Safety decision

ATC-20 Placard Mapping

The Applied Technology Council's post-disaster building safety placard system, encoded into the classifier's output.

GREEN
ATC-20
No damageSafe to enter

Building inspected. No restriction on use or occupancy.

Engineer action
Routine inspection only
≥ 0.55, class = no-damage
YELLOW
ATC-20
Minor damageRestricted use

Entry, occupancy, and lawful use restricted as posted. Repairs required.

Engineer action
Limited entry permitted
≥ 0.55, class = minor-damage · OR any prediction with confidence < 0.55
ORANGE
ATC-20
Major damageEngineer inspection

Significant structural damage. Requires engineer inspection before any re-occupancy.

Engineer action
No re-entry until engineer signs off
≥ 0.55, class = major-damage
RED
ATC-20
DestroyedUnsafe — DO NOT enter

Extreme hazard. Do not enter under any circumstances.

Engineer action
Posted unsafe
≥ 0.55, class = destroyed

Decision logic

The classifier predicts one of four damage classes. to_placard() maps that label to a placard — but only if confidence ≥ 0.55. Otherwise the system conservatively escalates to YELLOW for human review.

if confidence < 0.55:
    → YELLOW   (engineer review)

elif top_class == "no-damage":     → GREEN
elif top_class == "minor-damage":  → YELLOW
elif top_class == "major-damage":  → ORANGE
else:                              → RED   (destroyed)
Why confidence gating?
ATC-20 is a safety document, not a research metric. A wrong RED costs occupant trust; a wrong GREEN can cost lives. When the model is unsure, the placard always defaults to YELLOW so a human engineer makes the call.