Spaxiom: Global Experience Substrate for the Era of Experience

About This Section

This section positions Spaxiom as the foundation for a new kind of AI training corpus: a continuously growing, highest-resolution dataset of how people interact with buildings, devices, and robots. Unlike static web-scale corpora, this experience substrate is semantically typed, spatially grounded, and temporally explicit.

We explore how offline training, online retrieval (experience-RAG), and governance mechanisms transform sensor telemetry into a structured, model-ready corpus for the Era of Experience.

12. Spaxiom as a Global Experience Substrate for the Era of Experience

The original "Era of Experience" framing envisions agents that learn predominantly from trajectories of interaction with the world, rather than from static internet corpora. In reinforcement learning notation, an agent's experience can be written as a sequence of tuples

τ = ((s₀, a₀, r₀, o₀), (s₁, a₁, r₁, o₁), …)

where s_t is state, a_t an action, r_t a reward, and o_t an observation at time t.

Spaxiom extends this classical view by inserting a layer of structured events between raw observations and the agent. At each time t, the underlying sensors produce raw signals x_t, which Spaxiom transforms into an event set E_t:

x_t → E_t = {e_t⁽¹⁾, e_t⁽²⁾, …, e_t^(k_t)}

where each e_t⁽ⁱ⁾ is an INTENT-level object such as GaitInstability, CrowdFormation, UnderutilizedExit, or NeedsService.

12.1 Highest-resolution, regularly updated corpus of experience

Consider D deployed Spaxiom sites (hospitals, warehouses, offices, etc.), each generating event streams over time. For site d, let:

ℰ_d = {e_d,1, e_d,2, …, e_{d,n_d}}

denote the set of events emitted over some interval. The global corpus of experience events is then:

ℰ_global = ⋃_d=1^D ℰ_d

Because Spaxiom operates on high-resolution sensor grids (e.g., 4" floor pixels) and other dense modalities, and because it runs continuously at the edge, ℰ_global can approximate a highest-resolution, regularly updated corpus of how people interact with buildings, devices, and robots. Importantly, this corpus is not just unstructured telemetry; it is:

Semantically typed

Every event has a type and schema

Spatially grounded

Zones, coordinates, topologies

Temporally explicit

Start/end times, durations, windows

Licensable and auditable

Per-event metadata on source, policy, and privacy

Let |ℰ_global(t)| denote the number of events accumulated up to time t. Assuming each site produces events at average rate λ_d events/second, then:

𝔼[|ℰ_global(t)|] ≈ t · Σ_d=1^D λ_d

For many sites and long timescales, this becomes a continuously growing experience fabric whose size and diversity can rival or exceed static web-scale corpora, but now grounded in physical interaction.

12.2 Offline training on structured experience

This corpus can be used to train a variety of models:

World models that predict future INTENT events conditioned on current context and hypothetical actions.
Policy networks for control tasks (e.g., crowd management, energy control, fall prevention), trained on logged experience.
Representation learners that map event sequences to embeddings encoding situation, risk, or intent.

Because events encode semantically meaningful structure, many models can operate in a reduced state space. Suppose a raw sensor state x_t lives in ℝⁿ, but the INTENT state z_t (e.g., counts, scores, zone-level features) lives in ℝ^m with m ≪ n. A world model f_θ can be trained in the lower-dimensional space:

z_t+1 ≈ f_θ(z_t, a_t)

reducing sample complexity and compute requirements relative to predicting in raw sensor space.

12.3 Online serving: experience-RAG and safety priors

Online, agents can treat ℰ_global as a retrieval-augmented memory:

Given a current event cluster (e.g., emerging crowding near an exit), retrieve similar past episodes and their outcomes.
Use retrieved episodes to condition an LLM's plan (e.g., how to redirect people, adjust signage, dispatch staff).
Enforce safety priors by checking whether the current pattern matches known dangerous motifs (e.g., near-misses).

Formally, let q be a query embedding derived from the current situation, and let {z̃_i} be embeddings of past episodes. A retrieval step returns:

𝒩(q, K) = {top-K episodes by similarity to q}

and the agent conditions its policy π(a | q, 𝒩(q,K)) on those episodes.

12.4 Governance, attribution, and controllability

Finally, having a language-level representation of experience simplifies governance:

Attribution: each event knows its origin site, zone, and policy, enabling fine-grained revenue sharing and audit trails.
Policy filtering: certain event types or fields can be excluded from training or serving (e.g., per-region privacy rules), by manipulating schemas rather than raw logs.
Red-teaming & safety: dangerous patterns can be expressed as event-level formulas (e.g., combinations of DoorBlocked, CrowdFormation, and low-exit-usage) and used to filter or reweight training data.

In this sense, Spaxiom is not just another middleware layer; it is a deliberately designed substrate for the Era of Experience: a way to convert messy, heterogeneous sensor streams into a structured, governable, and model-ready corpus of physical-world experience.