ARC-ADR-041 — Pace-Layered RDF↔LPG Projection: a Drift-Sensing Operational Frontier and Dream-Consolidated Ontology Graduation¶

Field	Value
ID	ARC-ADR-041
Status	Accepted
Date	2026-05-30
Deciders	Hub owner (Nicky Clarke) — approved 2026-05-30
Supersedes	—
Superseded by	—
Tags	ontology, pace-layering, lpg, rdf, projection, drift, emergence, graduation, dreaming, annealing, ufo, bfo, ice, provenance

Context and Problem Statement¶

The ontology pipeline already decides most of the loop: the canonical store + reasoner (ARC-ADR-019), the relator-vertex / typed-role-binding representation for n-ary, metadata-bearing relations (ARC-ADR-016), the gate-first ingestion pipeline (ARC-ADR-030), and the sift-sort authoring loop where the LLM proposes and the formal layer disposes (ARC-ADR-032). Every existing projection runs ontology → code (forge, ARC-ADR-029), source → ontology (ingest), or staging-LPG → canonical-RDF (snap).

What is not yet decided is the other direction and its consequence: how the canonical ontology projects down into a fast operational Labeled Property Graph (LPG) that (a) enforces the ontology's schema + rules on high-speed data, while (b) letting operators attach dynamic, rigor-free metadata for operational tailoring — and how useful structure that emerges in that fast layer can graduate back up into the canonical ontology without corrupting it.

The organizing principle is pace layering (Stewart Brand): the ontology is the slow layer (meaning, rules, rigor, changes rarely); the LPG is the fast layer (operations, speed, dynamic properties that need no ontological rigor). Fast learns, slow remembers; fast proposes, slow disposes; the coupling between them must be loose enough that the fast layer can churn without whipsawing the slow one. The fast layer's distinctive job is to measure the value of, and changes in, meaning (semantic telemetry), and to note drift and emergence — which may, optionally, be harvested for attention.

The decision: what is the contract across the slow↔fast membrane — the down-projection, the admission rules for fast-layer extensions, and the up-graduation of emergent structure — such that the fast layer is freely tailorable while the slow layer stays congruent, parsimonious, and provable?

This composes ADR-016/019/029/030/032 and is grounded in the Labs notes Pace-Layering & the Dreaming Ontology, Ontology-Pipeline, Reification-and-Hyperedges, and Evidence as a Primitive. The design was pressure-tested by three independent foundational reviews — UFO/OntoUML, BFO 2020/OBO, and KR&R/knowledge-engineering — whose verdicts converged and drove the final shape (see Review findings).

Decision Drivers¶

#	Driver
D1	Pace separation. Slow ontology vs fast LPG, loosely coupled. Down = constraint; up = a slow, gated ratchet; shear is buffered (the fast layer absorbs volatility).
D2	*The fast layer enforces a shadow* of the rules, never the rules themselves.** Only A-Box instances cross the membrane; T-Box axioms (class expressions, restrictions, `rdf:List`/`owl:intersectionOf` chains) stay in the slow layer. The LPG carries cheap RDFS-level closure, not OWL-DL inferences.
D3	Extensions must be safe by construction. Operator metadata is `ops:`-namespaced, physically cannot shadow a canonical IRI, carries mandatory metadata (owner, purpose, created_at, provenance, TTL, stability, promotion_status), and is garbage-collected when unused.
D4	The fast layer measures value-of and changes-in meaning (semantic telemetry) — but those measurements are Information Content Entities about the record (IAO), never domain universals. Rigor-free + dynamical is correct, not a shortcut.
D5	Popularity nominates; classification decides. A fast statistical signal may flag a candidate; only a slow ontological classifier may assign a category. Letting a fast signal make a slow commitment is a pace violation and the root cause of anti-patterns.
D6	Category soundness. Graduation must branch (datum / intrinsic-moment / relational / occurrent / ICE / admin), not funnel everything toward "class"; disjointness axioms must be ON so miscategorization is rejected, not silently flattened.
D7	Congruence-first / propose-dispose (ADR-032 lineage). Operations propose meaning; the formal layer disposes; nothing auto-promotes; non-snapping candidates are retained + visible, never silently dropped.
D8	Round-trip honesty. The governed core is data-round-trippable only given an explicit invariant set; referential identity is a separate guarantee; the down-projection is asymmetrically lossy (binarized edges drop the time index), so relation reasoning stays in RDF/Common-Logic, never in the LPG shadow.
D9	Bitemporal + PROV audit (ADR-016 D6 lineage). Extensions, promotions, and dream-consolidations are all recorded as provenance events; no silent deletes; "evidence as a primitive."

