💥 Where the loop breaks

A loop sees only the metric you gave it. If you said “reduce CAC” (customer acquisition cost — how much it costs, on average, to bring in a new customer), the loop will find all the ways to bring this number down, including those that betray its real purpose: making a campaign cheaper by attracting people who will never buy, cutting post-sale support, promising more than it delivers. It’s not an execution flaw — it’s the loop doing exactly what you told it to, using a number that has become detached from the reality it’s supposed to represent.

The loop pursues a reference—the target you defined—and nothing inside it can ask whether that reference is right. A thermostat keeps the temperature at 20 degrees with impeccable precision; it will never ask itself whether 20 is the right temperature for that room on that day. It only knows how to measure the distance to 20 and act to close that gap. Questioning its own goal requires someone outside of the loop—because from the inside, the objective is the only ground there is.

A speed loop fights with a quality loop. A hiring loop puts pressure on a culture loop. Each one, examined alone, passes the test — it’s doing exactly what was requested. The problem appears only when you look at both together. It's like two poorly configured air conditioners in the same room: one cools because the temperature went up, the other heats because the temperature went down—both “working,” while the environment is unstable.

What to look at: the three loops are identical on both sides — the only difference is the two red lines on the right. They represent the real pressure that "moving faster" and "hiring faster" put on "maintaining quality." A single-loop mindset has no vocabulary to describe this line — it only exists when you draw all three loops in the same graph and connect what interferes with what.

Sensors go out of calibration. Data pipelines rot—a field changes its name, a filter stops running, and nobody notices. The metric definition changes behind the scenes (someone redefines "active user" and forgets to tell anyone). Verification gradually slips from checking reality to just comparing paperwork: the number matches the previous number, so "everything is fine." But both numbers could be equally wrong.

In a single-node loop, everything happens in the same context window — the body of text the model can see at once, along with the instructions, what has already been done, and what remains. Raw HTML from a search, a half-finished draft, reasoning from two iterations ago: all floating together when it’s time to decide the next step. That’s what produces fake review (the verifier approves because it’s in the same emotional context as the person who wrote it—covered in module 2.2) and “forgetting” the original instruction, which is still there, just buried in noise.

A loop reads sources one at a time because a loop is, by definition, one thing after another: discover, plan, act, verify, and repeat. It can’t open ten work streams at the same time and combine the results later — not because it lacks intelligence, but because its very structure is a queue. This is the single of the six limits in this module that no better prompt, sharper verifier, or amount of discipline can solve.

What to look at: the four boxes are the same on both sides — what changes is the shape. On the left, each box starts only when the previous one finishes; that’s how a loop works by definition. On the right, the same work opens up into fan-out (a node that launches several workstreams at the same time) and then converges at a join node (fan-in). This isn’t a prompt improvement — it’s a different structure.

Audite o loop em <caminho do arquivo ou descricao do processo>.
Responda em quatro blocos, sem suavizar:

1. METRICA — liste a metrica principal que este loop otimiza.
   Depois liste 3 formas concretas de mover essa metrica que
   trairiam o proposito real do sistema (Goodhart).

2. CEGUEIRA — diga quem, hoje, fica FORA deste loop e poderia
   questionar se o alvo dele ainda faz sentido. Se a resposta
   for "ninguem", diga isso claramente.

3. COLISAO — aponte que OUTROS loops do meu sistema esta
   metrica pressiona ou de quem ela rouba recurso/atencao.

4. MEDICAO — aponte onde a medicao deste loop pode estar
   decaindo sem alarme (sensor, pipeline, definicao mudada).

Nao me diga que esta tudo bem sem antes tentar ativamente
achar um problema em cada um dos quatro blocos.