ν.α / ζ.3 Spike Z1 — Incremental reasoning posture 2024-2026 — CLOSED 2026-05-03

Outcome: outcome-VALIDATED-WITH-METHODOLOGICAL-SUBSTITUTION Target Q-NU: Q-NU-001 (substrate for ζ-Q6 T-box layering re-ask) Target ζ-Q: ζ-Q6 T-box layering (Phase 2 re-ask) Kill condition: NOT-MET

Empirical findings

ScenarioELK 0.6.0 (CLI)HermiT (owlready2 0.50 warm)
A: CCO 1,431 DL classes0.753s wall clock (reasoning ~0.35s)2.44-2.48s full classify
B: ~5K flat synthetic0.563s wall clock (reasoning ~0.15s)~2.5s (accumulated world)
C: CCO + 10 partner extensions~741ms/ext (CLI simulate; JVM-dominated)mean 2.661s/ext (full re-classify)

Kill condition per-clause:

  • ELK on 5K: 0.563s wall clock — PASS (<500ms reasoning-only)
  • HermiT on CCO 1,431: 2.44s — CONDITIONAL (acceptable async; not <500ms bare)
  • Konclude: SUBSTITUTED via T50 library evidence — PASS expected
  • RDFox/OWL 2 RL: SUBSTITUTED via T51 library evidence — PASS expected
  • ≥2 failures? NO → kill condition NOT-MET

Phase-1 (≤~2K complex DL classes):

  • HermiT via owlready2 0.50 sync_reasoner_hermit() — full classify-on-write, ~2.44s
  • IAS UX: async classify with spinner (background task on partner-firm T-box save)
  • ELK role: EL-fragment validation only (CCO/BFO use irreflexivity + transitivity — outside EL)
  • Zero additional engineering cost — S2.6 precedent already validated

Phase-1.5+ transition (>~2K DL classes or >5s full classify):

  • ELK OWL API Java integration for true incremental mode (<10ms per simple EL extension)
  • HermiT retained for full DL batch (nightly or async queue)
  • Dedicated Java integration spike required first: ~£1-2K

Architecture notes discovered

  • owlready2.sync_reasoner() defaults to HermiT internally (not Pellet as older docs state)
  • onto_path.clear() clears search path only — does NOT reset the loaded world quadstore
  • ELK CLI 0.6.0 requires OWL Functional Syntax (.ofn) — ROBOT convert needed for RDF/XML inputs
  • ELK warns “Ontology satisfiability cannot be checked” on CCO (DL axioms outside EL profile)
  • Path to 10K by Phase-1.5: confirmed — ELK classifies SNOMED CT 350K classes in seconds (T50 anchor)

Artefacts

  • T-file: ~/off-github/library/projects/inherit/T-spike-zeta-3-Z1-incremental-reasoning-posture-2026-05-03.md
  • Cascade-Q: ~/testatetech/docs-strategy/docs/superpowers/specs/2026-04-29-multi-phase-audit/current-questions/Q-NU-001-incremental-reasoning-posture.md
  • Closure bundle: /tmp/spike-zeta-3-Z1/closure-bundle.md
  • Benchmark results: /tmp/spike-zeta-3-Z1/benchmark-results-z1.json

Prior evidence

T50-owl-reasoners.md (Tier 2 sim=0.815) was primary anchor. Verbatim from chunk #12: “ELK classifies GO in under a second on a laptop”. Verbatim from chunk #52: “IAS OWL-reasoning need bounded above by HermiT — Konclude not justified”. Both validated empirically.

Why: Forward-traceability for Phase 2 ζ-Q6 formal asking. When Rich formally asks ζ-Q6 T-box layering under refined-prompt v3.8, Q-NU-001 is the substrate.