← ClaudeAtlas

bayesian-aggregationlisted

Combining probabilistic judgments from several agents into one calibrated probability or distribution: verbal-to-numeric conversion, linear vs logarithmic opinion pools, extremizing, weights from calibration history (Brier and log score, shrinkage, Hedge), Bayesian model averaging, recalibration, correlated experts and effective N, multi-outcome and quantile pooling, with a stdlib script and worked numeric examples. Load whenever you merge probabilities, confidences or estimates from multiple agents.
rustcircle/c101 · ★ 0 · AI & Automation · score 65
Install: claude install-skill rustcircle/c101
# Bayesian aggregation Several agents gave probabilities for the same event. The pooled number must be better than any single input on average, honest about disagreement, and not fooled by five copies of the same model agreeing with itself. The procedure: make the inputs comparable, pick the pool, weight by track record, correct for correlation, extremize only when justified, sanity-check, report. Tool: [scripts/pool.py](scripts/pool.py) (stdlib Python 3.9+, JSON output): ```bash STACK="$HOME/.claude/stack"; [ -d "$STACK" ] || STACK="${CLAUDE_PROJECT_DIR:-.}" P="$STACK/.claude/skills/bayesian-aggregation/scripts/pool.py" python3 "$P" binary --p 0.6 0.7 0.9 [--w ...] [--d 1.5] python3 "$P" multi --dist 0.5,0.3,0.2 0.6,0.3,0.1 --labels A,B,C python3 "$P" clusters --p 0.9 0.9 0.6 --cluster m1 m1 m2 --rho 0.8 python3 "$P" weights --ledger .stack/calibration/ledger.jsonl [--scheme skill|hedge] python3 "$P" recal --ledger .stack/calibration/ledger.jsonl --agent fact-checker python3 "$P" neff --n 5 --rho 0.6 python3 "$P" verbal "very likely" "roughly even chance" python3 "$P" quantiles --q 80,120,200 100,150,260 90,130,400 --levels 0.1,0.5,0.9 ``` ## 1. Preconditions: is pooling even legitimate? - [ ] **Same question, same resolution criterion.** "Ships by June 30" and "ships in Q2" are different events. Rewrite every input against one resolution sentence, or do not pool. - [ ] **Same outcome space.** For multi-outcome questions all agents use the same exhaustive, m