Considered Options¶

Pace-layered membrane: drift-sensing frontier + dream-consolidated graduation (recommended). Three LPG zones (CANON / FRONTIER / CANDIDACY), one membrane enforcer, a two-stage graduation (NOMINATE fast / CLASSIFY→DISPOSE slow) with category gates, and slow-layer evolution modeled as dream consolidation (annealing + re-sparsification) rather than a per-item ratchet.
Collapse the layers — make the LPG be the ontology (or the reverse). Rejected: violates congruence-first; "storage-shaped truth"; the LPG proves nothing about meaning and cannot reason.
Read-only projection, no operator extensions — clean, but operations cannot tailor, and the whole point (capturing value/drift telemetry from real use) is lost.
Auto-promoting ratchet (frequency-thresholded, single funnel) — the naïve version the reviews refuted: popularity deciding category mints anti-patterns (relator-as-kind, mixin-as-identity, occurrent-as-continuant), thrashes, and corrupts canon.

Decision Outcome¶

Proposed: Option 1. The membrane is the ARC-ADR-032 "propose/dispose" philosophy generalized, with operations as a new proposer (alongside Cerebras-over-documents and structured lifting). Operations may propose meaning; only the slow ontological layer disposes.

Three zones in the LPG¶

Zone	Source	Mutability	Governed by
CANON	deterministic down-projection of canonical RDF; every element keeps its source `iri`	read-only (locked; drift-gated)	the ontology (ADR-016 representation)
FRONTIER	operator `ops:`-namespaced extensions + dynamic telemetry	writable, namespaced, TTL'd, GC'd	the admission enforcer
CANDIDACY	flagged emergent structure under sift/repair/validate	staging	the sift-sort loop (ADR-032)

The membrane (Projection & Admission Enforcer)¶

        ┌──────────── SEMANTIC / SLOW PLANE — canonical RDF · OWL · SHACL · reasoner (ADR-019) ────────────┐
        └───────────▲────────────────────────────────────────────────────────────────┬─────────────────────┘
        promote (gated ratchet / dream-snap, ADR-032)                                  │ project down
        ═══════════ PROJECTION & ADMISSION ENFORCER (the membrane) ═════════════════════▼══════════════════════
        classify every write · namespace · stamp PROV-O · validate (closed-world) · drift-detect · note · GC
        ──────────────────────────────────────────────────────────────────────────────────────────────────────
        OPERATIONAL / FAST PLANE — LPG     [ CANON read-only ]  [ FRONTIER ops: ext ]  [ CANDIDACY staging ]

Down-projection (CANON): RDF → LPG — only A-Box crosses (D2)¶

RDF construct	LPG projection	Rule
IRI resource	node + reserved `iri` (canonical join key)	invertibility precondition (but see D8 / C4)
`rdf:type`	label(s)	T-box hierarchy stays slow; carry RDFS-materialized type closure only
datatype property	node property + `_rdf_datatype` (and `_rdf_lang` for tagged) sidecar	preserves XSD/decimal/lang on round-trip
binary object property	typed edge	edge type via a namespace registry (collision guard)
n-ary / metadata-bearing relation	relator vertex + typed role-binding edges	= ADR-016 verbatim; bitemporal + PROV on the relator
blank node	skolemized IRI + `_rdf_was_blank` + source-graph scope	A-Box only; T-Box blank nodes are not projected
`owl:sameAs` cluster	one elected canonical node + `_rdf_same_as_aliases[]`	*closure runs before* projection** (else duplicate nodes)
inferred triples	tagged `inferred` overlay, excluded from mining	reason slow/open-world; never mine entailments as "novel"

