Skip to content

Particles Whitepaper — Revision History

The per-round change log formerly appended to whitepaper.md, relocated here so the whitepaper stays focused on the standard. These notes serve reviewers diffing one revision against another; first-time readers can ignore them.

Appendix: Revision History

Each whitepaper revision corresponds to a round of internal and / or external review. Round summaries are recorded here so reviewers diffing against a prior revision can see what changed and why without scrolling through unrelated front-matter.

v2.2 (Draft, May 2026)

Initial v2.2 substantive pass — added §3.3 "Subjects: The Knowledge Graph Backbone" introducing the canonical-entity catalogue and the LLM-resolved external-ontology alignment that had been load-bearing since v0.4 but unstated in the whitepaper. Expanded §3.4.1 from co-evidential-only to the full relation-kind registry (CO_EVIDENTIAL active; CONTRADICTS, BOOSTS, QUOTES, REPLIES_TO, MENTIONS reserved). Added §3.9 "Link-Shaped Sources Preserve Curation Signal" documenting link-shaped-post follow pipeline. Updated §3.1.2 (paragraph-bounded structural chunking), §3.5 (Logseq exporter + cross-exporter synthesis cache ), and §3.7 (calibration shipped mechanically ). Added the 1.0.0 stability commitment to the front-matter governance framing. §3 expanded from 10 design principles to 11 — readers comparing against v2.1 should expect a one-section drift from §3.4 onward.

v2.2 r2 — Gemini pre-review pass

Softened §1.2's framing of consistency checking from "solves" to "bounds"; hoisted §3.1.3's single-fixture baseline caveat earlier; added a combinatorial-context-limit trade-off to §2.2 and Risk #4; added a concrete Wikidata example to §3.3; clarified §3.7's pipeline-versus-data-coverage distinction; sharpened the §2.2 / Risk #9 framing of pre-rendered wiki articles as a compiled read-view over the particle store.

v2.2 r3 — ChatGPT pre-review pass

Added a §3.3 "identity is contractual, not solved" paragraph naming the disambiguation failure modes (name collisions, top-result brittleness, ontology drift) and the standard's posture (deterministic / overridable / provenance-tracked); restructured §3.4.1's prose to split governance (the registry's interop guarantee) from shipped functionality (CO_EVIDENTIAL only), removing the "shipping in name only" reading; reframed §3.9 from "editorial intent" to "curation signal" with an explicit non-goal disclaiming intent modelling and a paragraph on envelope-vs-target attribution; added §3.6 worked examples (near-duplicate news + SEC, conflicting numismatic weights); added a §6 KG-column caveat acknowledging that production triplestores carry provenance / confidence in practice; restructured the 1.0.0 stability paragraph into four explicit commitments and moved CLI migration mechanics out of the front-matter; hoisted the compiled-read-view framing into "What Particles Is"; added a §3.7 worked example reporting the real calibration run on the v0.21.0 Numista seed (ECE 0.0471 → 0.0953 at T = 10.0 ceiling).

v2.2 r4 — Claude pre-review pass + new code anchors

Restructured the §3.7 calibration vignette around three benchmark suites (Numista, Reddit, HN) rather than one, replacing the single-fixture caveat narrative with a three-row table that demonstrates calibration delivers ~85–89 % ECE reduction on LLM-driven extractors while admitting the bounded-fit ceiling is the binding constraint across all three runs. Anchored §3.3's "identity is contractual, operator-overridable" claim to the shipped particles subjects split verb (v0.43.0) rather than to a deferred R1.1 roadmap item. Anchored §3.9's curation-signal principle to the shipped particles corpus links list audit verb (v0.42.5) — the first downstream consumer of corpus_follow_edges, converting §3.9 from architecture-only to architecture-with-payoff. Clarified "What Particles Is" as "narrower than a formal-ontology knowledge graph" so the §3.3 sparse-property-graph framing doesn't read as self-contradictory. Added a §3.4.1 sentence explaining symmetric/asymmetric as the storage-invariant axis. Acknowledged the §3.9 depth-1 cap as shape-blind (a Reddit-to-Reddit follow is one hop just like a Reddit-to-news follow). Renamed the 1.0.0 stability axes to "namespace-frozen at 1.0" / "behavior-additive at minor" so the two senses of "frozen" no longer overload; added the unknown-fields-ignored backward-read story. Added a §4.3 "An operator's week" walkthrough so the architecture has an operational shape and not just an architectural one. Named the two-paths architecture (structural ops vs presentation ops) explicitly under §2 so readers don't have to assemble it across §3.6 / §3.7 / §4. Reframed the §1.1 Karpathy gist callout from "circular validation" to "dogfooding" — the SDK ingesting its motivating source is not evidence the SDK is correct. Moved the revision-history blocks to this appendix so the front-matter no longer competes with the substantive content for a cold reader's attention.

