{
  "replay_version": "1.0",
  "claim": "The same inputs produce the same verdict, and you can establish that yourself without an account and without seeing any of our logic.",
  "the_logic_is_not_published": "No route here returns the scoring source, the weights, the thresholds, the signal names or any intermediate value. The only thing published is a SHA-256 of the deployed source, which is one-way.",
  "how_to_test_us": [
    "POST /x/replay/challenge with any inputs object you like.",
    "Keep the input_hash it returns.",
    "Send the identical inputs again tomorrow, next month, next year, from anywhere.",
    "GET /x/replay/history?input_hash=... to see every run, each sealed separately.",
    "If the verdict ever moves under an unchanged code fingerprint, the engine is not deterministic and you have proof of it that we cannot withdraw."
  ],
  "why_black_box_is_stronger": "A published listing only shows what the code says. Repeated challenge shows what production actually does, over time, on inputs we did not choose.",
  "what_breaks_determinism": [
    "a wall-clock read inside the scoring path",
    "iteration over an unordered structure",
    "an unseeded random call",
    "any model call in the decision path - which is why most platforms cannot do this"
  ],
  "what_this_does_not_prove": "That a decision was correct, or that the inputs were honestly captured. Only that the same inputs still yield the same output under known code. Determinism is not fairness.",
  "rate_limits": "Repeat submissions of inputs already seen are never limited - retesting is the point. Novel inputs are limited, because bulk novel inputs are how a decision boundary gets mapped rather than how a claim gets tested."
}