A digitizer that learns from you

The short version of how the auto-digitizer gets better — and how your calls drive it.

Forge (the auto-digitizer) turns a logo PNG into a stitch file. It's good at the easy 80% and wrong in specific, repeatable ways. Instead of hand-tuning it forever, we built a loop: Forge records the decisions it makes, a human reviews or corrects them, and those corrections become rules that make the next run better. The Judge is where those human calls happen.

Forge makes a decision ─► it records what it decided + why (the features) │ ▼ a human reviews it in the Judge ─► you confirm or correct the call │ ▼ the correction is distilled into a proposed rule change │ ▼ [ gate: a trustworthy check + a human sign-off ] ─► the rule ships to production │ └────────────────────────────────────────────► the next digitization is better

Two rulers (why we measure twice)

Quality has two different questions, so we keep two separate measures — conflating them was the old mistake.

EDS — the offline compass. "Is this change moving us toward the experts?" It compares Forge's output to an expert's file for the same design. Great for choosing what to try next — but it needs an expert file, so it can't judge a brand-new customer logo.

PAS — the production gate. "Is this output shippable, or does it need a fix?" It judges Forge's output against the source artwork only (no expert file needed), and routes it: ship / minor fix / major fix / reject. This is the one that decides real jobs. We're building it now from your dispositions.

How you help — two quick tasks

Decision review

We show one region Forge stitched and its call ("I made this satin, not a filled panel"). You say: right, or it should've been a composite. Your calls sharpen a specific rule.

Open decision review →

Disposition

We show Forge's finished output beside the artwork. You route it: ship / minor / major / reject (+ a defect tag). These are the labels that build PAS, the production gate.

Open disposition →

A few dozen decision reviews and ~150–300 dispositions is enough to turn on the first automated proposals. Every vote is stored and attributed to you.

How a correction becomes a rule

Propose — the system reads your verdicts and drafts a specific rule change (a moved threshold), with your votes as the evidence.
Gate — nothing ships automatically. A proposal must clear a trustworthy quality check (PAS), a hold-out test on a different style, and a human sign-off.
Promote — a person accepts it; the rule lands in the knowledge base, stamped with its evidence, and Forge picks it up on the next run. That stamp is how we track that the loop is actually paying off.

Where we are

The loop is built and live: Forge records its decisions, both review surfaces are open, the proposal engine and the cost model run. The one piece still being assembled is PAS (the production gate) — and it's waiting on your dispositions. So the highest-value thing right now is simply: label a batch.

Full operator detail: PI-EMB-Codex/wiki/50-Automation/operator-runbook.md · the method: two-rulers-eds-and-pas.md · back to Judge