v2.2 r5 — Retrospective-parser truth sync (2026-06 review F1.1)

Synced §3.5 and §4.1 with the shipped reality the 2026-06 external review's F1.1 flagged: the retrospective onboarding path shipped in v0.31.1 as particles import vault, but both sections still described it as "not yet built." Both now describe the shipped mechanism (deposit existing notes unmodified, convert via the standard extraction pass). Retired the "treat markdown content as low-confidence particles with implicit provenance" mechanism promise from §4.1 and recorded why it was rejected rather than built: sentence-split prose yields context-dependent fragments instead of self-contained claims, and uniform stamped confidence is the Risk #11 metadata-theater failure mode. No design change — the sections now describe what the SDK does.

v2.3 (June 2026) — epistemic-stance thesis + prior-art pass (2026-06 review F1.1 / F1.2)

Added the four-paragraph epistemic-stance block to "What Particles Is": the founding premise (no fundamental fact/opinion distinction — only claims true for a group of observers, for a period of time); the two-scopes split (temporal scope carried on the particle, observer scope computed at read time through the trust layer, per the substrate-plus-lens invariant and the shareable lenses of the trust layer; the Cyc / Wikidata / nanopublications lineage with LLM extraction plus the trust model as the missing enabler; and the anti-relativism framing — the fact/opinion distinction recovered as a measurement, with the per-claim contestedness signal deferred as future work. Added the closest prior art the 2026-06 external review flagged as missing (F1.2): a Nanopublications column in the §6 comparison table, an extension of the fairness caveat (production nanopub datasets do carry confidence scores; the differentiator is standardized calibration provenance plus runtime trust weighting), a note on adjacent commercial agent-memory systems (Zep/Graphiti, Letta, mem0), the claim-granularity lineage (Cyc microtheories, nanopublications, micropublications) in §1, and the two corresponding entries in the techspec reference list. Reframed §4.1 to lead with the shipped linter surface and gave the "12 % self-contradictory" hook its scope caveat — same-source contradiction detection ships, cross-source detection is open work — re-ordering the honesty without deleting the vision (F1.1). Refreshed stale anchors: v0.2 → v0.61 implementation framing, one-suite → three-suite benchmark coverage (§3.1, §3.1.3, §3.7), the §3.4.1 registry as-of, and the §7.1 future-work hedge updated only where reality changed (interchange, federated query, and trust lenses have shipped; the adversarial multi-agent protocol has not). Stepped the document version to v2.3 rather than another r-pass because the thesis block is a substantive addition, not a review-response edit.

v2.3 r2 (June 2026) — roadmap-pointer sync

Editorial truth-sync, no design content changed. The change restructured the companion roadmap.md from RFC 2119-keyword prose sections into per-milestone gate tables plus a forward queue. The Reader's Guide row for the roadmap now describes that shape (tables with a lint-checked done column as the status record, plus the forward queue) instead of the retired section organisation. Two pointers claiming the reference extractors are "catalogued in roadmap.md" (§3.2 and Risk #8) now point at the techspec §14.4 reference-extractors table — the actual catalogue — since the roadmap only ever carried a handful of extractor work items, and post-restructure carries rows only.

v2.3 r2 — final owner review (June 2026)

