Adaptive Learning Queue / Pool Ladder

Type: algo Origin repo(s): reincarnation (canonical, survey items), vela (divergent variant, image library curation) Extraction readiness: needs extraction (algorithm is clean; persistence schema is shared across both variants) Depends on: append-only signal capture (user responses that drive re-pooling) Last reviewed: 2026-04-24

What it is

Items (survey questions, images, or any rankable unit) are distributed across named pools forming a ladder — e.g. INFINITY → D → C → B → A → PURGATORY — based on accumulated response signals. A background batch scorer runs every N hours, recomputes per-item virtue scores using seeded randomness + thresholds, and promotes or demotes items across pool boundaries. Users querying for the "next item" see a probability-weighted sample drawn from the current pools, so adaptive reordering happens naturally as more responses come in.

Dual-owner note. Reincarnation drives survey items; vela uses the same shape to curate its image library. Both owners are valid; learnings from one may apply to the other but they have diverged intentionally and neither needs to chase the other.

Data shape

• items — one row per rankable unit with current pool column (enum).
• item_signals — append-only per-user/per-item feedback events (P32).
• pool_transitions — append-only audit log of promotions/demotions (P31 append-only pool ladder).
• scorer_runs — metadata per batch run (started_at, items_scored, transitions_made).

UI / surface shape

• Admin view: pool populations (how many items in each), transition history.
• Consumer surface (/api/next-item?scope=X): returns a probability-weighted item from the current pools.
• Editorial override: operators can pin/demote manually (P28 sticky transitions so auto-scoring doesn't reverse curator intent).

Variants in the wild

• Reincarnation — canonical. Pools represent psychometric quality (item discrimination + response volume). Scorer uses item-total correlation.
• Vela — same ladder shape applied to image aesthetic coherence. Scorer uses entropy + Inception-v3 similarity (P27 cluster coherence).

Primary files in origin

• server/pools/ladder.ts (reincarnation) — pool definitions, transition rules
• server/pools/scorer.ts — batch scorer
• server/pools/sampler.ts — probability-weighted next-item sampler
• db/schema/pools.sql — shared schema shape (P31 append-only pool ladder)

Next-version notes

• The batch scorer cadence is fixed (every N hours); adaptive scoring (run when enough new signals accumulate) would be more efficient.
• Sticky transitions (P28) need more visibility — it's currently hard to tell whether an item's pool is curator-pinned vs. algorithm-placed.
• The reincarnation and vela variants share the persistence shape; worth a common package in the consolidated repo even though the scorer implementations differ.

Related patterns

• P31 — Append-Only Audit History with Pool Ladder
• P27 — Entropy-Based Cluster Coherence (vela variant)
• P28 — Editorial Lifecycle State Machine with Sticky Transitions
• P32 — Append-Only Signal Capture
• P34 — Multi-Stage Cron with Per-Stage Error Isolation (scorer orchestration)