Indra's Net

From Fazang to Relational Systems

Exploring

Can missing structural properties between older conceptual models and modern systems reveal useful engineering questions?

UNN-001 / research fileUpdated 18 Oct 2026

Question

UNN-001 / starting point

Can missing structural properties between older conceptual models and modern systems reveal useful engineering questions?

Source

What is inherited

A working source is Fazang’s Essay on the Golden Lion (華嚴金師子章), a Huayan text that offers a conceptual system for dependent formation, whole and part, one and many, identity and appearance, and mutual inclusion.

Indra’s Net belongs to the wider Huayan field of recursive relational imagery. Neither Fazang nor the metaphor is being treated as a computer scientist, an algorithm, or a prediction of AI, databases, Web3, GPUs, or fractals.

Interpretation

What we are doing with it

We decompose the source into structural properties, compare those properties with modern technical systems, and look closely at what is absent as well as what appears similar.

The working method is called Gap-Driven Research. It is an emerging research habit, not a finished framework or a major new brand.

A small relational netSix connected nodes reflect one another, with a highlighted centre node showing how a local question can carry relationships with it.sourcecontextvalueriskmemoryquestionrelation

A diagram for thinking with relationships. It is a prompt for the experiment, not a map of the world.

Keep the categories apart.

Epistemic key

Source
What the inherited text says.
Interpretation
Our reading of its structure.
Analogy
A comparison, not an equivalence.
Hypothesis
A technical possibility that may fail.
Experiment
A controlled way to test it.
Evidence
Only what has been observed.
Unknown
What remains unresolved.

Gap-Driven Research

Working method

If an older conceptual model describes A + B + C + D while a modern system implements A + B, investigate C + D without assuming they are valuable or computable.

  1. Identify the source’s structural properties.
  2. Compare them with present technical systems.
  3. Mark what is present and what is absent.
  4. Ask why each absent property is absent.
  5. Classify it: undesirable, expensive, impossible, already solved, or a possible gap.
  6. Build the smallest experiment that could weaken the idea.

From a lion to a test

Source → interpretation → experiment

The research trail is interpretive rather than historical proof. Each step changes the question before it becomes an experiment.

  1. Golden LionBegin with Fazang’s account of dependent formation, identity, appearance, whole, and part.
  2. One and manySeparate an individual thing from the categories and relations through which it is understood.
  3. Ten MysteriesFollow Huayan accounts of mutual inclusion, interdependence, and layered relation.
  4. Indra’s NetUse the net as recursive relational imagery, not as a literal technical architecture.
  5. Modern analogiesCompare carefully with fractals, holography, graphs, networks, and distributed systems.
  6. Missing propertiesLook for structural properties that the modern comparison does not reproduce.
  7. Measured questionTranslate one possible gap into a falsifiable decision or computing experiment.

The sequence is our research trail. It does not claim that Huayan philosophy predicted modern computing.

Similar is not the same

Analytical decomposition

Five distinctions keep a useful analogy from quietly becoming an equivalence.

  • Self-similarity

    A pattern can resemble itself across scales. That does not mean every local part contains or reflects every other part.

  • Local / global representation

    A whole may be reconstructed or inferred from local information. Holography is a useful analogy here, not an equivalence.

  • Mutual interdependence

    Changes propagate through relationships. This is closer to questions in networks, distributed systems, graphs, and multi-agent systems.

  • Local autonomy

    Components remain distinct while influencing the behaviour of a larger whole.

  • Recursive embedding

    Local and global relationships repeat across levels or scales without requiring literal copies of the whole.

These five properties are our analytical decomposition. They are not an accepted mathematical definition of Indra’s Net or the Golden Lion.

The missing gap

Working hypothesis / not yet tested

What if the interesting part is not where an ancient metaphor resembles a modern system, but where it does not?

Modern distributed systems can show local autonomy and network relationships without giving every node a meaningful global representation or repeating local/global structure across levels.

That omission is not evidence of a missed breakthrough. It is a prompt to ask why the property is absent and whether a task-specific approximation would be useful.

Possible reasons for the gap

  • storage cost
  • communication overhead
  • synchronisation
  • system complexity
  • information-theoretic limits
  • no demonstrated benefit
  • an unexplored combination

Two research branches

UNN-001 research split

Relational Intelligence

Branch A / Experiment design

Does preserving relationships improve decisions when the same facts would otherwise be retrieved in isolation?

Keep the existing retrieval comparison, but score the qualities that matter instead of treating a longer answer as a better answer.

Test

  • Use the same sources and cross-domain decision prompts.
  • Compare isolated retrieval with context-linked retrieval.
  • Inspect where additional relations help, distract, or invite unsupported inference.