The two-stage graduation (the ratchet, split on pace)¶

FAST ── STAGE 1: NOMINATE (continuous · statistical · rigor-free) ───────────────
  value (freq · fan-out · centrality · demand)  +  drift (SHACL-strain ·
  cardinality divergence · sense-split)  → time-series + hysteresis → FLAG
  (the measurements are ICEs — they NEVER graduate; they only point)
        │ a flagged candidate crosses the membrane
        ▼
SLOW ── STAGE 2: CLASSIFY → DISPOSE (rare · gated · categorical) ────────────────
  A. CATEGORY TRIAGE  ─┬─ ADMIN (cache/UI/TTL) ────────────▶ stays fast, never a term
     (BFO)             ├─ ICE about the record (score/tenant)▶ IAO annotation only
                       ├─ OCCURRENT/TEMPORAL (verified_by…) ▶ OCCURRENT ladder
                       └─ WD (world-descriptive) ─┐
  B. WD FORK ─────────────────────────────────────┼─ datum ─────────▶ stays a property
     (UFO)                                         ├─ intrinsic moment ▶ «quality» | «mode» (1 bearer, NOT an edge)
                                                   └─ relational ─┐
  C. RELATION GATE ────────────────────────────────┬─ formal (taller-than/part-of) ▶ typed edge, never a relator
     (UFO+BFO)                                      └─ material ▶ reify + pick realist shape:
                                                      relational-quality | roles+process | process+ICE (+ bind time-index)
  D. TYPE SORTER ── sortality → kind/subkind | category/mixin (req. ≥2 kinds);
     (UFO+OBO)        rigidity → subkind | phase/role; + genus-differentia,
                      one asserted is_a → BFO via reused CCO/IAO, opaque ID, reuse RO
  E. DISPOSE ── L2 anti-pattern · L5 Alloy · L3 reasoner (disjointness ON) · L4 SHACL ·
                CQs · instance dedup + duplicate-universal check · HITL signs the STEREOTYPE
                            ▼
                CANONICAL TERM ──▶ re-projected DOWN into CANON (read-only)

The recommender (Stage 1) nominates; the classifier (Stage 2) decides. FK-smell, frequency, and fan-out are hints, never gates.

Slow-layer evolution = dream consolidation, not a per-item ratchet¶

The slow layer does not react to each candidate. It mostly sleeps, and periodically dreams: an offline, attention-budgeted consolidation that replays harvested drift/emergence, expands the possibility space (entailment, recombination), then anneals toward a low-energy configuration and re-sparsifies (prunes redundancy, merges duplicate universals, deprecates dead structure). The "energy" being minimized is description length + uncovered drift, subject to hard consistency + category well-formedness — so the UFO/BFO gates are not a checkpoint, they are the shape of the basin the graph settles into. Dreaming is free; crystallization is gated and recorded (every consolidation is a PROV-O event: merged X,Y→Z, pruned W, because…). Re-sparsification is the demotion/GC path; the cooling schedule is the hysteresis that prevents thrash. Full treatment: Pace-Layering & the Dreaming Ontology.

