ValidStitch's validation engine is mostly deterministic rules against measurable thresholds, with a smaller AI-assisted layer on top. Below: the split, and the failure modes the engine doesn't catch yet.
What's deterministic
- Hoop fit (geometric check against the selected hoop's safe-stitch area)
- Stitch density (count per mm² against the fabric threshold)
- Long satin stitches (length measurement against the safe-length threshold)
- File-format integrity (header parse, stitch-record completeness, version match)
- Color sequence (count of color blocks, presence of companion palette file)
- Total stitch count and design extent (recomputed from actual stitch data)
Deterministic checks are predictable. The same file produces the same findings every time. The thresholds are documented per rule and you can adjust them per project.
What's AI-assisted
- Severity ranking when multiple findings interact (e.g., 'high density AND long satin' is worse than either alone)
- Suggested fix selection when more than one deterministic remediation is valid
- Detection of stitch patterns likely to snag based on training over flagged production runs (this one is genuinely model-based)
AI-assisted findings are clearly labeled in the report. They surface confidence (low / medium / high) and never override deterministic errors.
What we don't catch yet
- Thread-color clashes with the garment color (we know the design colors; we don't know the blank)
- Operator-handled issues like hooping technique, machine timing, bobbin tension drift
- Wear patterns specific to a particular garment design (high-stretch areas, seams)
- Customer-intent issues (the design is correct but isn't what the buyer wanted)