# Stage 110: Bayesian Lift & N_eff (aq110_compute_bayesian_attribution / DB-UNIT-09)

Computes 0-call paired within-root_task_id lift over bare_baseline, Empirical-Bayes shrinkage, win rate, and Kish effective sample size across 6 factor dimensions.

- Atomic micro-queue: `aq110_compute_bayesian_attribution` (visibility: `public_commons`)
- Backing database unit: `DB-UNIT-09` (`Agent Fees, AFK Budget Holds, CPOP Review & Q12 Royalty Clearinghouse`)
- Database cell: `ledger-db.teleon.dev` (`teleon_core.agent_fee_tariffs`, `teleon_core.agent_fee_quotes_and_holds`, `teleon_core.agent_fee_settlement_splits`, `teleon_core.agent_fee_savings_evidence`, `teleon_core.agent_fee_csr_tax_ledger`, `teleon_core.cpop_orthogonal_review_verdicts`, `teleon_core.teleon_double_entry_royalty_journal`)
- Claim row hydration: `single_self_contained_flat_task_row` (100% context inline, normalized writeback)
- Track 2 Sister queue (`SISTER-09`): `sq09-improve-ledger-queue.teleon.dev` (Tune off-peak AFK discounts, Shapley/Q12 attribution weights, Hebbian decay, and CUSUM drift windows.)
- Lanes: Paired Delta vs Bare, Empirical-Bayes Shrink, Kish N_eff Leaderboard

## Execution engines & AI Agent freedom contract

- **Deterministic (0 model calls)**: `q12_ledger_and_cusum_engine_v1` (Paired delta, Empirical-Bayes shrinkage & Kish N_eff engine (0 calls))
- **Hybrid (<=1 model call)**: `cpop_evidence_audit_summarizer_v1`
- **AI Agent & Bot Fleet (1..6 calls)**: `cross_family_independent_reviewer_v1` (Winning factor synthesizer & default-policy recommender)
  - Contract: **Instructions + Resources + Freedom to search for new resources and new ways not bound by instructions**

## Database invariants

- Ledger and fee-hold rows run under SERIALIZABLE ACID_COMMIT_KERNEL_POLICY with zero inline agent middleware.
- A producer family never approves its own work (reviewer_family != producer_family).

## External & internal bot fleet actions

- Compare mutated siblings strictly against their paired bare_baseline within each root_task_id
- Shrink noisy small-N factor estimates toward the grand prior via Empirical Bayes
- Discount correlated monoculture runs via Kish effective sample size N_eff

## Sister self-tuning areas

- Tune off-peak AFK congestion discounts and 6-bucket Q12 royalty weights (DIM-08)
- Calibrate Hebbian learning rate and CUSUM drift sensitivity on historical anomalies (DIM-01)
- Add deterministic Sybil-collusion graph detectors (DIM-06)