Dowsing Chain — router

The pendulum doesn't guess. It swings toward what is actually there.

Dowsing looks like intuition from the outside. Inside, it is a disciplined sweep: every candidate task and every detected agent must be backed by observed evidence. You may not invent work. A dowsed task with no evidence behind it is a hallucination wearing a pendulum.

Selection table — read exactly one

Your situation	Aspect file
"Find something worth doing" / vague goal to scope / between tasks with standing self-assign authority / triaging a codebase	task-dowsing.md
Lost / ancient / unmarked tree / Pascal / assembly / COBOL / ROM / floppy image	excavate.md
Building the enlistable-agent roster / "what agents are on this machine?" / pre-Steal-Chain scan / an unknown CLI needs profiling	system-dowsing.md
A source contradicts another (docs vs code, comment vs behavior, agent claim vs observation, memory vs terminal) and you must rule	lie-detection.md

Multiple situations at once → run them as separate selections in sequence (excavate often triggers lie detection when a comment disagrees with a binary).

Chain-wide laws (apply in every aspect)

1. Evidence per swing — every finding cites a quote, a file:line, or a command output captured this session. "Seems like" findings are discarded before they reach any report.
2. Read-only chain — dowsing observes; it never fixes, installs, or signs in. Anything worth changing becomes a task candidate for the client (or the waterfall), never an in-place edit "while scanning".
3. One artifact per session — a dowsing run ends in exactly one output (candidate table / roster / ruling / SURVEY.md), recorded in the ledger. An unrecorded sweep is lifespan spent on nothing (Vow of Worthy Spend).
4. Skipped probes are declared — inapplicable or unrunnable probes are listed as skipped in the artifact, so absence of findings is never mistaken for absence of problems.
5. No home language — references/language-agnostic.md. The files pick the language.