Substrate, shelter, redundancy, and risk

\(\Phi\) (substrate integrity) and \(\Psi\) (shelter coefficient) are not fixed averaging functions. They are graph-indexed composition operators: the same symbols multiply the middle term of the FPE, but the rule that maps child health or enclosure buffering into a scalar depends on topology — whether components are in series (all required), parallel (substitutable), or mixed.

This essay is the canonical clarification for that distinction. Formal case studies: to_publish/05_substrate_and_shelter.md. Derivation: information_persisting_systems.md §2.5, §4.4.

1. What \(\Phi\) and \(\Psi\) are

Both are dimensionless factors in \([0,1]\) that gate \(\mathcal{R}^{(L)}\):

By convention smaller \(\Psi\) is better shelter (less noise transmitted). Larger \(\Phi\) is better substrate (healthier composition graph).

They are independent channels of collapse: substrate failure and shelter failure have different signatures (05_substrate_and_shelter.md §3). Conflating them — treating alliance membership like organ health, or averaging country \(\mathcal{R}\) into humanity’s \(\mathcal{R}\) — produces specific wrong predictions.

2. Composition laws at the extremes

The FPE fully specifies behaviour at limit topologies. Mixed real systems sit between these limits; identifying which limit applies to which edge is the empirical work.

2.1 Substrate \(\Phi\) — series vs parallel

Let \(\{\mathcal{R}_i\}\) be persistence ratios of level-\((L-1)\) sub-IPS, and \(\mathcal{C}\) the set of critical constituents (see §4).

Theorem 5.3 (IPS paper): failure of any critical child breaks Markov-blanket factorisation at level \(L\) → \(\Phi \to 0\) regardless of shelter.

Non-critical constituents may enter via weighted summaries far from collapse; near collapse the minimum over \(\mathcal{C}\) dominates.

2.2 Shelter \(\Psi\) — independent channels vs overlap

Let \(\Psi_j\) be the fraction of environmental noise transmitted through enclosure \(j\) on a given channel.

Shelter compounds across distinct channels (product). Substrate gates on critical paths (minimum). That asymmetry is the central compositional claim.

3. Redundancy as a design choice

An IPS (person, firm, state, trained model) does not get a fixed \(\Phi\) or \(\Psi\) from physics alone. It chooses substrate topology and invests in shelter edges subject to cost:

Risk in this framework is not a separate mystical variable. It is sensitivity of \(\mathcal{R}^{(L)}\) to which graph edge fails first:

• Series substrate → tail risk (step-function collapse).
• Parallel substrate → smoother loss, higher baseline cost.
• Thin shelter on a noisy environment → \(\mathcal{E}_\Sigma\) rises continuously at the node.

There is no free lunch: redundancy is insurance on \(\Phi\) or \(\Psi\), priced in \(\omega\), \(\Gamma\), and foregone efficiency.

4. Critical constituents

A sub-IPS \(i\) is critical (\(i \in \mathcal{C}\)) when its failure breaks

— the conditional independence that defines the Markov blanket (IPS paper eq. 2.1). Then \(i\) enters \(\Phi\) through the minimum, not the mean.

Operational identification (open problem): domain knowledge (cardiology, grid engineering, ecology), bow-tie / keystone structure [Csete & Doyle, 2004], percolation on the substrate graph, or — speculatively — transfer-entropy thresholds. The framework is explicit that which nodes are critical is not deduced from the FPE alone; it is read off the graph.

A nominal part (empty department, symbolic UN vote, idle expert in MoE) contributes nothing to \(\Phi\) or \(\Psi\) unless it does thermodynamic or informational work on the blanket.

5. Worked mappings

5.1 Person

• \(\Phi\): critical organs in series → \(\Phi_{\mathrm{person}} \approx \min_i \mathcal{R}_i\) over \(\mathcal{C}\) (circulatory, respiratory, nervous, …). Microbiome: partly redundant.
• \(\Psi\): body, household, employer, healthcare — product over channels. Losing one job does not kill; losing one heart does.

5.2 Humanity and the UN

• Country \(\mathcal{R}_i\) does not aggregate to humanity \(\mathcal{R}\) by arithmetic mean.
• Countries (plus biosphere, infrastructure) are substrate at the humanity scale → \(\Phi\) is graph-shaped; critical nodes (food system, climate buffer, nuclear stability) dominate.
• UN is primarily \(\Psi\) (planetary coordination shelter), not the substrate aggregation rule. See ips_geopolitics.md. Humanity’s \(\mathcal{R}\) is not the mean persistence of UN members.

5.3 Globalization (late 20th century)

A deliberate substrate topology experiment:

Nearshoring and friend-shoring are attempts to re-purchase \(\Phi\)-redundancy after discovering which edges were in \(\mathcal{C}\) all along. See ips_economics.md.

5.4 Fractal MoE (software)

Designed IPS (IPS paper §6.8; nanochat theory):

Each FractalMoE node learns phi_criticality_logit (unless fixed via moe_phi_criticality_learnable=False). High \(\kappa\) raises dissolution pressure when any active child has low \(\mathcal{R}\); low \(\kappa\) plus moe_redundancy_coef encourages redundant expert signals. See implementation §5.3.

6. Log-space decomposition

When diagnosing mixed failure, use:

Each channel’s contribution to loss is readable from the appropriate log-derivative. A reef can lose \(\log \Phi\) (bleaching) and \(\log \Psi\) (warming ocean) simultaneously.

7. Open problems

1. Closed form for mixed substrate graphs — reliability theory on labelled graphs; cut-set enumeration.
2. Criticality criterion — general rule for \(\mathcal{C}\) without domain experts.
3. Correlated shelter channels — product overestimates buffering when \(\Psi_j\) share noise sources.
4. Learned aggregates in ML — when to train min-like vs mean-like \(\Phi\) for faithful collapse dynamics.

8. Related documents