amazon-earnings-trainium-commodity]] (Stratechery, Ben Thompson). Two ways to build durable profit: Apple's moat model, or the commodity-cost-structure model where the market-clearing price is set by the worst-cost supplier. AI is a commodity market with demand exceeding supply, so the low-cost provider earns commensurate profit. Trainium's 30% better price-performance translates into margin advantage.
- Macro framework - [[2026-04-30-not-boring-scarce-assets-abundance-driven-scarcity]] (Not Boring, Packy McCormick). Christensen restated: when modularity / commoditization kills profits at one stage, attractive profits emerge at an adjacent stage. The macro chassis under both Entis's pricing pivot and Thompson's commodity dynamics.
Why the convergence matters for the SC piece. Three independent, credible sources, three different angles, same conclusion: code generation and the compute it runs on are both becoming commodity. The adjacent layer (judgment, targeting, verification) is where the premium goes. The SC piece doesn't need to argue the macro frame - it can stand on the triad as established premise and spend its word count on the dbt-specific landing. The triad is the reader's permission structure to take the targeting-system claim seriously; without it, the piece reads like RDCO inventing convenient theory.
Triad citation move in the piece itself: name the three sources in a single short paragraph, lean weight on Packy's framework, then turn immediately to the dbt-specific claim. The piece should not try to re-explain the triad - assume it as backdrop and move.
Suggested headline + lead
Headline options (in order of recommended use):
- "Your dbt models are about to get Tickered." (Founder's coinage. Mystery in the verb. The piece IS the explanation. Best for SC's earned-curiosity reader.)
- "The handwritten dbt model." (Federico da Montefeltro reference. Quieter, more crafted. Works if the piece leans into the Christensen / abundance-scarcity history.)
- "What dbt Copilot can't write." (Most direct, most SEO-friendly, lowest-curiosity-tax. Probably best as the email subject line even if the article uses #1 as title.)
Lead paragraph candidate (1-2 sentences, plain register, no mini-essay declarative):
Your next dbt model is going to be written by an LLM. 76% of them already work first try. The interesting question isn't whether AI writes the code - it's who decides whether the model is actually right, and what "right" even means in your warehouse.
That lead does three things: opens with a fact (concrete, citable), pivots immediately to the verb the piece wants to land on (decides), names the missing thing (acceptance criteria) without using jargon. Matches the X-voice cadence - 3 sentences, last one earns the read, no semicolons, no em dashes.
Open follow-ups
- What does the MAC reference implementation against an AI-generated dbt model look like as a worked example? The piece would benefit from one concrete walked-through example - a generated model, the MAC criteria applied, the failure surfaced, the human judgment captured. Likely needs a 1-hour engineering session before draft.
- Does Packy's "Tickered" framing apply to Sanity Check itself? When AI-generated data newsletters proliferate, what makes this newsletter the handwritten codex? Worth surfacing in the closing graf as a tongue-in-cheek meta beat (the founder's voice, the cross-niche pattern, the work-in-public layer). Don't write the meta-beat as analysis; write it as a wink.
- Is there a Christensen-vault paper the piece should anchor on directly, not via Packy? Right now the piece quotes Packy quoting Christensen. One layer of remove. If there's a clean Christensen source on Conservation of Attractive Profits that's accessible, citing it directly would tighten the piece by 200 words. Low-priority - Packy's restatement is sufficient.
- What's the LinkedIn / X distribution shape for this piece? The headline "Your dbt models are about to get Tickered" works on X (curiosity, specific noun, the verb-as-mystery). On LinkedIn it may need a less-jargony subtitle. Worth pre-planning before draft.
- Does the piece end on MAC (the product) or on the pattern (targeting systems)? Defensible either way. Ending on MAC makes the CTA crisp but reads more salesy. Ending on the pattern keeps Sanity Check's lighthouse-not-megaphone posture intact. Recommend the latter; have the founder confirm before draft.
Sources
Vault canonical:
- [[2026-04-30-not-boring-scarce-assets-abundance-driven-scarcity]] - Packy McCormick, Not Boring, 2026-04-30. The macro framework. Already names the candidate SC angle in section 4.
- [[2026-04-30-rdco-thesis-targeting-systems-feedback-loops]] - founder's canonical thesis articulation, 2026-04-30.
- [[2026-04-29-every-compute-is-new-cash]] - Laura Entis, Every, 2026-04-29. Demand-side compute-economics piece.
- [[2026-04-30-stratechery-amazon-earnings-trainium-commodity]] - Ben Thompson, Stratechery, 2026-04-30. Supply-side compute-economics piece.
- [[2026-04-24-targeting-system]] - RDCO concept page.
- [[2026-04-24-three-decision-algorithms]] - RDCO concept page; already states the MAC-as-judgment-bridge move.
- [[2026-04-23-unhobbling]] - RDCO concept; the Kingsbury rebuttal on human-authored acceptance criteria.
Web:
- Steven Alber, "The Verifier Layer: As AI Writes the Code, Human Judgment Becomes the Scarce Asset" - Planet Pulsar, 2026-04-19. [[https://www.planetpulsar.com/the-verifier-layer-as-ai-writes-the-code-human-judgment-becomes-the-scarce-asset/]]
- Reliable Data Engineering, "AI-Generated dbt Models Are Actually Good Now (I Tested 50 of Them)" - Medium, 2026-01-26. [[https://medium.com/@reliabledataengineering/ai-generated-dbt-models-are-actually-good-now-i-tested-50-of-them-b87bd82bc7c2]]
- dbt Labs, "A new era of data engineering: dbt Copilot is GA." [[https://www.getdbt.com/blog/dbt-copilot-is-ga]]
- dbt Labs, "How AI helps with data modeling." [[https://www.getdbt.com/blog/ai-data-modeling]]
- O'Reilly, "Comprehension Debt: The Hidden Cost of AI-Generated Code." [[https://www.oreilly.com/radar/comprehension-debt-the-hidden-cost-of-ai-generated-code/]]
- Qodo, "Building the Verification Layer." [[https://www.qodo.ai/blog/building-the-verification-layer-how-implementing-code-standards-unlock-ai-code-at-scale/]]
- Packy McCormick, "Vertical Integrators: Part II" - Not Boring. [[https://www.notboring.co/p/vertical-integrators-part-ii]]
Related
- [[2026-04-30-not-boring-scarce-assets-abundance-driven-scarcity]]
- [[2026-04-30-rdco-thesis-targeting-systems-feedback-loops]]
- [[2026-04-29-every-compute-is-new-cash]]
- [[2026-04-30-stratechery-amazon-earnings-trainium-commodity]]
- [[2026-04-24-targeting-system]]
- [[2026-04-24-three-decision-algorithms]]
- [[2026-04-23-unhobbling]]
- [[feedback_no_derivative_sanity_check_pieces]]
- [[2026-04-15-mac-anchor-article-draft-v1]] - the existing MAC anchor article; this SC piece is the macro-framework upstream companion