sherlocklisted

# sherlock Data before theory; several theories at once; kill by evidence, not by preference; converge only when one theory explains every clue. Keep the trail so a dead branch is never re-entered. ## The loop 1. **Frame the case.** Rewrite the ask as one answerable question with a stopping condition, and write what a *no* looks like. Split what is **known** (observed) from what is **assumed** (told, inferred, remembered); each assumption gets a status and an *if wrong* in the casebook. "Find out how X works" has no stopping condition; "which of caching / precompute / streaming makes X return in 50ms at 10M rows?" does. 2. **Observe before you theorize.** Collect clues first — logs, code, traces, pages, dates — and log each as `E<n>` with its source, its date, a grade, and whether it was **observed** or **inferred** (see the casebook format). Record **absences** as clues (the log line that should be there and isn't; the feature the docs never mention). Record the **odd detail** — the one fact that doesn't fit the obvious story is the lever, not noise. Never edit or "fix" while observing. 3. **Write the theories.** At least three `H<n>`, always including a boring one (config, cache, stale build, different user) and one that questions the premise (the report is wrong, there are two causes, the docs describe a previous architecture). For each, write what it predicts that the others do not, and the cheapest test that would show it false. Keep an explicit `H0: none of the above