Improving long-window recall: complementary approaches

[alltheseas]

Problem

Algorithm recall drops from 83% at 7d to 24% at 1yr. Relay retention explains only ~1/4 of this (77% median availability at 1yr). The rest is relay selection and discovery. No single fix closes the gap — but five complementary layers do.

Approach stack

Each layer addresses a different bottleneck. No contradictions between them.

Layer	What it solves	When it helps	1yr recall	Status
Stochastic scoring (Welshman)	Spreads queries beyond popular relays	Session 1, no state	24%	Deployed in Coracle
NIP-66 liveness filtering	Stops wasting connections on dead relays	Session 1, no state	~24% + efficiency	#14
NIP-65 archive marker	Discovers relays not in current list	Session 1, no state	Unmeasured	nips#2243
NIP-66 retention-depth	Informs relay selection before learning	Session 1, no state	Unmeasured	#21
Thompson Sampling	Learns which relays actually deliver	Session 2+	92%	#19, benchmarked

Key findings

• Relay retention is decent — 77% median at 1yr across 277 relays (Relay history problem: 77% availability vs 24% recall at 1yr #21, deno task retention)
• Historical NIP-65 walk-back is infeasible — kind-10002 is replaceable, old relay lists are gone (bench/probe-historical-nip65.ts)
• Selection > discovery — Thompson reaches 92% using only current NIP-65 relays, meaning most of the gap is selection quality, not missing relays
• Archive + retention-depth improve session 1 — the 24% starting point before Thompson learns

Related issues and PRs

• NIP-66: client liveness reporting via library opt-in #14 — NIP-66 client liveness reporting (Layer 0+1)
• Layer 2: Recall tag for NIP-66 client observer reports #19 — NIP-66 recall tag (Layer 2)
• Relay history problem: 77% availability vs 24% recall at 1yr #21 — Relay history problem: 77% availability vs 24% recall
• Add retention profiling, fix report accuracy, reframe real ceiling #22 — Add retention profiling, fix report accuracy, reframe real ceiling
• nips#2243 — NIP-65 archive marker proposal