Measures

  • surfaced trade-offs
  • contradiction detection
  • dependency awareness
  • decision usefulness
  • hallucination or unsupported inference

Relational Efficiency

Branch B / Planned / not yet tested

Can useful global computational behaviour emerge from local representations with substantially less data movement?

The goal is not to place a complete global copy at every node. It is to test whether a small local representation can reconstruct, infer, retrieve, or access enough global context for a task.

Test

  • Establish where local inference time and memory are actually spent.
  • Compare full context, a strong conventional reduction method, and only then a future relational representation.
  • Measure resource change and task quality together.

Related research areas

Related does not mean originating here.

  • hierarchical representations
  • sparse compute and selective activation
  • cache reuse and data locality
  • structured retrieval
  • local / global embeddings
  • graph and shared latent representations
  • task-aware context reconstruction

Golden Mac

Branch B / fixed consumer test bed

PLANNED / NOT YET TESTED
Hardware
Apple Mac mini · Apple M4 · 32 GB+ unified memory
Workload
Local LLM inference, beginning with MLX / MLX-LM where practical

This is a fixed consumer test bed for application-level effective performance. It does not increase the machine’s physical FLOPS, and no speedup has been observed.

First baseline

  • Begin around a 4B quantized model, for example 4-bit, then consider larger models.
  • Use prompt lengths around 512, 2,048, and 8,192 tokens with controlled output lengths.
  • Record warm-up runs, repeated measured runs, medians, model details, and configuration.

Record together

  • time to first token
  • prefill tokens / second
  • decode tokens / second
  • total latency
  • peak memory and memory pressure
  • context processed
  • data movement where measurable
  • output quality or task accuracy

Task-Aware Partial Context Reconstruction

Can a smaller task-relevant representation preserve enough relationships to answer correctly without processing the complete context every time?

  1. Full context baselineGive the model the complete relevant source context.
  2. Strong conventional baselineUse an established retrieval or context-reduction method.
  3. Experimental relational representationA future method intended to preserve useful local/global relationships while processing less context. It has not been designed yet.

What success would actually mean.

Evidence ladder

“Make the Mac faster” is not a success criterion. Performance claims wait until a strong baseline is beaten and the result can be reproduced.

  1. Stage 0Reproduce the baseline

    Establish a stable, inspectable local inference baseline.

  2. Stage 1Locate the cost

    Identify where inference time and resources are actually being spent.

  3. Stage 2Reduce one resource

    Reduce context, latency, memory, data movement, or redundant computation without unacceptable quality loss.

  4. Stage 3Beat a strong baseline

    Outperform a conventional method on at least one meaningful workload.

  5. Stage 4Reproduce the gain

    Repeat the improvement across different prompts, models, or tasks.

Counterexample

The useful objection

  • Relational representations may require more computation than they save. Global awareness may increase communication overhead, and compression may destroy the information the task actually needs.
  • Retrieval overhead may exceed inference savings. Existing techniques may already capture the useful properties, or theoretical constraints may prevent the desired local/global behaviour.
  • The metaphor may be beautiful as a model and operationally expensive as a method. Producing no useful engineering insight is a valid result.

Findings

Evidence status

NOT YET TESTED

  • No empirical finding yet. The source boundary, analytical decomposition, two research branches, test bed, baselines, metrics, counter-hypotheses, and staged success criteria are now specified.

Open Questions

What remains open

  • Which structural gaps are undesirable, expensive, impossible, already solved, or genuinely open?
  • What counts as a relevant relationship, and when does preserving it improve a decision rather than create noise?
  • How little global context can a local representation carry while still supporting a specific task?
  • Can any reduction in context, latency, memory, data movement, or redundant computation survive a strong conventional baseline without unacceptable quality loss?

Build

Optional future layer

  • Build one transparent harness for Branch A with identical prompts, sources, and scoring across flat and relational retrieval.
  • Establish a Golden Mac baseline for normal Apple Silicon local inference before proposing an experimental representation.
  • Keep prompts, context lengths, outputs, measurements, quality checks, and failed runs inspectable.

A build here would remain experimental. It is not part of the production Decision Clarity tool unless separately stated.

Loose pages from the workbench

Lab notes

Method

The resemblance is not the result. The gap is the experiment.

Next test

Golden Mac Experiment 001 begins with a reproducible full-context baseline, not a speedup claim.

Question

What would it mean to preserve the connections that make a fact matter?

Counter-hypothesis

Maybe the net is beautiful as a model and operationally expensive as a method.

Boundary

No claim is being made that ancient thinkers invented the internet, AI, or modern systems theory.

Atmosphere, not evidence.

Visual plates from the SisiTheFox room

Sisi and Xiaoyu working together in a quiet workspace
A surface for the next test.
Sisi working alongside a client
Human judgment remains part of the system.
Sisi and Xiaoyu looking out at a natural landscape
The model still needs an edge.
Sisi and Xiaoyu reading together

Keep the question open.

The next page