CounterCall

CounterCall trains football referees to identify the one observable fact that changes a decision—then prove they can transfer the principle to a minimally changed incident.

Comment

Inspiration

Referee education often measures whether someone remembers a law or reaches the final decision. In a real match, the difficult skill is identifying the observable fact that makes the law apply.

I built CounterCall independently in Moldova to make that reasoning visible and testable. It is not a federation product or endorsement.

What it does

CounterCall is a counterfactual reasoning coach for football referees.

A learner first commits to a decision and names the decisive observable cue. Only then does CounterCall reveal a near-identical incident in which exactly one decision-relevant fact has changed.

The learner must transfer the principle, not simply remember an answer.

  • Coaching is withheld until the learner commits.
  • Each pair changes one declared decision fact.
  • Mastery requires the correct call and the decisive cue on both incidents.
  • The experience separates Law, Fact, and Judgement.
  • Correct guessing without reasoning is not counted as mastery.

How I built it

The product is a responsive browser application with a Node.js server and a Cloudflare Worker-compatible deployment boundary.

The deterministic scenario engine owns correctness and mastery. GPT-5.6 Sol is used through the OpenAI Responses API with strict Structured Outputs to provide concise, personalized coaching after the learner's decision is graded.

The model cannot change the grade. If the API key is unavailable, a request fails, or the live-call limit is reached, the same lesson continues with an explicitly labelled deterministic fallback.

I used Codex throughout product shaping, implementation, API integration, testing, adversarial review, and UX validation.

Challenge overcome

The central challenge was using a generative model without making assessment unreliable.

CounterCall solves this by separating roles: deterministic logic evaluates the answer, while GPT-5.6 explains the misconception and proposes the next practice step. A model failure can reduce personalization, but it cannot silently change the result.

Impact

The one-fact counterfactual pattern can extend from football handball offences to challenges, offside involvement, advantage, and disciplinary thresholds. The same learning design can also apply beyond football wherever a person must transfer a principle across almost identical cases.

Built With

Share this project:

Updates