crucible-investigation-methodology

# Investigation Methodology — 6 execution patterns > **Self-Evolving Skill**: If any pattern here fails in practice (wrong results, wasted compute), update the section AND append to `references/evolution-log.md`. Don't defer. These 6 patterns executed in service of `a-research-foundations`. They are the "how" to the foundations' "why". Apply in roughly this order for a new hypothesis. --- ## 1. LLM-native data representation — quintile tokens Before asking an agent to "look at" numerical market data, encode as per-bar token sequences using **rolling quintile ranks** within a causal window. **Canonical schema**: ``` idx dir body_q range_q dur_q uwick_q lwick_q loc sess fwd+H... ``` Each quintile is `1..5` in a causal 200-bar rolling window (see Skill A §1). Agents can spot motifs like `+1:5:1|+1:5:1|+1:5:1` (three consecutive fast big-up bars) that are invisible in float-space. Context-budget rule: 60 KB tokenized stats-table fits in agent context; 67 MB raw bars don't. Full reference: `findings/methodology/01-llm-native-data-representation.md`. --- ## 2. Serial adversarial gates (A/B/C/D/E protocol) Before trusting any in-sample positive, survive 4-5 independent gates in series. | Gate | Question | Catches | | ---------------------------------------- | -------------------------------------------- | ---------------------------------------...