Is 'append-only log as source of truth, record separated from verdict' a genuine cross-domain convergence, or a two-point coincidence?

This capture directly answers the open question logged at question-append-only-ledger-cross-domain-convergence, raised after the vault noticed that CANONIC (AI content governance) and ActiveGraph (AI agent memory) — two 2026 papers in subfields that do not cite each other — both built an append-only log as ground truth with judgment held separately outside it (see claim-canonic-deliverable-is-an-append-only-evidence-ledger and claim-activegraph-rebuilds-babyagi-on-event-log). The question asked for either a third independent instance or a common-ancestor check. This run found both, and they point the same direction: the pattern recurs, but it is not a mysterious independent-invention coincidence — it is the diffusion and re-application of an already-named, two-decade-old architectural pattern (event sourcing / CQRS) into several 2025-2026 AI subfields, plus at least one older provenance tradition C2PA's designers also drew on.

Claim: ActiveGraph's own paper attributes its append-only-log architecture to pre-existing "event sourcing" and CQRS patterns, not independent invention

Claim type: historical/attribution (what the primary source itself says about its own origins). Tier required: 1-2. Tier obtained: 1.

Nakajima's ActiveGraph paper ("The Log is the Agent," arXiv:2605.21997, 21 May 2026) explicitly disclaims novelty for the append-only-log-as-source-of-truth idea itself: "The borrowed ideas are not new in isolation. Event sourcing and command-query responsibility segregation are established patterns in data systems, and reactive recomputation of derived state from a changing input is the core of incremental dataflow and of spreadsheet engines. The contribution here is the recombination: applying an event-sourced, reactive substrate to the specific problem of long-running agentic systems." This directly weakens the "independent convergence" framing for one of the original two points: the author names the shared ancestor himself.

provenance: Nakajima, Y. (2026). "The Log is the Agent: Event-Sourced Reactive Graphs for Auditable, Forkable Agentic Systems." arXiv:2605.21997, p.3. https://arxiv.org/abs/2605.21997

Claim: Martin Fowler formalized "event sourcing" — an event log as the official record, with application state as a derived/cached projection — in a December 2005 essay, roughly 21 years before CANONIC and ActiveGraph

Claim type: historical/definitional (naming who formalized a settled term and when). Tier required: 3-4 acceptable for definitional use, but this claim is load-bearing for the "common ancestor" argument, so it is held to the escalated Tier 1-2 floor for surprising/contested/load-bearing historical claims. Tier obtained: 1 (author's own venue).

Fowler's "Event Sourcing" page (martinfowler.com/eaaDev/EventSourcing.html, dated 12 December 2005, byline Martin Fowler) states the pattern in terms matching both 2026 papers' framing: "the event logs can be the official record and databases can be built from them whenever needed," and "we are persisting two different things[:] an application state and an event log." Elsewhere the page frames it as a choice — "The official system of record can either be the event logs or the current application state" — but the "log-as-record" option is the one both CANONIC and ActiveGraph took. This gives the pattern a documented, findable common ancestor in mainstream software-architecture writing that predates both 2026 instances by two decades, which is the kind of check the routed question specifically asked for.

provenance: Fowler, M. "Event Sourcing." martinfowler.com/eaaDev, 12 December 2005. https://martinfowler.com/eaaDev/EventSourcing.html

Claim: CANONIC's own related-work section situates its append-only ledger inside an existing lineage of provenance/immutability systems (C2PA, Open Policy Agent, model/data cards, study preregistration, git, Bitcoin), not as a novel invention

Claim type: historical/attribution (what CANONIC's own text claims about its precedents). Tier required: 1-2. Tier obtained: 1.