Affected Layers / Repos¶

Layer	Repo	Impact
backend-core	nickpclarke/backend-core	Owns the operational LPG, the membrane/admission enforcer, the FRONTIER + CANDIDACY zones, the Stage-1 recommender, and the dream-consolidation orchestrator; projects canon down and snaps promotions up to Fuseki (ADR-019).
canonical store	Fuseki (ADR-019)	Holds slow-layer truth; runs reasoner (gUFO ‖ BFO) + SHACL; the only arbiter of meaning.
forge	nickpclarke/agentarmy-forge (ADR-029)	Emits the deterministic CANON down-projection DDL/mapping (drift-gated), mirroring the RT5 "generate DDL from modelgen" pattern.
persistence	ArcadeDB / (candidate: Postgres + Apache AGE)	Fast-layer LPG. AGE is a strong candidate: JSONB for rigor-free metadata, SQL for value/drift measurements, pgvector for sense-drift/dedup, RLS/triggers as the admission membrane (OQ3).
(agents)	hub `.claude/agents/`	`ontologist-ufo`/`ontologist-bfo` own the category classifier + the energy landscape (anti-pattern catalog, disjointness); `knowledge-engineer` owns the recommender, identity resolution, and dream runs.
(cross-cutting)	docs/contracts.md	New rows for the projection/mapping contract + the FRONTIER extension-metadata schema.

Pros and Cons of the Options¶

Option 1 — Pace-layered membrane (recommended)¶

Pros: congruence-first preserved by construction (operations cannot mutate canon meaning); the fast layer is freely tailorable and is exactly where value/drift telemetry belongs; reuses ADR-016 representation, ADR-030 gate, ADR-032 propose/dispose, ADR-019 reasoning; the dream model gives thrash-resistance (cooling = hysteresis), demotion/GC (re-sparsification), and category-soundness (energy landscape) "for free." Cons: more machinery (membrane, recommender, dream orchestrator); the down-projection is lossy for relations (mitigated: reason in RDF/CL, treat LPG relations as a shadow); dream auditability + re-projection shear must be engineered (OQ4).

Option 2 — Collapse the layers¶

Pros: one store, no membrane. Cons: storage-shaped truth; no reasoning; the exact anti-pattern the north-star forbids.

Option 3 — Read-only projection, no extensions¶

Pros: simplest, no admission risk. Cons: operations cannot tailor; no telemetry; defeats the purpose.

Option 4 — Auto-promoting frequency ratchet¶

Pros: fully automatic. Cons: popularity deciding category mints anti-patterns at scale; thrash; corrupts canon — refuted by all three reviews.

Review findings that shaped this design¶

Three foundational pressure-tests converged on one verdict: the naïve single-funnel ratchet let a fast-layer signal make a slow-layer commitment. The fixes are gates, not a teardown.

UFO / OntoUML. The ladder must branch, not funnel: add the intrinsic-moment branch (a property that "gains its own attributes" is a «mode», not an edge); make the terminus a sortality/rigidity sorter (kind is rare and identity-bearing — never the default), not "a class." "Popularity should nominate; classification must decide." Replace gate "FK-smell → edge" with a material-vs-formal ontological gate. Upgrade claim C2: share the whole category classifier across projection-time and promotion-time, not just the reify rule.
BFO 2020 / OBO. Add the Category Triage gate first: WD (world-descriptive → domain-eligible) vs ICE (information about the record → IAO annotation only) vs ADMIN (directives → never a term) vs OCCURRENT/TEMPORAL (→ a parallel occurrent ladder; do not fossilize processes into relator vertices). Split rung-3 "relator" into relational-quality / interdependent-roles+process / process+ICE; force a time-index decision (BFO relations are ternary). Enforce OBO authoring discipline at mint (genus+differentia, single is_a to BFO via reused CCO/IAO, opaque ID, reuse RO). Keep disjointness axioms ON so miscategorization is unsatisfiable.
KR&R / knowledge-engineering. "Lossless via preserved IRIs" is too strong — define a minimal invariant set (datatype, lang, multi-valued, named-graph, sameAs sidecars) and run owl:sameAs closure before projection. Mining needs guards (FK-smell false positives; ops-bookkeeping exclusion list; support/lift; type-entropy; rare-but-genuine override). Promotion needs hysteresis (time-series, sustained window) + a deprecate-not-delete path. Usage evidence must be tamper-evident (hash-chained, time-series, not a snapshot). Re-projection needs change-log/epoch drift detection; GC needs tombstone + resurrection to avoid silent loss of still-referenced extensions.

