nanodevice_flakedetect_review

nanodevice_flakedetect_review — Visual Validation Protocol

Pure agent workflow using existing skills and MCP tools. No custom scripts needed.

When to Use

After the commit step has inserted polygons into KLayout. Review validates the actual committed result, not just overlay images.

Prerequisites

• KLayout running with committed polygons (from the commit step)
• Background microscope image loaded (from the commit step)
• Combine output available: <out>/combine/overlay_raw.png, overlay_lut.png, combine_report.json
• Alignment report available: <out>/align/alignment_report.json

Review Protocol

Step 1: Screenshot the committed layout

Use MCP screenshot to capture the current KLayout viewport with polygons on the background image:

mcp__klayoutclaw__screenshot

Read the resulting PNG to see all layers overlaid on the microscope photo.

Step 2: Inspect individual layers

Use the display skill to isolate each material layer for focused inspection:

# Show only top_hBN (layer 10/0) against the background
python skills/display/scripts/show_only.py 10/0

# Screenshot, then inspect

# Show only graphene (layer 11/0)
python skills/display/scripts/show_only.py 11/0

# Show only bottom_hBN (layer 12/0)
python skills/display/scripts/show_only.py 12/0

# Show only graphite (layer 13/0)
python skills/display/scripts/show_only.py 13/0

# Restore all layers
python skills/display/scripts/toggle_layer.py 10/0 11/0 12/0 13/0 --on

Take a screenshot after each show_only to compare each polygon against the flake features.

Step 3: Compare with combine overlays

Read the overlay images from the combine step for cross-reference:

• <out>/combine/overlay_raw.png — contours on desaturated raw image
• <out>/combine/overlay_lut.png — contours on LUT image (if available)
• <out>/combine/mask_composite.png — color-coded mask overlay

These show the contours before coordinate transform. The committed KLayout polygons should match.

Step 4: Structured visual assessment

Answer these specific questions about the KLayout screenshots:

Q1 — Top hBN boundary: "Do the top_hBN polygons (layer 10/0) follow the flake boundary visible in the background image?" Rate: tight fit / acceptable / poor fit

Q2 — Graphene containment: "Is the graphene polygon (layer 11/0) fully inside the top_hBN polygon (layer 10/0)? Any graphene outside top_hBN indicates a problem." Rate: fully contained / mostly contained / significant leakage

Q3 — Graphite alignment: "Does the graphite polygon (layer 13/0) align with the dark strip visible in the image?" Rate: aligned / offset / missing

Q4 — Bottom hBN coverage: "Does the bottom_hBN polygon (layer 12/0) cover the correct underlying hBN region?" Rate: good / partial / wrong region

Q5 — Overall: "Rate this result: excellent / good / needs work."

Step 5: Check quantitative metrics

Read the reports for numeric cross-checks:

From <out>/align/alignment_report.json:

• Forward Chamfer < 3 um → good, > 5 um → investigate
• IoU > 0.7 → good, < 0.5 → poor
• Outside fraction < 0.1 → good, > 0.2 → poor

From <out>/combine/combine_report.json:

• ECC correlation > 0.9 → raw/LUT registration is reliable
• Transform summary confirms correct warp applied per material

Step 6: Decision

Visual Assessment	Metrics	Decision
excellent/good	metrics OK	PASS — detection complete
good	metrics borderline	PASS with note
needs work (alignment)	high chamfer / low IoU	FAIL → re-run align with different rotation or parameters, then redo combine → commit → review
needs work (detection)	—	FAIL → re-run specific detect script with adjusted parameters, then redo combine → commit → review
needs work (commit)	—	FAIL → re-run commit (coordinate transform or layer error)

Step 7: Document

Record the decision and reasoning. If FAIL, specify which step to retry and what to change.

Vision Feedback Principles

1. Structured prompts, not open-ended. Specific questions get reliable answers.
2. Show raw image as context. The background image provides ground truth.
3. Budget calls. Max ~10 vision calls per review cycle.
4. Vision for coarse judgment, metrics for precision. Vision distinguishes "completely wrong" from "roughly right." Use numeric metrics for fine distinctions.
5. Source images as ground truth. Compare against the original microscope photos when debugging.

Retry Guidelines

• Alignment retry: Try a different rotation from the sweep candidates, or widen the scale range.
• Detection retry: Adjust thresholds or use --cluster-id overrides. Check source image quality.
• Max retries: 2 per stage. If still failing after 2 retries, flag for manual intervention.