A withhold-only ranked fair-play rule that protects the ladder without ever calling a person a cheater, backed by submission-time telemetry instead of self-reported clocks.
How the 90-node campaign was pinned to existing problems, test-count coverage, topic balance, and a stable curriculum contract.
How assessment scoring moved from a raw score toward public competencies for problem solving, planning, AI direction, verification, and workflow autonomy.
The shipped gates that decide whether a completed run can become a self ghost, join the human ghost pool, or enter the solution bank.
How a cost-ordered model fallback chain keeps candidate assistance, hints, analysis, and review copilots fast, cheap, and available when a single provider has a bad day.
How UI assessment review moved from code-only inspection toward DOM snapshots, viewport coverage, console issues, and visible artifacts.
A small analytics methodology for counting real product activity across practice, puzzle rush, solved problems, assessments, and classroom sessions.
How the special-judge sweep separated genuinely many-valid-answer tasks from the exact-match tail, then shipped input-aware checkers where the contract was clear.
A storage-first cost model for replayable coding matches, written before keystroke playback became part of the product surface.
How an expensive lobby fetch turned two players queuing into thousands of reads, and how moving selection back to the server fixed the cost shape.