← back to pin list

14c. The check list — what is fully automated, what needs a human

kerfmaster pin list · Preflight — how files come in

intentmixedWhat we want to build. Never claimed the present, so it cannot be stale — and must never be read as a description of the code.

Content last changed 2026-08-21 — computed from the item itself, not typed.

Split deliberately into what a machine can decide alone and what it can only flag. Anything in the second list that gets silently auto-decided is a future scrapped part.

⚑ Build state, 2026-08-03. This list is the DESIGN of the intake gate, not a description of one. It is typed intent and every "BLOCKING" below is a decision about what the gate should do, not a report of what it does. What exists today, with the code behind it: %PDF magic bytes and legacy-PostScript .ai rejection (vec_io.py:278-279), chained ASCII85+Flate decoding, fill-vs-stroke op counts and clip-path counts (vec_io.py:290-293), live-text detection (vec_io.py:284), open contours + duplicates (qc.analyze), sub-kerf specks (vec_pipeline.py:151), island nesting depth (emit._depths), the 80"×160" envelope (vec_pipeline.envelope_check) and the scale gate (ScaleUnknown, item 14b). Not built at all: self-intersection detection, minimum-feature-vs-kerf, encryption check, producing-application check, embedded-raster (Do) flag, flattened-export detection, and magic-byte sniffing for SVG or DXF. Nothing in the list BLOCKS except the scale gate and the envelope; everything else that runs is reported.

A. Fully automated — file integrity

• Magic bytes match the extension. Built for PDF/.ai only: vec_io.read_pdf_paths requires %PDF and raises "not a PDF-backed file (legacy PostScript .ai?)" otherwise (vec_io.py:278-279). <svg and DXF section markers are not sniffed — intent.
• Not encrypted / password-protected.
• Content streams decode, including chained filters — newer Illustrator (29.1) uses ASCII85Decode + FlateDecode, older (25.2) just Flate.
• Contains vector geometry at all; is not a raster wrapped in a vector container.
• Producing application recorded, where the file names it.

B. Fully automated — scale and machine limits

• Scale indicator present (see 14b). BLOCKING.
• Stated dimension vs measured bounding box, cross-checked. BLOCKING on mismatch.
• Fits the 80"×160" bed. BLOCKING.
• Smallest feature vs kerf — any slot or hole narrower than the kerf is uncuttable. Intent, not built: nothing measures minimum feature width and nothing blocks on it.
• Holes below the pierce-standoff limit — reported, not blocking.

C. Fully automated — geometry health

• Open contours and gap sizes, against the auto-close threshold (item 9).
• Duplicate and overlapping segments — one line per cut.
• Branches — built (qc.analyze emits branch and dangling flags). Self-intersection detection does not exist: grep -rn intersect cadmaster/*.py returns nothing, so a self-crossing contour posts silently and the comp side is undefined on it. This line read "QC step 1, done" — the most dangerous claim in the list, and the reason this pass exists.
• Zero-length and sub-kerf degenerate specks (item 12).
• Island nesting depth — what will need bridges (item 2b).
• Live text vs outlined text. Detection is built — vec_io.py:284, live_text = any(re.search(r"\bBT\b", s) for s in streams). Note the operator: it is the BT text BLOCK, not Tj/TJ as this line used to say. It is reported only (build_vec_board.py:339) — a file with live type posts today. BLOCKING is intent.
• Fill ops vs stroke ops counted (f, f*, B vs S, s) — feeds the traced-raster test below.
• Embedded raster images present (an Image XObject drawn with Do) — flag.
• Clip paths (W n) — ignored as artboard clipping, but reported as ignored.
• Flattened-export detection — a DXF with no bulges anywhere and long runs of tiny near-equal straight segments is almost certainly an Illustrator export. Flag and ask for the original (14a).

D. Automated detection, human decision

• Traced raster. Thousands of small filled paths, no strokes, many near-identical shades — the garage-cover sample has 7,812 filled paths. Detectable with high confidence, but the call to reject belongs to a person, because a legitimately intricate design looks similar to a machine.
• Borderline heal chains — where spike-vs-sustained is genuinely ambiguous. Already our rule: flag, never silently guess.
• Near-duplicate contours — mistake or deliberate double line?
• Crop marks, registration marks, stray guides that survived into the geometry.

E. Cannot be automated — intent

• Is this outline the cut line, or a drawn representation of a thick stroke? Only the customer knows.
• Cut vs engrave vs score per contour — until the colour convention exists (item 8), this is a declaration, not a measurement.
• Comp on or off per boundary — already established as a per-job declaration, not inferable from geometry (items 4b, cutter-comp convention).
• Whether forty tiny islands are art or an accident.

Design rule for the whole list: every automated check either passes, blocks with a plain-English question, or produces a note on the report. There is no fourth behaviour, and in particular there is no "quietly fix it".

⚑ Build state update, 2026-08-21 (supersedes the 2026-08-03 block above where they disagree): preflight.py wires the checks into the actual gate with the three 14-outcomes. Newly built: magic-byte sniff for DXF (text + binary sentinel), DWG (AC10xx), PDF/.ai, SVG, PNG/JPEG/GIF/BMP raster, plus extension-vs-content mismatch reporting; empty/undecodable-file blocking; contains-no-geometry blocking; stated-finished-dimension cross-check — BLOCKING on mismatch outside ±2% (assumed tolerance, provenance in the file), with the rotated-90° case caught and asked about rather than blocked; the 80×160 envelope BLOCKING in either orientation; flattened-export detection (14a — no native curves + 200-plus tiny lines → “do you still have the original?”); exactly-doubled-segment scan, O(n) by endpoint+centre hash (found 349 real doubled semicircles in TL-1016-82 that the sampled overlap scan under-reported as 1; the O(n²) partial-overlap scan measured 12 s on the novi lettering and deliberately stays a heal-stage QC job, the report says so); open contours / gaps / dangling / branches classified against the pin-9 auto-close thresholds; sub-kerf specks and under-standoff holes by loop longest-dimension (the report itself states this is a PROXY — true slot-width vs kerf is still not measured). Still not built: self-intersection, encryption check, embedded-raster (Do) flag, producing-application capture, and running .ai/PDF/SVG through this flow — vector art is recognised and BLOCKS with the scale question, routing note included, until the vector front-end is wired in.