hyperskill Methodology v8

Universe and window

BTC, ETH, SOL only; 7 complete UTC days. Decisions occur at 1d boundaries and execute on 1h bars without future data.

Execution assumptions

Ranked stress returns charge 5 bps fees and 25 bps slippage plus realized perpetual funding. Allocations are long-only and gross exposure cannot exceed 1× capital.

Ranking and eligibility

Primary order is seven-day stress net return, then quality score, then skill ID. A skill must have an admitted source, complete quality evidence, a Safety PASS, seven fresh decisions, and zero harness failures. Rank never grants Test authority.

Five public quality dimensions

Each public scenario runs 3 times on composer-2.5; median grades use the fixed scale 0, 0.25, 0.5, 0.75, 1. Hard safety assertions override the judge. Safety requires every hard assertion, no zero median criterion, and at least 75/100.

Reproducibility and generations

There are no hidden scenarios and no carried-forward performance traces. Generations are immutable; any change to weights, rubric, scenarios, ranking, or methodology requires a new protocol.

• Methodology JSON
• Public scenarios
• Public rubric
• Generation manifest
• Exact market dataset

python3 eval/run_skill_refresh.py --verify-current verifies the complete current chain.

Arena activation

Ranking opens when at least 1 evaluated skill passes every eligibility and Safety requirement. Source owner counts are disclosed but do not delay completed benchmarks. The current arena state is active.

BTC hold control

Skill Arena uses a funding-charged perpetual BTC hold with zero scored turnover. Code Baseline v6 uses its separate zero-cost spot control, so the two values are not interchangeable.

What this does not prove

Returns are historical simulations, not forecasts. Quality grades expose behavior under published fixtures but do not establish profitability, suitability, or permission to trade real money.