Claims scorecard¶

Claim	Verdict
C1 property → edge → relator → class	Refuted as a single funnel → replaced by triage + WD-fork + sortality sorter (multi-destination)
C2 same rule at projection & promotion	Upgraded → one shared category classifier, invoked both directions across the membrane
C3 ops propose / formal disposes	True with teeth → disposal must judge category (not just consistency) and run with disjointness ON, else it corrupts by miscategorization
C4 lossless via preserved IRIs	Split: transport round-trip is true given the invariant set; referential identity needs separate dedup + duplicate-universal gates; down-projection is asymmetrically lossy (time-index)
C5 reason-RDF-open / validate-membrane-closed	Defended (the strongest idea) + scope OWL-vs-Common-Logic (temporal/ternary claims reason in CL; SHACL on the RDFS-materialized graph, not assuming OWL entailments)

Open Questions¶

What wakes the dreamer — clock (cadence) vs pressure (sleep when accumulated drift / free-energy crosses a threshold) vs attention (sleep only when budget is spent). This also sets the loop's time constants: drift-as-alarm wants low deadband/high recall; value-as-prioritizer wants high deadband/slow. The hub owner's lean (2026-05-30): just note drift + emergence; harvest for attention — i.e. observation-first, optional consolidation, not an aggressive controller.
ICE handling — a separate governed information-artifact ontology (IAO/CCO-aligned) for the telemetry/record-metadata, vs lightweight annotations. Keeps the domain ontology clean (orthogonality) either way.
Substrate — Postgres + Apache AGE for the fast layer (JSONB rigor-free metadata · SQL value/drift measurements · pgvector sense-drift/dedup · RLS/triggers as the admission membrane) vs a native LPG (Neo4j+GDS) vs a both-models engine (Neptune/ArcadeDB). The slow layer needs an RDF/OWL/SHACL + reasoner store regardless.
Dream auditability + shear — crystallization must be gated (only consistent, category-sound configurations) and recorded (PROV consolidation events with a merge/prune map); re-projection after a dream must let the fast CANON zone rebind gracefully (bitemporal, no whiplash).
Energy function — the MDL/parsimony weighting; whether category-soundness enters as a hard constraint (infeasible region) or a steep soft penalty; competency-question coverage as the explanatory-power term.

ARC-ADR-016: relator-vertex + role-binding — the CANON down-projection of n-ary relations, and the reify gate reused at promotion time.
ARC-ADR-019: the slow-plane store + reasoner (the disposer; where dreams crystallize).
ARC-ADR-030: the gate-first ingestion pipeline — operations becomes another upstream proposer behind the same gate.
ARC-ADR-032: propose/dispose, the holographic LPG staging graph, quarantine, PROV-O lineage — reused for CANDIDACY and crystallization.
ARC-ADR-029: forge — emits the deterministic CANON projection DDL.
ARC-ADR-026: raw-vs-business split — the FRONTIER/CANON analogue and same-as/identity-resolution patterns.
ARC-ADR-001: HITL — adjudicates the stereotype at crystallization, not just the schema diff.
Labs: Pace-Layering & the Dreaming Ontology, Ontology-Pipeline, Reification-and-Hyperedges, Evidence as a Primitive, Governance in the Model.

Revision History¶

Version	Date	Author	Change
0.1	2026-05-30	Claude Code (assisted)	Initial Proposed — pace-layered RDF↔LPG projection with a drift-sensing operational frontier and dream-consolidated graduation; folds in the UFO / BFO / KR&R pressure-test findings; composes ADR-016/019/029/030/032.
0.2	2026-05-30	Claude Code (assisted)	Renumbered ARC-ADR-038 → ARC-ADR-041 (038 retained by Unified Process & Time; 039 = Foundations; 040 = Graph Visualization). No content change.
0.3	2026-05-30	Hub owner	Accepted. Pace-layering architecture approved. Open questions (dreamer trigger, ICE handling, substrate, dream auditability, energy function) remain documented follow-ups.