Frozen Fits, Layered Corrections

"Never overwrite what the machine said. Freeze it, and lay your correction on top where you can always peel it back off."

Freeze What the Machine Said

When a detector proposes a fit, Loomis stores that raw output frozen — an immutable record of exactly what the machine saw. It is never edited in place. Whatever happens later, the original detection remains as a faithful account of the proposal, because the moment you overwrite it you lose the one thing that made it a record: its unchanged-ness.

Corrections Are a Layer, Not an Overwrite

When a human nudges a landmark the detector misplaced, that adjustment does not reach in and change the frozen fit. It is stored as a separate correction layer — recoverable, reversible, and sitting on top. The fit the geometry actually uses is computed from the two: the frozen detection with the corrections applied over it. The original stays intact underneath, so a correction can be undone, re-examined, or thrown away without any risk to the record it modifies.

Provenance Stays Honest

Because the machine's proposal and the human's edits live in separate layers, the engine can always answer three distinct questions: what did the detector propose, what did a person change, and what did the geometry ultimately use? That is provenance, and it is only possible because nothing was destructively merged. Overwrite the detection with the correction and all three questions collapse into one lossy blob that can no longer tell you who was responsible for what. Keeping the layers apart is what keeps the story auditable.

The Same Shape, Everywhere

This is one instance of a pattern that repeats across the whole engine and the whole family: freeze the source of truth, layer edits on top, recompute the view. The grader in Track 6 preserves your drawing attempts append-only, never mutating a past attempt. Collections in Track 7 organize sessions through a meta-overlay while the session files on disk stay untouched. Pippa's own conversation logs are an append-only ground truth with rebuildable derived mirrors. Once you see the shape here — frozen fit, correction layer, computed effective view — you'll recognize it as the family's default answer to 'how do we allow edits without losing the truth?'

Code

Frozen source, correction layer, computed view·python

# The detector's output is FROZEN -- an immutable record of the proposal.
detection = detector.detect(photo)         # stored once, never mutated in place

# Corrections are a SEPARATE layer -- recoverable and reversible.
corrections = load_corrections(session)    # e.g. {landmark_id: new_xy, ...}

# The effective fit is COMPUTED, never a destructive merge:
effective = apply(detection, corrections)  # 'detection' stays intact underneath

# Provenance is always answerable:
#   detection   -> what the MACHINE proposed
#   corrections -> what the HUMAN changed
#   effective   -> what the GEOMETRY actually used
# Overwrite 'detection' in place and all three collapse into one lossy blob.

Exercise

Find a place in your own work where you edit a record in place and lose its previous state — a config you overwrite, a field you mutate, a file you save over. Redesign it as immutable source plus a correction/override layer, with the working value computed from both. List what you gain: undo? audit trail? the ability to say what changed and who changed it?

Hint

The tell is any operation where the old value is simply gone after the edit. Converting it to 'keep the original frozen, store the change as a layer, compute the effective value' costs you a little computation and a little storage, and buys undo, provenance, and reproducibility. That trade is why Loomis — and the whole family — pays it by default.