| name | optics-correct |
| description | Run the correction-as-FEEDBACK loop on an optical bench — surface the proposed fixes, then judge each against user intent (refuse / partial / accept) rather than blindly auto-fixing. Use when asked to "fix the bench", "what's wrong and correct it", "clean up the alignment/energy", "review for problems and fix", or "/optics-correct". Never silently "improve" a bench the user built on purpose.
|
optics-correct: surface, judge, then apply
Corrections are advisory feedback, not commands. A "fault" is often the experiment: a crossed
analyzer IS an extinction measurement, a retro-reflection IS a cat's-eye / Michelson end mirror, an
underfilled figure IS a sub-aperture probe. Blind auto-fix breaks intent. Mirror the physics-honesty
gate: surface the issue, weigh intent, decide.
Workflow
- Get the proposals —
propose_corrections(). Each carries: issue, element, detail,
severity, suggested_fix, the tool that would apply it, maybe_intentional_if, and
fault_confidence (0..1: genuine fault vs design choice).
- Judge each one — for every proposal, ask did the user ask for this on purpose?
- Read
maybe_intentional_if. If it holds given the user's stated goal → REFUSE (report it,
change nothing).
- High
fault_confidence (e.g. energy_violation ~0.9, a config bug) → lean ACCEPT.
- Low
fault_confidence (e.g. crossed_polarizer ~0.3, usually intentional) → lean refuse, or
confirm with the user before touching it.
- Mixed → PARTIAL: apply the safe part, flag the rest.
- Apply the accepted ones — use each proposal's
tool (align_element / set_param / place_relative
/ design_4f …). Apply one, re-trace, re-check — don't batch blindly.
- Re-verify —
propose_corrections() / diagnose() again; confirm the accepted issues cleared and
nothing new broke. inspect_beam / inspect_element to confirm the physics is now right.
- Report the judgment — list what you ACCEPTED (and the measured before/after), what you REFUSED
(and why it's intentional), and what you left PARTIAL. The reasoning is the deliverable, not just the
diff.
Disciplines
- Never auto-apply a low-confidence proposal without confirming intent.
- Each accepted fix that moves an element makes the trace non-byte-identical — say so.
- This is the same honesty you'd want for physics: state assumptions, show the numbers, don't pretend a
refused issue doesn't exist — surface it so the user decides.
