ONN + ORTSF — a field note on testable claims and proxy controllers

An edited extract from the 2026-02-03 working session. The underlying meeting notes are iterative; this is the version that survived three passes of critique.

Historical note (2026-07-10). This Feb-2026 planning note was written before the programme's higher-order thesis was audited. Its instinct holds up well — it already treats ORTSF as a conceptual/proxy layer and flags that the empirical harness is not public. What changed since: the central higher-order claim resolved to a scoped No-Go boundary, and the control side reduced to a standard delay margin. Read the hypotheses below as of their date; see the current ONN research status.

The session's question#

Can the ONN + ORTSF pipeline be reduced to a small set of testable claims, each backed by a minimal experiment, without pretending to reproduce implementation details we do not have?

The honest answer turned out to be yes, but only if ORTSF is treated as a conceptual layer and swapped for a well-chosen proxy controller for experimental purposes. The theoretical stability and topological-invariance claims of the original framework are strong; the empirical harness that validates them is not yet fully public.

What we anchored to#

Three papers formed the backbone of the review:

• ONN + ORTSF framework — arXiv:2506.19277. The primary source. Frames ONN as an ontology-driven neural representation and ORTSF as a control layer that compensates delay while preserving phase margin.
• ONN for topologically conditioned constraint satisfaction — arXiv:2601.05304. Useful specifically for understanding how ontology schema constraints shape embedding consistency.
• Constructive Lyapunov functions via topology-preserving neural networks — arXiv:2510.24730. The bridge to classical stability analysis.

Across the three, the conceptual unification is stronger than the empirical evidence base: large-scale validation under realistic (bursty, heavy-tailed) delay distributions is the gap worth closing.

Three testable hypotheses#

• H1. Topology-consistent ONN embeddings improve control stability under stochastic delay.
• H2. ORTSF-style compensation yields higher phase margin than a Smith predictor baseline in the heavy-tailed delay regime.
• H3. Ontology schema complexity has a U-shaped effect on robustness — too sparse and the topology fails to constrain; too dense and schema bias dominates.

H3 is the one most worth stress-testing; it's the claim most likely to be wrong in its strongest form.

An experiment ladder#

Three rungs, each minimal, each isolating one variable:

1. Topology-only reasoning — no delay. Measure embedding consistency across schema variants.
2. Reasoning + control at fixed delay — classical delay, fixed $\tau$. Measure phase margin and tracking error against a baseline MPNN + Smith predictor.
3. Reasoning + control at stochastic delay — the interesting regime. Start with log-normal, then move to bursty and bimodal.

Evaluation matrix (draft)#

Acceptance thresholds (draft)#

• Phase margin $\ge 25^\circ$ under bursty delay.
• Tracking error $\le$ baseline $+ 10\%$ under the worst delay distribution.
• Embedding consistency $\ge 0.85$ (normalised).
• Runtime $\le 20$ ms/frame on the target hardware.

None of these are final. They are the numbers at which we'd accept the hypotheses and move on.

Risk register#

• R1. ORTSF implementation details insufficient for reproduction. Mitigation: treat ORTSF as a conceptual layer and build a proxy controller — a predictive controller with tunable delay-compensation gain — as a stand-in. Document what is ORTSF and what is proxy.
• R2. Dataset mismatch — scene graphs exist for SLAM benchmarks, but not for delayed-control tasks. Mitigation: build a synthetic graph simulator coupled to a delayed-control plant.
• R3. Ontology-schema bias. Mitigation: include a sensitivity analysis across schema depth, branching factor, and relation cardinality.

One concrete deliverable#

An evaluation-matrix dashboard — a single table that auto-updates the four metrics above against ontology complexity and delay distribution. The artefact is small; the discipline of maintaining it is the point.

This note is lightly edited from the session's internal minutes. The underlying artefacts — action items, ideas backlog, iterated findings — remain in the research archive.