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 C is non-negotiable. Never report a positive without time-split OOS.

Classify verdicts:

• All pass → validated-cross-asset (promote)
• A+B+C pass, D fails → validated-asset-specific (narrow scope, don't kill)
• C fails → kill and archive with resurrect_if: conditions

Full reference: findings/methodology/05-serial-adversarial-gates.md.

3. Multi-lens agent synthesis (4-5 parallel specialists)

When brute force is infeasible, launch 4-5 specialist agents in parallel with distinct analytical lenses. Convergence = validation; divergence = diagnostic.

Canonical lens set:

1. Pattern-motif / n-gram
2. Regime / trend classifier
3. Morphology / wick hunter
4. Information theorist / surprisal
5. Hidden-signal hunter / skeptic critic — ALWAYS include; prevents confirmation bias

Rules:

• Same raw data, different lens prompts
• Each agent briefed: "test against shuffled-null, estimate multiple-testing burden, require null-test z > 2 AND per-year stability"
• 5th agent explicitly asks: "prior 4 agreed; do you concur or dissent?"
• ≥3/5 convergence = likely real; only 1 agent = likely lens-specific artifact

Disagreement is information. If two agents find different signals, test both independently.

Full reference: findings/methodology/04-multi-lens-agent-synthesis.md.

4. Per-trade enrichment for loss postmortem

When a signal works sometimes and fails sometimes, convert aggregate question → row-level question.

Pipeline:

Step 1 — Run signal across full history → collect N trade outcomes
Step 2 — Compute 20-30 causal features at each trigger bar
Step 3 — Emit parquet (per-trade × feature matrix)
Step 4 — Ship to multi-lens agents (§3)
Step 5 — Each agent hunts filters
Step 6 — Evaluate filters via shuffled-null + OOS

Must-have feature categories (at least one from each):

• Trend/regime (SMAs, slopes, autocorrelations)
• Volatility regime (rv, range compression, Kaufman ER)
• Pattern context (quintiles at trigger + lags)
• Exhaustion/cluster (bars-since-last-signal, signal-count-in-window)
• Position (dist-to-swing-high ATR-normalized)
• Calendar (hour, weekday, month, session)

Session example: turned +0.178 bps baseline → +0.514 bps filtered (2.9× lift).

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

5. Agnostic-null cascade (orthogonal null retest)

When a signal passes one null type, retest with an orthogonal null before trust. Example from session:

• Phase F-B shuffled the vol-forecast column → +0.370 signal "lost" to z=−5.15 under that null
• Phase C shuffled the trigger mask (proper null) → +0.439 signal passed at z=+5.74

If Phase C had been the only test, we'd have trusted a potentially wrong null. If Phase F-B had been the only test, we'd have missed a real signal.

Rule: whenever a positive emerges, identify AT LEAST ONE orthogonal null and retest. Common orthogonal pairs:

• Feature-shuffle AND mask-shuffle
• Permutation AND block-bootstrap
• Shuffled-selection AND parametric (e.g., binomial tail)

If signal passes both, trust more. If it passes one but fails the other, investigate why — the divergence tells you something about the signal's mechanism.

6. Compute orchestration (pueue + wrapper scripts)

Long parallel sweeps go through pueue on gpu-host-1 with a specific discipline.

Anti-pattern (silently fails, 0-second "successes"):

ssh gpu-host-1 '~/.local/bin/pueue add -- bash -c "VAR=1 python script.py"'

Correct pattern:

cat > /tmp/run.sh <<'EOF'
#!/usr/bin/env bash
set -euo pipefail
export VAR="${VAR:-default}"
cd /tmp
exec uv run --python 3.14 --with numpy --with pandas python -u /tmp/script.py
EOF
scp /tmp/run.sh gpu-host-1:/tmp/
ssh gpu-host-1 'chmod +x /tmp/run.sh && ~/.local/bin/pueue add --group mygroup --label "job" -w /tmp -- /tmp/run.sh'
ssh gpu-host-1 '~/.local/bin/pueue wait <id> --quiet; tail -50 ~/.local/share/pueue/task_logs/<id>.log'

Key points:

• Wrapper script (never inline bash -c) — pueue's sh -c wrapping strips inline quotes
• python -u for unbuffered stdout (otherwise logs stay empty for 40 min)
• pueue restart --in-place <id> (not bare restart, which creates new task IDs)
• Use fork multiprocessing context to share pre-computed arrays via COW

Full reference: findings/methodology/08-compute-orchestration-pueue.md.

See also: Skill(devops-tools:pueue-job-orchestration) for extended patterns.

Confirmation counts (provisional)

Post-Execution Reflection

After invoking this skill:

1. Did a pattern catch a bug, save time, or produce a validated finding? Increment confirmed in the table above; note the session/audit folder in references/evolution-log.md.
2. Did a pattern fail (misled you, wasted compute)? Demote in the table with a brief note; add to evolution log with link to the failing session.
3. New execution pattern emerged? Draft a new section + append to evolution log.
4. Compute anti-pattern caught you again? Re-check §6 wording — clarify the trap.