The almost-free tier

Four of Cadence's ten workloads - yesterday's DAU, monthly revenue by plan, cards that failed three times, the weekly reconciliation - passed both questions from session 1. The logic is writable, the definitions are aligned. That makes them deterministic, and deterministic work has a superpower no probabilistic tool can buy: you verify it once, and then it is simply correct, every run, forever.

This session exists because that pile is where routing discipline pays fastest. In the simulator, moving just these workloads off the LLM and onto queries is a big part of how the true monthly cost falls from $2,315 at baseline to $224 with the free tier on - while trusted answers go from 0 out of 10 to 4 out of 10. Cheaper and more trustworthy at the same time is not a trade-off. It is what fit looks like.

• Almost free at the margin. Building the query and arguing about the definition costs real hours. Running it again costs cents, and often literally nothing.
• Free of verification debt. A verified query does not need an eval set, a monitoring owner, or a cleanup budget. Those line items exist only above this tier - session 5 prices them.
• Not free of maintenance. Definitions drift and dashboards sprawl. Part 3 is honest about that, because a course that only flatters its favorite tier is an advertisement.

BigQuery's on-demand price is $6.25 per TiB scanned, and the first TiB every month is free. A 10 GB scan is 10/1024 of a TiB, so about six cents - and in most real months it lands inside the free TiB and costs nothing at all.

Now push the same 10 GB through a frontier model as input. Text runs roughly four characters per token, so 10 GB is on the order of 2.5 billion tokens. At around $5 per million input tokens, that is $12,500 and up in input alone, before the model has said a word back. Same bytes, same question, five orders of magnitude apart. All figures as of Sept 2026 - prices move, so date-stamp yours too.

The price gap is the headline, but the trust gap is the lesson. A query is correct by construction: the logic was checked against the data once, and the same logic runs every time after that. Anyone who can read SQL can audit it. When it breaks, it breaks loudly - a failed job, an empty table - not fluently.

A probabilistic answer, however good, has to be re-earned on every run. That is what an eval set is: rent, paid monthly, on trust the deterministic tier gets for free. This is why the routing test from session 1 asks about verifiability before anything else - it is the single property that decides whether trust is a one-time cost or a subscription.

Cassie Kozyrkov, Google's former chief decision scientist, draws the boundary this course routes by: do not use machine learning for problems that are "defined by human-made rules in the first place". If a human wrote the rule, a human can write the code - directly, checkably, once. ML earns its place only when you cannot articulate the rules yourself and you have the data to learn them from.

Her broader point is that reaching for the fancy tool first is not ambition, it is a tax: the simplest approach that actually works is the one that ships fastest, audits cleanest, and leaves budget for the problems that genuinely need learning. Cadence's deterministic pile sits entirely on the human-made-rules side of her line - every one of those four workloads has a definition somebody wrote down in a meeting.

Rogati's layers, bottom to top: collect, move/store, explore/transform, aggregate/label, learn/optimize, and AI at the tip. Her point is not that the tip is bad - it is that the tip is self-actualization, and self-actualization on an empty stomach is called hallucinating. An org that cannot agree what "active user" means is not one clever model away from anything.

The pyramid is also a budget map. The lower rungs are exactly the almost-free tier from Part 1: verified once, run forever. Every dollar of trustworthy AI at the tip is standing on rungs somebody funded and maintained. Skipping them does not save the money - it moves the cost into cleanup, which is the most expensive line in session 5's true-cost math.

• Collect and move/store are healthy if one system of record exists per fact and nobody exports to spreadsheets to "fix" the numbers.
• Explore/transform is healthy if metric definitions live in code that is reviewed, not in the heads of the two people who have been there longest.
• Aggregate/label is healthy if the dashboards people actually open answer the questions people actually ask - Part 3 covers what happens when they drift apart.
• A hole in any rung is a routing fact. It does not mean "no AI ever". It means the AI above the hole inherits the hole, plus interest.

Cadence has 500 support tickets a month and a leader who wants to know what customers are angry about. The free-tier reflex is a keyword count: tally "crash", "sync", "billing". And it fails in a specific, predictable way: keyword counts miss every paraphrase. The customer who writes "the app eats my notes" is describing data loss, and the string "data loss" appears nowhere in the ticket. The count comes back confident, complete-looking, and wrong about exactly the tickets written by the angriest, least jargon-fluent customers.

This is a genuine no on Q1 from session 1: there is no writable rule for meaning. Language work belongs on the language tier, and session 4 prices what doing it properly costs and earns.

• Rules impersonating judgment. A 300-line regex for extracting action items from meeting notes is not the free tier - it is judgment work in a deterministic costume. It carries the maintenance burden of code and the accuracy of a guess, which is the worst seat in the house.
• Dashboard sprawl. Every unanswered question spawns a tile; nobody deletes tiles. The DAU incident happened partly because the right tile existed and nobody could find it among the forty others. A dashboard nobody opens is a query nobody verified recently.
• Stale definitions. The query stays correct against the definition it encodes - while the business moves the definition in a meeting the data team was not in. Verified once is a superpower with one clause: it holds until the definition changes, so definition changes must reach the code.

Honest routing cuts both ways. The free tier deserves the deterministic pile - and it has to keep earning it with maintenance, pruning, and definitions that stay aligned.