How the numbers are made.

The reasoning score (90–150)

The official daily run is 12 questions. Your score is computed as: 90 base points, plus your accuracy ratio (correct ÷ total) times 52, plus a speed bonus of up to 8 points. The speed bonus scales linearly against a 45-second-per-question target — finish the full run instantly for +8, use the full budget for +0.

That puts a perfect fast run at 150 and an all-miss run at 90. Single-puzzle modes (Daily Sprint) use fixed anchors instead: 138 for a solve, 104 for a miss. For official attempts the server recomputes the canonical score, rank, percentile, and speed bonus from the submitted tally — the number on the board is server-derived, not whatever the client claims.

What "modeled" means

Your percentile and #N rank come from a fixed reference-population model, not a live census of players. The percentile ladder is a hardcoded mapping from accuracy bands to percentiles, and the rank is your position in a fixed reference population of 1,000,000: rank = 1,000,000 × (100 − percentile) ÷ 100.

Concretely, the hardcoded ladder is: 100% accuracy → 99.9th percentile, ≥92% → 99.2, ≥83% → 97, ≥67% → 90, ≥50% → 74, ≥34% → 58, below that → 37. It is a deliberately conservative editorial model, and it will be recalibrated as real attempt data accumulates.

The daily leaderboards are different: those are real, live rankings of actual submitted attempts, deduped by player. Only the #N-of-1,000,000 rank and the top-X% percentile are model outputs — which is why they always carry the "modeled" label.

How AI baselines are determined

Lines like "Frontier models usually solve this" come from a static per-puzzle flag stored with the puzzle itself. The flags are curated from model evaluation runs and updated periodically with content updates — no live model is called during your run, and the flag does not adapt to model releases in real time.

The "AI misses" count on your result simply counts your correct answers on puzzles flagged as ones current model baselines often miss.

Protections and known limits

What the server enforces today: one official attempt per day per player key; the canonical score, rank, percentile, and speed bonus recomputed on the server for every official attempt (the numbers the client claims are discarded); and consistency checks on the submitted tally — correct answers may not exceed the run length, the AI-miss count may not exceed correct answers, and the run length must be one the official daily ramp actually produces, so a fabricated run length is rejected rather than scored. Every official matrix item also ships a human-readable proof, a formal rule, and a checksum tied to the configured answer.

What remains client-trust today, disclosed honestly: the server now independently re-verifies your answer count from your submitted choices against its own answer key, but this re-verification is being validated in shadow first — the count is checked and logged, while your score is still derived from the submitted correct/total tally (server-side enforcement is rolling out after that validation). The answer key also still ships in the app bundle, a known residual we are closing with a server-held question bank, so until enforcement is on a modified client could still report a correct/total it did not truly earn. Guest identity is also device-based, so an unconnected player identity could be spoofed. A server-held question bank plus stronger identity binding are the planned next steps; connecting an email account already binds your results to a verified account.