# Calibration > Develop judgment through repeated contact with outcomes you can inspect. **Curriculum phase:** Practice **Lesson:** 05 **In this lesson:** - How calibration — the felt sense of what works — develops through cycles - Why calibration shifts when you switch models, change domains, or grow - The signs of good and poor calibration - How to sharpen calibration faster through deliberate practice You've built the environment — the mycelium. Now: how do you develop the judgment to know what's working, what isn't, and how to adjust? That's calibration. --- ## Definition Calibration is **the developing sense of what works, what doesn't, and how to adjust.** It's not a skill you learn once and keep forever. It's a felt sense that evolves through cycles. Through experience. Through paying attention. After enough cycles, you start to know — without thinking about it — whether the AI understood you. Whether you're on track or drifting. Whether to push forward or change direction. That knowing is calibration. --- ## How Calibration Develops Calibration doesn't come from reading about it. It comes from doing it. **Early cycles (1–20):** You're feeling your way. Everything is learning. You don't know what works yet, so you try different approaches. Some land, some don't. **Pattern recognition (20–100):** Patterns emerge. You notice: this kind of input produces this kind of output. This approach works for this domain. You start to anticipate what the AI will do before it does it. **Instinct (100+):** You move with confidence. Adjustments are small and precise, not big pivots. You can articulate what works and why. You've forged a blade. But calibration is never done. It's always refining. --- ## Why Calibration Changes Calibration is contextual. It shifts when: - **You switch models** — Claude works differently than GPT, which works differently than others. Your calibration has to adjust. - **The model updates** — a new version is better in some ways, different in others. Your calibration is off until you cycle through it again. - **You change domains** — calibration for writing code is different from calibration for writing essays. - **You grow** — after 1,000 cycles, your standards are higher. You're more discerning. Your calibration evolves with you. This isn't a problem. It's healthy. Calibration should evolve. --- ## Signs of Good Calibration - You're rarely surprised by the output (in bad ways) - You know what to ask for, and you get it roughly on the first try - Adjustments are small and precise - You can articulate what worked and what didn't - Your cycles feel like progress, not random wandering --- ## Signs of Poor Calibration - Every output surprises you, usually negatively - You ask for X and consistently get Y - Adjustments are big and clumsy - You can't articulate what's wrong, just that something is - Cycles feel like noise, not progress --- ## Calibration Protects the Truth As models become more capable, weak output gets easier to detect — but plausible output gets harder. A capable model can produce a convincing interpretation of almost anything. Inside a strong mycelium, it can connect a new fragment across your entire project in seconds. This is mycelial uptake (Lesson 4), and it is enormously powerful. It can also be wrong. The model's ability to make a connection _meaningful_ does not make the connection _true_. Calibration is what lets you feel the difference between: - recognition and seduction - a living connection and a merely clever one - specificity grounded in context and specificity invented to impress you - an artifact that carries your fingerprint and one that imitates its surface This is where the work remains required. Your judgment is not a quality-control step performed after creation. It is an active ingredient in creation. Each correction teaches the shared network what matters and keeps plausible invention from hardening into truth. > **The model can multiply meaning. Calibration decides which meaning is allowed to live.** --- ## Developing Calibration Faster You can't skip calibration. It requires cycles. But you can sharpen it faster: **Pay attention.** Most people rush through cycles without noticing what's happening. Slow down. **Articulate your reactions.** Don't just think "that's good" or "that's bad." Write it down. Explain why. This forces clarity. **Repeat deliberately.** Work in the same domain for a while. Get calibrated there before moving to a new one. **Keep notes.** Write down what you're learning: "This model is good at X. It struggles with Y. When I ask for Z, I get better results if I do A first." Calibration is more flexible than following instructions. Instructions break when conditions change. Calibration adjusts. --- ## The Blade Think of calibration like a blade. You don't buy it sharp. You forge it. Each cycle refines the edge. After enough cycles, the blade is sharp. But it still needs maintenance. Stop practicing and it dulls. Move to a new domain and you're resharpening for different material. The blade is yours. Nobody can give it to you. You forge it through the work. --- ## Exercise 1: Calibration Audit **Time: 30 minutes** Pick five recent cycles where you got good results and five where you got poor results. Compare them: - How was your input different? - What was your state? How present were you? - How focused was your focal distance? - Was the fingerprint — the mark of presence from Lesson 2 — visible in the output? Write down the patterns you notice. --- ## Exercise 2: Deliberate Practice **Time: 1–2 weeks** Pick one domain you want to build calibration in — writing technical docs, generating code in a specific language, brainstorming, whatever. Over two weeks, run 20 cycles in that domain. After each cycle, note what you tried, what happened, and what you'd do differently next time. After 20 cycles, reflect: What do you know now that you didn't at cycle 1? How has your sense of what works evolved? --- Calibration develops through cycles. But what happens to the patterns you discover? Without something to lock them in, they're fragile. That's the ossification layer.