Sketch, Spec, Swap

The constraints a local AI imposes on student work (standard label placement, architectural lettering, structured Markdown spec, four named sections) aren't workarounds for a primitive model - those constraints are the curriculum. The discipline of producing work that an AI can read IS the discipline of producing work that any future reader (a peer, a teammate, an interviewer, future-you) can read.

This recasts the whole pipeline: the QR-corner, the block-print, the spec template, the rubric vocabulary of gap-types - all the precision rituals - are not concessions to the machine. They're the same rituals an architect, a draftsman, a Product Requirements author, a technical writer has always observed. Machine-readable IS human-precise.

Apply this when a student pushes back on the precision rituals ("why do I have to use block print?"). Answer: because that's what makes you a writer whose work travels. The AI is a stand-in for every future reader.

The Materials Checklist on the Lab Scene plan is the shipped, sub-runnable, day-of version. This is the messy list where new items get tried out before they graduate to the plan (or get cut).

• Lab journal. Bound, every student has one. Goes home and comes back. The take-home spec lives in here; rough notes from the scratchpad can be taped/clipped in. Question: composition book vs. spiral vs. graph-paper-bound. Lean composition.
• Ruler. Probably 6″ clear plastic ruler. Helps the sketch and reconstruction steps look like actual technical drawings rather than freehand boxes. One per student, kept in the journal.
• Architectural lettering guide. Possibly. Inexpensive plastic stencil drafters and architects used to produce uniform, machine-readable block print before CAD - the style is architectural lettering (single-stroke, all-caps, sans-serif); the tool is a lettering guide or lettering template (mid-century brand: Leroy, Wrico, Rapidesign). Reinforces the block-print convention the lesson already establishes and makes the take-home spec OCR-friendly without expecting students to have professional handwriting. Open question: one shared per group, or one per student?
  • Printable lettering practice sheets. Free + scalable companion to the physical guide: PDF printouts of the architectural-lettering alphabet (and the digits 0-9) in light-gray ghosted letters that students trace over on the first pass, then reproduce freehand below. One sheet per student, refillable from the supply shelf. Doubles as a CFU artifact - the practice sheet itself is something a vision-LLM can grade for letterform consistency. Could also stand in for the physical guide entirely if the per-student stencil cost runs too high.
• Graph paper. Already on the checklist (~5mm grid). Need to confirm a sourcing channel - bulk Amazon vs. school print shop.
• Pencils. Already on the checklist. Mechanical vs. wood? Mechanical reduces sharpening time; wood is cheaper to bulk-buy. Probably wood with a class-shared sharpener.
• Sticker sheet of QR-coded ID labels. One Avery 5160 sheet per student, issued at start of semester. See the Submission Pipeline section below for the architecture.

When an item earns its place, copy a short telegraphic line into the Materials Checklist section of the Lab Scene plan and delete it from this list. The drafts themselves stay here so the reasoning is recoverable.

Note: draft lists like this one - from every Lab Scene's Notes page - can be collected and processed into a single Course Materials List that covers the whole semester. This page contributes its students, the next triplet's contributes its own, and the union is what gets ordered.

Per-classroom items, ordered once and reused across every triplet. Same draft-list discipline as the student kit; same downstream destination.

• Document scanner. Owned; campus copier as backup. 30 ppm duplex eats a 100-student class in ~7 minutes.
• Drop boxes. Two per Lab Scene today (Sketch + Reconstruction). Plain stackable bins, labeled. Need a stable home in the room and an end-of-day collection ritual.
• Pencil sharpener. Wall-mounted or desk-clamp; placed away from the supplies queue so it doesn’t bottleneck.
• Backup supply shelf. Extra graph paper, extra rulers, extra lettering guides, extra pencils, extra Avery sheets for replacement label strips. Self-serve so students don’t need to interrupt to ask.
• Display for the Scene Schedule. Project the 6-beat timetable, or a permanent poster? Probably project per-scene so it can be swapped easily, but a generic Act 1 schedule poster could anchor the wall.

Four artifacts per student, four times a week for some sections. Volume is the constraint, not depth.

• Type the Four-Section Spec + the closing reflection. Submit as Markdown on the site. LLM rubric pass against rows (a), (b), and (d).
• Photograph the sketch + peer reconstruction. Vision-LLM first pass extracts labels and counts sections; teacher spot-checks.
• Lean on the rubric's design: row (b) judges spec precision against the peer's actual reconstruction. The Beat 5 swap does most of the assessment work for the teacher.
• Pre-build a feedback bank tied to the four named gap-types (ambiguity, omission, implicit assumption, observational error). LLM tags each artifact; teacher approves the tags. No per-student feedback-paragraph authoring.

The Beat 3 sketch and the Beat 5 reconstruction are paper-only by design. Submission flow:

1. Each student gets one Avery 5160 sheet of pre-printed labels at the start of the semester (30 labels = ~15 scenes worth of sketch + reconstruction labels per student).
2. Each label carries both a human-readable ID (e.g. ID #047) and a small QR code encoding student_id:section:semester. QR is what the pipeline actually reads; the number is so the student can confirm at a glance.
3. Standard placement: top-right corner, always. Taught once at onboarding; reinforced by the lesson's "convention is the contract" frame. Lets the pipeline crop a known region instead of label-hunting.
4. Two drop boxes at the front of the room - one labeled Sketch (Beat 3), one labeled Reconstruction (Beat 5). Box implicitly identifies the artifact type, so the label doesn't need to.
5. End-of-day scan: campus copier (30 ppm duplex) eats a 100-student class in ~7 minutes.
6. Pipeline: crop top-right -> decode QR -> vision-LLM on the body -> rubric tag -> flag for review. Typed spec + reflection arrive through the website form with student ID already baked in - they don't need this paper flow.

Volume: ~200 paper sheets/week to scan for one section; capture isn't the bottleneck - the borderline-rubric review queue is what to design for. The teacher dashboard should surface the AI's low-confidence calls, not present every artifact for grading.

For "does the Structure list match the drawn elements" the task is UI-element detection + label reading, not free-form description. Block-print + graph-paper conventions in the lesson plan are precisely what makes this feasible.

Next step: spin up llama.cpp + Vulkan path on the MinisForum dev box; pipe a sample sketch + Structure list to Qwen2.5-VL 7B first (fits on the 890M iGPU), compare to LLaVA-1.6 13B once on the 4090 rig.

• Submission UX: native form on muggsofcompsci.net or Google Form bridge for the typed artifacts? D1 storage either way.
• Photo capture: do students upload from their phones (and we accept the off-policy phone use for the upload moment) or do we put one classroom scanner / document camera in the lab?
• What does the teacher review pass actually look like? A dashboard that surfaces the LLM's low-confidence calls and the borderline rubric scores - not a per-student grading queue.
• Feedback return loop: in-class verbal in the next scene, written digital, or both?