diff --git a/STYLE.md b/STYLE.md
index 9c59746ff..ee4c1a683 100644
--- a/STYLE.md
+++ b/STYLE.md
@@ -10,25 +10,94 @@ TLDR:
 
 We want to decouple metrics and analysis from the core codebase as much as possible. Metric- and figure-generation code should be encapsulated in `spd/metrics.py` and `spd/figures.py`.
 
+### Pure Functions & I/O
+- Keep I/O at the edges of your code
+- Make as many functions as possible pure (no side effects, deterministic output from inputs)
+- Push file operations, logging, and network calls to the highest reasonable level in the call stack
+
+### Data Structures
+- **Don't use bare dictionaries for heterogeneous structures**
+  - **Good**: `{user_id: user_data}` - homogeneous values indexed by keys
+  - **Bad**: `{"tokens": [...], "loss": 0.5, "metadata": {...}}` - use a dataclass or typed dict instead
+- This makes code more maintainable and catches errors at type-check time rather than runtime
+
 ### Fail Fast (Negative Space Programming)
 Code should fail immediately when assumptions are violated, preventing bugs from propagating.
 
-If there's an assumption you're making while writing code, assert it.
-- If you were right, then it won't matter
-- If you were wrong, then the code **should** fail.
+**Assert liberally:**
+- If you have an invariant in your head, assert it
+- If you're afraid to assert, your program might already be broken
+- Never soft-fail or recover silently from unexpected states
+
+**Don't handle errors gracefully without good reason:**
+- If your program isn't working correctly, it shouldn't be running—you should be fixing it
+- Exceptions should be exceptional; let the program crash and surface the real issue
 
 ## Type Annotations
 - Use jaxtyping for tensor shapes (though for now we don't do runtime checking)
-- Always use the PEP 604 typing format of `|` for unions and `type | None` over `Optional`.
+- Always use the PEP 604 typing format of `|` for unions and `type | None` over `Optional`
 - Use `dict`, `list` and `tuple` not `Dict`, `List` and `Tuple`
-- Don't add type annotations when they're redundant. (i.e. `my_thing: Thing = Thing()` or `name: str = "John Doe"`)
+- Don't add type annotations when they're redundant (i.e. `my_thing: Thing = Thing()` or `name: str = "John Doe"`)
 - Prefer `match` over `if/elif/else` chains when dispatching on conditions - more declarative and makes cases explicit
 
+### Encode Invariants in Types
+Write your invariants into types as much as possible to offload validation to the type checker:
+
+- **Joint dependencies**: If you need both `a` and `b` together, or neither:
+  ```python
+  # Bad: independently optional
+  a: str | None
+  b: int | None
+
+  # Good: coupled in the type system
+  config: tuple[str, int] | None
+  ```
+
+- **Make invalid states unrepresentable**: If tracking dependent states requires lots of assertions, encode the dependency in types instead
+
+### Distinguish None from Empty Collections
+Differentiate "no data" from "empty data" when semantically meaningful:
+
+```python
+# Bad: ambiguous
+def process(items: list[str] | None = None): ...
+
+# Good: explicit semantics
+def process(items: list[str] | None):
+    # None = data not loaded yet
+    # [] = loaded but empty
+    ...
+```
+
 ## Tensor Operations
-- Try to use einops by default for clarity.
+- Try to use einops by default for clarity
 - Assert shapes liberally
 - Document complex tensor manipulations
 
+## Function Design
+
+### Default Arguments
+Default arguments are a code smell more often than they're useful:
+
+- You should have a **very good reason** for having a default value for an argument
+- Especially problematic when a caller also defaults to the same thing—this creates hidden coupling
+- Keep defaults high in the call stack rather than deep in implementation details
+- Often it's better to be explicit at the call site than to hide behavior in defaults
+
+**Ask yourself**: "Does this default encode domain logic that should be visible to the caller?"
+
+## Code Hygiene
+
+### Delete Unused Code
+- If an argument is always the same value, strongly consider removing it as a parameter
+- Delete dead code immediately—don't comment it out
+- If code isn't being used, the repository history will preserve it
+
+### Remove Unnecessary Logs
+- Logs should provide actionable information
+- Remove debug logs that don't help users understand what's happening
+- Keep logs at module boundaries and for genuine diagnostics
+
 ## Comments
 
 Your first instinct should be: "If I couldn't write any comments, how would I write this code?"