The owner's final read before the 1.0.0 cut, closing the milestone gate. Reframed the thesis block's second paragraph around the facts-on-the-particle / judgments-in-the-lens line: the particle carries temporal facts; recency decay is a read-time judgment (operator-level today; lens migration deferred), and the trust layer's distrust-only scope is disclosed in place (the positive endorsement half is a later decision's child). Led the use-cases paragraph with the belief-ledger framing the README and docs index adopted, naming the first-party use — the AI system's own belief substrate — ahead of the linter and wiki surfaces. Simplified §4.1 to describe what is built and what is intended, dropping doc-history narration per the standing editorial rule that this appendix is the only home for draft history. Relabelled the §6 cost row to "Time to first useful store" (the old label priced Particles' ingestion against competitors' modelling). §6 fairness fixes: the nanopublication human-readable cell softened to Partial (viewer tooling on both sides; the substrate difference stated in-cell), and the agent-memory note now concedes episode-level references and validity intervals, narrowing the differentiator to the immutable content-addressed corpus, calibration provenance, and re-extraction. P1-1 (the thesis resting on the deferred contestedness instrument) was resolved by prioritizing the instrument rather than softening the claim — promoted to a proposed ADR.

v2.3 r3 (June 2026) — relation-registry + version-anchor truth-sync

Editorial truth-sync, no design content changed. Replaced the stale reference-SDK version anchors (v0.61) with version-free phrasing in the §1 design-vs-implementation note, the §3.4.1 registry guarantee, and the §3.7 calibration vignette — the implementation has since moved past that release, so the prose no longer pins a specific SDK version. The v0.43.0 "since" / "measured at" anchors stay: they record fixed historical milestones, not an as-of-now state, and the earlier v0.2 → v0.61 mention in this appendix is itself history, left verbatim. Integrated the narrative relation kinds into §3.4.1, which had predated them: the registry table gains PART_OF / SEQUENCE_IN rows (both asymmetric, ACTIVE) and the end-to-end guarantee now names all three ACTIVE kinds instead of describing CO_EVIDENTIAL as the only shipped one.

v2.3 r4 (June 2026) — contestedness instrument shipped (P1-1 closure)

Replaced the thesis block's "(Surfacing this per claim — a contestedness signal — is deferred work.)" parenthetical with a pointer to the now-shipped mechanism: contestedness is the max−min spread of effective confidence across the viewer's policy set (local policy + each adopted lens, each evaluated standalone), computed at read time and surfaced in query responses, prose exporters, and lint — disclosure, never a discount ( techspec §6.9). This closes 2026-06 review P1-1 by delivering the instrument the thesis promised rather than softening the claim: the fact/opinion distinction is now recovered as a per-statement measurement, the sibling of stance distribution. No design content otherwise changed.

v2.4 (July 2026) — §4 rewritten for the agent-memory arc (R1.14–R1.16)

Rewrote §4 "Primary Use Cases" end to end. The section predated the R1.14–R1.16 agent-memory arc and still described the product as "Linter for Engineers, Wiki for Everyone"; it now leads with the launch wedge — AI memory you can audit and trust — organised as two front doors over one substrate, with the shared-store bridge ("one epistemic store serves both the human operator's second brain and the AI agent's memory") as the architecture narrative beneath the wedge and the consent mechanisms (write-authority boundaries via asserted_by, open §6.6 reconciliation, read-time lenses) stated explicitly. §4.1 became the audit-and-harvest loop: the particles init claude-code lifecycle hooks, the drift-gated MEMORY.md projection, the first-run particles audit census as the activation moment, the nightly particles memory consolidate dream cycle with the mutable local-source refresh and utility rank-lift , the --as-of bitemporal read lens, and the reference memory-server compatibility façade. §4.2 became the cited-projections door, absorbing the former wiki-view section plus the exporter family, the Obsidian lint-callout plugin , and the retrospective vault import with its rejected-parser rationale carried forward. §4.3 "An operator's week" was updated to the shipped cadence: continuous zero-touch harvest, nightly scheduled consolidation, weekly curate/review/export. The §2 "two paths through the inversion" callout and Risk #2's mitigation column were synced to the new door names.