ML Proposes, Geometry Disposes

"The model gets a vote on where the eyes are. It gets no vote on what the answer is."

The Line

Loomis draws one bright, non-negotiable line through itself. Above it lives everything learned: the detectors may propose where the landmarks are. Below it lives everything deterministic: geometry alone defines the construction, the pose, and the grade. A neural network may suggest the fit; it may never draw the answer or set the score. State it as law, because the whole engine's trustworthiness is cantilevered off it: a learned component may propose a fit, but it may never define the answer key.

Why the Line Holds Reproducibility

The model is the only non-deterministic thing in the system. By confining it strictly to 'propose the landmarks' and making everything downstream pure geometry, Loomis stays reproducible — the same photo yields the same construction — even though it contains machine learning. The determinism boundary and the propose/decide boundary are the same boundary. Let the model reach past the line and touch the pose or the grade, and reproducibility dies instantly, because now the answer depends on a black box that need not repeat itself.

Errors on the Right Side of the Line

Placing the learned part where it only proposes also puts its mistakes where they're cheap. A bad landmark fit is visible — the reprojection error rises, the overlay looks off — and correctable — a human nudges a point and the deterministic geometry re-runs exactly. Compare that to a mistake baked into a black-box decision: invisible, uncorrectable, unreproducible. The design choice is really about where errors are allowed to live. Loomis keeps them on the proposal side, where you can see them and fix them, and out of the answer, where you can't.

A Pattern for Trustworthy AI

This is far bigger than drawing. It's a general recipe for building systems that are both smart and trustworthy: let the learned component propose, and let a deterministic layer dispose. A generator drafts; a verifier or a set of rules gates. An LLM suggests an action; a deterministic check authorizes it. The neural part earns its keep by handling the fuzzy, high-dimensional proposal that rules can't; the deterministic part supplies the guarantees the model can't. 'ML proposes, geometry disposes' is Loomis's instance of a pattern you can reuse anywhere you need learning and trust in the same system.

Code

The line, made explicit in code·python

# THE LINE. Learned ABOVE it, deterministic BELOW it. Nothing crosses down.

# --- learned: may PROPOSE, may be wrong, is correctable ---
landmarks = detector.detect(photo)                        # a neural net's proposal
landmarks = apply_human_corrections(landmarks, edits)     # a recoverable layer

# --- deterministic: DEFINES the answer -- pure geometry, reproducible ---
pose  = solve_pose(landmarks, CANONICAL_HEAD)   # exact
box   = build_construction(pose)                # exact
grade = grade_against(user_drawing, box)        # exact

# The model never touches pose / box / grade. That is precisely why one photo
# has exactly ONE answer -- and why the whole thing survives containing ML.

Exercise

Take an AI system you use or have built and sort its outputs into two piles: things the model merely PROPOSES (a draft, a ranking, a suggestion) and things it effectively DECIDES (an action taken, a value stored, a user-visible result). For one item in the 'decides' pile, sketch how you'd move it behind a deterministic check or a human confirmation — turning a decision back into a proposal. What guarantee would you gain?

Hint

The trustworthy-AI move is to shrink the 'decides' pile: every decision you can demote to a proposal-plus-deterministic-gate is a place you regain reproducibility, auditability, or safety. Loomis shrank its 'decides' pile to nothing — the model only ever proposes — which is exactly why its output can be an answer key at all.