linkedparticles-core 1.138.1

linkedparticles-core 1.138.1 — the Particles Client layer

The store-free Client layer of the Particles reference
implementation: schema models with calibrated-confidence
invariants, extraction of claim-granularity candidate particles,
conformance validation, and the pure interchange codec.

This repository is a published view of a private development upstream,
advanced by per-release scrubbed exports. Pull requests are reviewed here
and land by import, with each author's identity and sign-off preserved
verbatim; CONTRIBUTING.md documents the flow.

Changes in 1.138.1:

Added

• The memory benchmark's ablation knobs are now flags —
  particles benchmark memory gains --top-k, --qa/--no-qa,
  --consolidation, --dedup-judge and --reuse-stores, so each ablation
  named in deferred list is one re-run with one knob flipped.
  --consolidation runs the dream cycle's pass list against each
  scratch store before retrieval (the §11 controlled instrument);
  --dedup-judge applies the co-evidential judge so the ranker collapses
  paraphrase pairs inside top-k. Both are recorded on the run tuple, so an
  ablation arm can never be mistaken for a headline run.
• --reuse-stores replays a persisted store set instead of re-extracting,
  turning a read-side ablation arm from a full re-extraction into zero
  write-time LLM calls. A run given --store-dir now stamps that set with the
  write-side tuple that produced it, and --reuse-stores refuses a set built
  under a different dataset, question selection, extraction/embedding model, or
  write-time reconciliation config — a replay is only cheap if it is still
  comparable.
• --no-qa reports the retrieval stage alone, which over reused stores
  makes no LLM call at all.

Changed

• The benchmark's cost estimate covers the ablation passes. A
  --consolidation run projects its capped probe ceiling; a --dedup-judge
  run discloses the judge fan-out as an unbounded component it cannot project,
  and now requires confirmation on that basis alone rather than sailing past
  the call-count threshold at "~0 calls".
• The recorded run tuple carries every knob an ablation can flip — the cap, extract-time duplicate suppression, exact-duplicate auto-merge,
  and the content-age decay rule in effect for CONVERSATION. Two arms of one
  ablation previously recorded byte-identical tuples.

Fixed

• A memory-benchmark run now discloses when decay is inert. The shipped
  content_age_decay.sources table configures no CONVERSATION rule, so every
  particle the benchmark mints scored at recency factor 1.0 — a decay on/off
  comparison would have read as "decay changes nothing" when decay never ran.
  The run says so in its own quality notes.
• run_consolidation accepts a per-store cycle-lock path. The lock's scope
  is the store, but it defaulted to one global path — so a caller holding many
  stores concurrently had every store after the first record skipped.

Particles: provenance-tracked, calibrated-confidence knowledge for AI agent systems — https://linkedparticles.org