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

Symptom that you fell into the trap: apparent signal strengthens monotonically with loc quintile; collapses when you test adjacent cells.

Full reference: findings/methodology/02-label-leakage-bar-local-scaling.md.

---

3. Shuffled-null design (3 null types — get the right one)

Shuffled-null tests are mandatory before trust, but the choice of what to shuffle is a design decision.

Hypothesis class	Shuffle WHAT	Session example
"Feature X predicts outcomes"	Shuffle the feature values	Phase F-B (used wrong null, "falsified" a real signal)
"Trigger pattern fires at informative times"	Shuffle the trigger mask (preserve fire-rate, move locations)	Phase C (validated ngram_triple_fast_up at z=+5.74)
"Filter improves selection"	Shuffle which trades pass the filter	Phase L-C (evaluated filters against N-size random draws)

Rule: ask "what is the alternative hypothesis, in one sentence?" If you can't state it, you don't know what you're testing.

Common mistakes:

• Using feature-shuffle when testing a trigger pattern → destroys temporal structure the pattern depends on → real signal looks worse than shuffled noise
• Under-tight null (null std huge relative to observed effect) → no statistical power
• Over-tight null (too few permutations) → unreliable z-estimates; use ≥100 for z<3, ≥1000 for z<2

Full reference: findings/methodology/03-shuffled-null-design.md.

---

4. Agent significance corrections (z-scores are overstated 2-3×)

LLM agents systematically overstate z-scores. Treat agent-reported p-values as upper bounds.

Three overstatement patterns:

1. Ignored multiple-testing burden: agent tests 25 variants, reports z=2.43 vs nominal 1.96 threshold. True Bonferroni threshold is sqrt(2 * ln(N)) — for N=25 that's z>2.8.
2. Confused sample-mean z with binomial-proportion z: 53.5% vs 50% on N=840 gives z≈2.0 not 4.2.
3. Extremum-of-K treated as single test: "top combo from 17,280" has expected null-max null_mean + null_std × sqrt(2 ln K) ≈ null_mean + 4.5σ. An observed tw that's below that expectation is not a finding.

Always verify:

• How many implicit tests did the agent run?
• Re-derive z yourself: (real - null.mean) / null.std
• Bonferroni threshold for K tests: z > sqrt(2 * ln K)

Trust thresholds:

• z > 5, N > 500: likely real, test further
• z in [3, 5]: promising, mandatory gate validation
• z in [2, 3]: suspect, require adjacent-cell gradient + null test
• z < 2: treat as null

Full reference: findings/methodology/09-agent-significance-corrections.md.

---

5. Record-keeping discipline (append-only ledger + audit folders)

Every investigation — positive or null — must produce a permanent, discoverable record.

3-layer architecture:

findings/
├── evolution/
│   ├── evolution.jsonl              # append-only ledger
│   └── audits/
│       └── YYYY-MM-DD-slug/
│           ├── CLAUDE.md            # navigator
│           ├── verdict.md           # plain-English conclusion
│           ├── CHRONICLE.md         # narrative (for major findings)
│           ├── <reproducer>.py      # script that regenerates headline numbers
│           └── <artifact>.json      # raw telemetry
└── methodology/                     # universal principles

Ledger entry fields: id, date, status, supersedes, superseded_by, headline, key_numbers, evidence (file paths), sha256_results.

The supersedes pattern: when a later finding replaces an earlier one, ADD a new entry with supersedes: "OLD-ID"; UPDATE the old entry with superseded_by: "NEW-ID". Do NOT delete the older audit folder.

Full reference: findings/methodology/07-record-keeping-discipline.md.

---

6. Post-mortem-before-abandon

Before declaring a signal dead, enrich every trade with causal pre-entry features and hunt filters on individual losses. A "sometimes works" signal is often a filterable signal in disguise.

Pipeline:

1. Run the signal across full history; collect N trade outcomes
2. Compute ~20-30 causal features at each trigger bar
3. Emit per-trade parquet + CSV (one row per trade)
4. Ship to multi-lens agents (see Skill B)
5. Each agent hunts filters that separate winners from losers
6. Evaluate filters against shuffled-null (see §3)

Kill-selectivity metric: losers_killed / max(1, winners_killed). < 1.0 = harmful; 1.0-1.2 = marginal; 1.2-1.5 = useful; > 1.5 = strong.

Session example: +0.178 bps baseline → +0.514 bps after Phase-L filter. 2.9× lift from enrichment-driven filter hunt.

Full reference: findings/methodology/06-per-trade-enrichment-postmortem.md.

---

Confirmation counts (provisional, as of session ca9d7ffa)

Principle	Confirmed	Notes
1. causal-feature-invariant	18+ (every phase)	Fundamental; drop only with proof
2. label-leakage	2	Directly caught spurious "lower-rejection-at-bottom"
3. shuffled-null-design	4	Phase F-B wrong-null, Phase C right-null, Phase L filter-null, Phase M mgmt-null
4. agent-sig-corrections	5+	Combinatorialist, transition-asymmetry, trade-mgmt agents all overstated
5. record-keeping	5 ledger entries	Full chain for NGRAM3FU-STRADDLE
6. post-mortem	1	Phase L delivered the filter; needs re-confirmation on other campaigns

Higher confirmed = more trustworthy. Principle 6 has only one confirmation and should be treated as provisional.

---

Post-Execution Reflection

After invoking this skill:

1. Did applying a principle catch a bug or false positive? Increment its confirmed count in the table above; note the session where it fired in references/evolution-log.md.
2. Did a principle fail (bad guidance)? Demote it in the table; add a superseded_by pointer in references/archive/ with resurrect_if: conditions.
3. New trap that isn't covered? Draft a new section here and append to the evolution log.
4. Never silently move on. This skill's value compounds only if reality-corrections flow back.