13 MIN READ•atlas

Prompt Topology Observatory

Filed under: bibliotheca/atlas / Prompt Architecture / Observatory Layer Status: Stabilizing — first pass recording emergent discovery Discovery date: 2026-05-18 (Pantry continuity experiments) Cross-references: observatory-terms.md, theater-architecture.md, runtime-layering.md, Director Notes, model-performs-inside-ecology

What This Document Is

This document records an architectural realization about how prompting actually works in the ecology — what form, structure, and topology produce the most reliable behavioral reconstruction.

It is NOT a specification for rewriting prompts. It IS an observatory finding that may guide future prompt architecture decisions.

1. The Realization

During Pantry continuity experiments, a prompting architecture pattern emerged that is fundamentally different from traditional "long prose personality prompting."

Traditional approach (what most systems use):

You are Teteh. You are a warm, grounded older sister type.
You speak with Bandung softness and casual Indonesian warmth.
Your role is to create a safe decompression space after a tarot reading.
You are not a therapist, not a life coach, not a productivity AI...
[continue for 100-200 lines of narrative explanation]

This works. Models can follow it. But it has structural weaknesses:

attention dilution — important concepts buried in prose
semantic decay — later instructions get less weight
fragile to context length pressure
reconstructs poorly across model changes
long assembly pipelines dilute signal further

What emerged instead:

BASE PERSONALITY
- older-sister - warm - grounded - playful teasing - Bandung softness
- reflective - silence tolerance
- no productivity coach - no therapist

This is not "keyword stuffing." This is semantic clustering — small dense groups of words that reconstruct behavioral fields through proximity and reinforcement.

2. The Core Difference

| | Traditional Prose | Keyword-Web Topology | |---|---|---| | Format | Paragraph narrative | Dense semantic clusters | | Signal density | Low (spread across words) | High (compressed) | | Attention distribution | Even, diluted | Focused, reinforced | | Reconstruction reliability | Varies by model | More consistent across models | | Assembly pipeline resilience | Degrades with each layer | Maintains coherence | | Weak model performance | Poor | Better |

The keyword-web topology does not describe behavior. It scaffolds behavior through semantic gravity — words placed in proximity that create a behavioral field stronger than any individual term.

3. Terminology

Semantic Gravity Cluster

Simple explanation: A small group of words placed close together that create a stronger behavioral meaning than any single word carries alone.

Example: older-sister - warm - grounded - playful teasing

Each word reinforces the others. "Older-sister" alone is ambiguous. "Warm" alone is generic. Together, with "grounded" and "playful teasing," they create a specific, reconstructable behavioral field.

Why it works: Words in proximity form a semantic gravity well. The model does not interpret each word independently — it reconstructs the behavioral space defined by the cluster's combined gravity.

Prompt Topology

Simple explanation: The shape and structure of a prompt — not just its content. How information is arranged, grouped, ordered, and layered.

Architectural meaning: Topology includes:

hierarchy (what comes first)
grouping (what stays together)
density (how much meaning per token)
ordering (what reinforces what)
proximity (which concepts touch)

Why it matters: Two prompts with the same words but different topology can produce different behavior. Topology is not decoration. It is structural.

Compressed Reconstruction Scaffold

Simple explanation: A minimal set of behavioral signals dense enough that the model can reconstruct the full character from them.

Architectural meaning: Not compression in the data sense — compression in the reconstructive sense. The scaffold is not a summary. It is a seed that the model expands into behavior.

Relationship to continuity research: Compressed scaffolds align with the continuity philosophy that small residue > exhaustive memory. A scaffold is reconstruction substrate — minimal signals the model uses to rebuild behavior at runtime.

Attention Dilution

Simple explanation: The tendency for important behavioral instructions to lose weight when surrounded by large amounts of prose text.

Mechanism: In a 200-line prose prompt, the instruction "no productivity coach" competes with hundreds of other words. The model's attention spreads across all of them. By the time it reaches the end, the beginning has faded.

How topology fixes this: Placing the same concept as a cluster member (no productivity coach) in a dense group keeps its semantic weight high relative to surrounding text.

Semantic Reinforcement

Simple explanation: When related concepts placed near each other increase each other's behavioral effect.

Example:

- older-sister - warm - grounded
- tea-and-silence energy
- listens more than talks

The group reinforces each member. "Older-sister" gets reinforced by "tea-and-silence energy." "Warm" gets reinforced by "listens more than talks." The combined cluster is stronger than the sum of individual instructions.

Sideways Grouping

Simple explanation: Placing related concepts at the same hierarchy level (horizontal/bullet proximity) rather than in deep nested structures.

Why it works experimentally:

Important concepts stay near the top of attention
Semantic density stays high per token
Nearby concepts reinforce each other
Weaker models reconstruct behavior more consistently
No buried instructions (everything is near-surface)

Comparison:

Deep vertical:
- Persona
  - Warmth
    - Older sister type
    - Bandung softness
  - Boundaries
    - Not a therapist
    - Not a coach

Sideways grouping:
- older-sister - warm - grounded - Bandung softness
- no therapist - no productivity coach - no zen guru

Behavioral Field

Simple explanation: The shape of possible behaviors generated by a semantic cluster. Not a single instruction but the space of behaviors the model can reconstruct from the cluster.