CANONIC (Hadley, arXiv:2607.05410, 10 June 2026) does not claim to have invented "record separate from verdict" as an idea. Its Related Work table (§6) positions itself against named prior systems on the same axis — including C2PA, whose entry reads "media files... at capture or edit... cryptographic manifest bound to the file... capture device or editor" as the locus of control — and the paper credits git directly for "an immutability guarantee through the append-only commit graph that records every state transition." It also places its ledger inside a longer historical sequence it names explicitly: "the corresponding historical layers in computing infrastructure accreted over decades: the Unix kernel in 1969, POSIX operating-system standards in 1988, the iOS App Store distribution model in 2008, and Bitcoin's immutable ledger in 2009." Like ActiveGraph, CANONIC frames its own move as recombination of known immutable-record precedents applied to a new problem (AI content governance), not first invention of the append-only-record-plus-external-judgment shape.

provenance: Hadley, D. (2026). "CANONIC: Governance Is Compilation." arXiv:2607.05410, §6 (Related Work) and §4. https://arxiv.org/abs/2607.05410

Claim: A third, non-citing 2026 paper in a distinct subfield (scientific-workflow reproducibility for LLM-driven action models) independently applies the same append-only-log-as-ground-truth-plus-separated-judgment architecture

Claim type: technical mechanism (how a third system actually works). Tier required: 1-2. Tier obtained: 1.

R-LAM (Sureshkumar, arXiv:2601.09749, cs.SE, submitted 12 January 2026) — which cites neither CANONIC nor ActiveGraph, and sits in a third subfield (reproducible scientific-workflow automation for "Large Action Models") — independently reaches for the identical shape. The append-only mechanism: "Every executed action produces a provenance record that is appended to a global execution trace," and "an action that is not logged is treated as non-existent[,] ... ensur[ing] that all state transitions contributing to a result are observable, auditable, and reproducible." The record/verdict separation: the framework logs and replays deterministically but explicitly leaves scientific judgment to a human outside the log — "The framework is designed to assist, not replace, human scientific judgment. LAMs are treated as decision-support components whose outputs are mediated, constrained, and auditable, rather than as independent scientific actors," and "Human oversight is required to define permissible actions, validate execution policies, and interpret results." This satisfies the routed question's own stated resolution criterion — "a system in yet another subfield... that reaches for append-only-log-as-truth with the record/verdict separation, arrived at independently" — though, notably, R-LAM's own text does not cite "event sourcing" by name, unlike ActiveGraph's explicit acknowledgment.

provenance: Sureshkumar, S. (2026). "R-LAM: Reproducibility-Constrained Large Action Models for Scientific Workflow Automation." arXiv:2601.09749, §V(C), §VIII(B). https://arxiv.org/abs/2601.09749

Further leads

• C2PA (Coalition for Content Provenance and Authenticity) is a fourth candidate domain for the pattern, but on closer look it doesn't cleanly match: C2PA manifests are cryptographic hash-bindings verified at capture/edit time, not a continuously-appended event log with read-time reconciliation. Worth a dedicated capture if a primary C2PA spec source is checked directly. (Lead only — not verified this run; per CANONIC arXiv:2607.05410 §6 Related Work table.)
• Open Policy Agent / HashiCorp Sentinel and study-preregistration are two more named precedents CANONIC's own Related Work table (§6) cites as partial precedents for "gate at admission, verdict elsewhere" — worth checking each against the append-only-log criterion specifically, since CANONIC's table groups them by "gates at admission" broadly, not by append-only-log mechanism narrowly. (arXiv:2607.05410 §6)
• CQRS (Command Query Responsibility Segregation) is named by ActiveGraph alongside event sourcing as a co-established pattern; not separately researched this run — would need its own primary-source history check (likely also traces to Fowler/Greg Young circa mid-2000s). (arXiv:2605.21997 p.3)
• Scientific-workflow provenance tooling predating 2026 (e.g., the Kepler and Taverna workflow-provenance systems referenced in older e-science literature) is a plausible deeper common ancestor for R-LAM specifically, distinct from software-engineering event sourcing — not checked this run.