AI Skill Report Card
Designing Waste Sorting Choice Architecture
Quick Start14 / 15
Given a waste station contamination problem, run the loop:
- Audit baseline for 14 days → compute $C_B = \frac{\text{Non-recyclable mass in recycling bin}}{\text{Total recycling bin mass}} \times 100$
- Diagnose friction via neutral observation → classify into bias registry (Automatic Pilot, Status Quo Inertia, Ambiguity Aversion)
- Engineer interventions: co-locate bins (≤0.5m), restrict apertures (120mm circular for cans, 400×35mm slot for paper, foot-pedal for landfill), apply visual anchoring
- Test via 30-day RCT: 5 control sites vs. 5 treatment sites
- Validate: accept if Treatment shows ≥40% relative contamination drop AND absolute $C_B$ < 15%
- Monitor: re-audit every 180 days (3-day sample); reset to Phase 1 if $C_B$ climbs above 15%
Recommendation▾
Add a third example showing a failure case where the RCT doesn't meet success criteria, to illustrate what happens next concretely
Workflow14 / 15
Progress:
- Phase 1a: Run 14-day unobtrusive baseline audit, calculate $C_B$
- Phase 1b: Conduct cognitive friction diagnostics, map observations to bias registry
- Phase 2a: Apply convenience geometry specs (co-location radius, friction distancing, visual centerline)
- Phase 2b: Apply aperture restriction + visual anchoring specs
- Phase 3a: Set up RCT (5 control sites, 5 treatment sites, 30 days)
- Phase 3b: Test against success criteria (≥40% relative drop, <15% absolute $C_B$)
- Phase 4: Schedule 180-day automated audit cycle; define loop-reset trigger
Step details:
- Baseline audit (3.1): Exactly 14 consecutive calendar days. Capture daily $W_L$ (landfill mass) and $W_R$ (recycling mass) before processing. Never shorten this window — weekly variance requires full-cycle coverage.
- Friction diagnostics (3.2): Observe from a neutral, non-interfering distance. Tag each anomaly to exactly one of the three registered biases — don't invent new categories without updating the registry.
- Intervention engineering (4.1–4.2): Treat the spatial/geometric specs as hard constraints, not suggestions. If a site's layout can't meet a spec (e.g., space for 3.0m friction distancing), flag it for custom design review rather than silently deviating.
- RCT (5.1): Control and treatment groups must be mirrored in size (5 and 5) and comparable in traffic/context — mismatched site profiles invalidate the comparison.
- Validation (5.2): Both conditions are required (≥40% relative drop AND <15% absolute). Passing only one does not qualify the intervention for standardization.
- Loop reset (6): This is automatic, not discretionary — if the 180-day check shows $C_B$ > 15%, immediately route the site back to Phase 1 rather than patching the existing intervention.
Recommendation▾
Include a bias registry definition table (Automatic Pilot, Status Quo Inertia, Ambiguity Aversion) with brief descriptions since they're referenced but never defined
Examples15 / 20
Example 1: Input: A corporate office kitchen has a single large bin; staff report "almost everything goes to landfill." Output: Run 14-day baseline → likely diagnosis: Status Quo Inertia (single convenient bin) + Automatic Pilot Mode (open-top, no friction). Intervention: install docked recycling+landfill pair within 0.5m, add foot-pedal lid to landfill only, 120mm circular aperture on cans bin. RCT across 5 kitchens vs. 5 unmodified kitchens for 30 days. Report relative contamination drop and absolute $C_B$.
Example 2: Input: A municipal park station passed its RCT 8 months ago but recent complaints suggest increased trash in recycling. Output: Trigger 180-day audit cycle check (3-day sample). If $C_B$ > 15%, flag site, reset to Phase 1 — do not re-engineer apertures directly. New baseline audit will reveal the new friction mutation (e.g., users learned to pry open the pedal lid manually, or signage faded).
Recommendation▾
Provide a template for the RCT report or audit output format to make the deliverable more concrete and actionable
Best Practices
- Always let induction (observation) precede deduction (design) — never skip straight to intervention based on assumed biases.
- Treat all numeric thresholds (14 days, 0.5m, 120mm, 40%, 15%, 180 days) as fixed protocol constants, not tunable parameters, unless explicitly revising the SOP.
- Document which specific bias each intervention element targets — this keeps the design traceable and auditable.
- Keep control and treatment site selections truly randomized and context-matched to preserve RCT validity.
Common Pitfalls
- Don't shorten the 14-day baseline or 30-day RCT windows to save time — short windows miss weekly/behavioral variance and invalidate results.
- Don't rely on text labels as a primary intervention; the protocol treats them as ineffective against rapid-transit, low-attention disposal behavior.
- Don't declare success on relative improvement alone — a 40% drop from a very high baseline can still leave absolute contamination above the 15% ceiling.
- Don't skip the 180-day re-audit cycle; behavioral adaptation to physical friction is expected and must be re-diagnosed, not assumed static.
- Don't patch a failed site in place — a contamination spike must route back through full Phase 1 induction, not a quick Phase 2 fix.