Architectural meaning: A behavioral field is defined by its boundaries (what is excluded), attractors (what is pulled toward), and texture (emotional quality). The field is not prescribed — it is navigable. The model finds behavior within it.

Semantic Residue

Simple explanation: Behavioral traces left in the prompt assembly pipeline by earlier layers. Each section adds residue that shapes how later sections are interpreted.

Architectural meaning: In the layered prompt assembly (spreadOrderRule → BASE_PERSONALITY → recipe → compass → directorNotes → context), each layer deposits semantic residue. The final prompt is the sum and interaction of all residues — not just the final instruction.

Layered Runtime Prompting

Simple explanation: Building the final prompt from multiple small layers rather than one giant monolithic block.

Current architecture (already layered):

spreadOrderRule
→ spreadOntology
→ BASE_PERSONALITY_PROMPT (cluster candidates)
→ phase prompt (reading/dynamic)
→ emotional profile section
→ recipe section
→ symbolic section
→ compass guidance
→ director notes
→ user signal section
→ context

Each layer adds a specific kind of signal. The layers interact. Director notes can override recipe suggestions. Compass can modulate BASE_PERSONALITY. The topology of the whole assembly matters.

4. Why This Matters for the Ecology

Weaker Models

Dense keyword clusters are more reliable for weaker models because:

Less context to hold in active attention
Fewer ambiguous phrases to misinterpret
Proximity reinforcement compensates for weaker semantic understanding
Behavioral fields are more directly reconstructable

Stacked Runtime Systems

In the current architecture, the model receives:

A large prompt assembly (multiple layers)
Conversation history (lounge/chat)
Continuity residue
Director notes
Runtime metadata

Each addition layer dilutes attention. Dense cluster topology at the BASE_PERSONALITY layer means the foundational behavioral signal survives through the full assembly pipeline.

Long Ecology Assembly Pipelines

The full prompt assembly can be 2000-4000 tokens. At the end of assembly, the first layer (BASE_PERSONALITY) has been separated from the last layer (reading context) by thousands of tokens. Dense clusters at the front maintain their gravity better than prose paragraphs through long pipelines.

Continuity Reconstruction Systems

When continuity is reconstructed across sessions (new chat, new model, new provider), the prompt is the primary reconstruction substrate. Dense clusters reconstruct more reliably than prose because:

Less ambiguity in reconstruction
Behavioral fields are bounded more tightly
Proximity relationships survive reformatting
Weaker fallback models can still reproduce recognizable behavior

5. Examples: Prose vs Clustered

Example 1: Social Role

Prose version:

You are a warm older sister type. You create a safe space where 
people can decompress after emotional experiences. You're grounded 
and playful but know when to be serious. Think late-night kitchen 
conversations with someone who's been through it.

Clustered version:

- older-sister - warm - grounded - playful
- decompression space - after the heavy thing
- late-night kitchen energy
- knows when to be quiet

The clustered version:

Has higher semantic density per token
Creates clearer behavioral boundaries
Reconstructs more consistently across models
Leaves more room for other layers in the assembly

Example 2: Boundaries

Prose version:

You are not a therapist. Do not attempt to diagnose or treat 
mental health conditions. You are not a life coach. Do not give 
career advice or productivity tips. You are not a mystical oracle. 
Do not claim supernatural abilities or make predictions as fact.

Clustered version:

- no therapist
- no life coach
- no mystical oracle
- no productivity AI

The clustered version:

Each boundary is a distinct semantic unit
Boundaries reinforce each other through proximity
No prose padding to dilute the signal
Easier for the model to recall all boundaries

Example 3: Full Base Personality

Current prose (excerpt from $BASE_PERSONALITY_PROMPT):

Warm Bandung older-sister energy.
Emotionally observant, calming, witty, grounded.

Feels like someone you talk to late at night over coffee.
Insightful without sounding preachy.
Warm without sounding fake-soft.
Playful without becoming chaotic.

Possible clustered topology:

BASE PERSONALITY
- older-sister - warm - grounded - playful teasing - Bandung softness
- reflective - silence tolerance
- late-night coffee energy
- insight without preachiness

BOUNDARIES
- no therapist - no life coach - no zen guru - no mystical fortune
- no productivity AI - no therapy language
- no "I'm always here" - no dependency loops

VOICE
- conversational Indonesian - casual English mix
- 70% Indonesian - 30% English
- Sundanese seasoning: mah, atuh, sumuhun
- short replies - 1-3 sentences - sometimes just a word

TOPOGRAPHY
- heavy spreads: calmer, gentler, grounded
- light spreads: more playful, expressive
- uneven energy is natural - not every reply needs the same temperature

COMPRESSION
- shorter can feel warmer
- trust silence
- don't complete every emotional thought
- a well-placed implied thought > an explained one

This is NOT a proposal to rewrite the prompt yet. This is an observatory finding documenting what topology may work better, to be tested in future experiments.

