crucible-research-foundations

Research Foundations — 6 epistemic disciplines

Self-Evolving Skill: This skill improves through use. If a discipline's guidance fails in practice or a new trap emerges, update the relevant section AND append to references/evolution-log.md. Don't defer.

Read these in order. The first three (causal, labels, nulls) are the hardest prerequisites — violating any of them silently invalidates every downstream result.

---

1. Causal-feature invariant (bars[:i])

Every feature f[i] used at trigger/decision bar i must be computable using only bars[0:i] — never bars[i], never bars[i+1:]. Violation produces look-ahead bias; findings silently become worthless.

Canonical pattern:

for i in range(n):
    lo = max(0, i - window)
    wind = values[lo:i]  # EXCLUSIVE upper bound — no peeking
    f[i] = compute(wind)

Note lo:i (exclusive), not lo:i+1. This discipline "feels off by one" but is correct.

Verification test (add to every new feature function):

def test_causality(fn, n=1000):
    bars = generate_test_bars(n)
    f_orig = fn(bars)
    bars_mod = bars.copy()
    bars_mod[500:] *= 2  # perturb the FUTURE
    f_mod = fn(bars_mod)
    assert np.array_equal(f_orig[:500], f_mod[:500]), "look-ahead detected"

Silent-bug signature: impossibly clean results (tw > 10 bps on FX, win rate > 70%, OOS matches IS perfectly).

Full reference: findings/methodology/10-causal-feature-invariant.md.

---

2. Label-leakage (bar-local scaling kills window leakage)

Forward labels must be scaled to the triggering bar's own range, NEVER to a window-wide scale. Window-relative labels are tautological.

Trap: If you label fwd+H = UP when close[i+H] - close[i] > window.span/20, then when close[i] is near window.min (loc=B), fwd=UP is near-automatic. Agents will report spurious "signals".

Fix: use bar-local triple-barrier labels:

r = high[i] - low[i]   # THIS bar's range, not window's
tp_level = close[i] + tp_mult * r
sl_level = close[i] - sl_mult * r
# walk forward, exit at first tp/sl/expiry