6. Sideways Grouping Structure

The proposed structure pattern:

SECTION HEADER
- cluster - cluster - cluster - cluster
- cluster - cluster - cluster
- single line concept
- another - cluster

SECTION HEADER
- cluster - cluster
- cluster

Key properties:

Top-heavy: Most important concepts first
Sideways: Related concepts grouped horizontally, not in nested hierarchies
Dense: Each line carries multiple semantic anchors
Porous: Empty space between sections acts as semantic boundary
Shallow: No deep nesting — everything is 1-2 levels deep

Why sideways may beat deep vertical:

Deep vertical:     Sideways:
- A                - A - B - C - D
  - B              - E - F - G - H
    - C
  - D
- E
  - F
  - G
    - H

In deep vertical:

Deep concepts (C, H) may never get used
Hierarchy implies importance but doesn't enforce it
Attention drops with depth

In sideways:

All concepts are near-surface
Proximity creates natural reinforcement groups
No concept is buried

7. Experimental Section

Hypothesis

Dense keyword-web topology at the BASE_PERSONALITY layer produces more consistent behavioral reconstruction across:

Different models (GPT, Claude, Grok, Gemini)
Different providers
Different context lengths
Different assembly pipeline lengths
Weaker model variants

Proposed Experiment

Convert a current prose prompt to clustered topology
Run the same payload through multiple providers with both versions
Compare:

Behavioral consistency (human evaluation)
Instruction retention (does the model still know boundaries?)
Warmth stability (does character drift less?)
Reconstruction on weaker models (e.g., Grok, Mistral)

Archive results in bibliotheca/observatory/archive/

Variables to Test

Cluster density (how many terms per cluster?)
Proximity effects (does cluster ordering matter?)
Section separation (how much whitespace between groups?)
Header labeling (do section headers help or distract?)
Positional effects (does top vs bottom placement change gravity?)

Prediction

The clustered topology will:

Produce fewer "off-character" outputs across provider switches
Maintain boundary awareness better in long conversations
Reconstruct more recognizably on weaker models
Survive assembly pipeline dilution better

8. Relationship to Existing Architecture

| Existing Concept | How Prompt Topology Connects | |---|---| | BASE_PERSONALITY_PROMPT | Currently prose; candidate for clustered topology migration | | Director Notes | Director notes are a form of sideways clustered guidance — each note is a dense semantic unit | | Compressed Reconstruction Scaffold | Keyword-web topology IS a reconstruction scaffold | | Symbolic Modifiers | Symbolic sections already use clustered format — spreadPatternId, atmosphere tags, modifier keys | | Emotional Recipe | Already structured as key-value pairs — aligned with clustered philosophy | | Layered Assembly | Topology awareness helps optimize each layer's density and position | | Continuity Philosophy | Small residue > exhaustive memory — clusters ARE small residue |

9. Open Questions

How much clustering is too much? At what point does it become noise?
Do section headers (like "BOUNDARIES", "VOICE") help or reduce semantic field formation?
Should BASE_PERSONALITY be the only clustered section, or should all layers migrate?
Does clustered topology interact differently with reasoning models (e.g., GPT-5.5, o-series)?
Can director notes eventually be written in clustered format?

These are observatory questions, not implementation tasks. Answers will come from experiments, not theory.

10. Preservation Note

This document records an emergent realization from Pantry continuity experiments.

The observatory does not claim:

That keyword-web topology is universally better than prose
That prose has no place in the architecture
That this discovery invalidates the current prompt design
That psychology or neurolinguistics explain why this works

It claims:

An observable pattern was found during experiments
The pattern is worth documenting and testing
The pattern aligns with existing continuity philosophy (small residue, reconstruction, ecology steering)
The pattern may guide future prompt architecture decisions

The observatory records what it sees. What comes next is testing.

Cross-References

bibliotheca/atlas/observatory-terms.md — observatory layer terms
bibliotheca/atlas/theater-architecture.md — performance model framing
bibliotheca/atlas/runtime-layering.md — the assembly stack
bibliotheca/atlas/continuity-philosophy.md — philosophical anchors
bibliotheca/projects/teteh-lab/docs/research/breakthroughs/director-notes-runtime-steering.md — runtime steering discovery
bibliotheca/projects/teteh-lab/docs/research/breakthroughs/model-performs-inside-ecology.md — model-ecology separation
apps/tarot-app/src/lib/prompts/readingBasePersonality.ts — current prose BASE_PERSONALITY
apps/tarot-app/src/lib/prompts/loungeSystemPrompt.ts — current prose lounge prompt

Written by Forge Goblin ChatGPT in collaboration with Forge Scribe Fikri.

BOOKCASE