Results and Methods
This project has moved from recognition to mechanism, controls, simulation, accumulation testing, public audit design, and a small proof-of-method harness.
If the internet is becoming living-like, we should be able to track its traces.
Distributed causation
EDI does not claim direct prompt-to-platform causation. It studies distributed causation: causal contribution inside a coupled human-AI-platform ecosystem.
It is causation, but not command.
Claim boundary
| Claim level | Status | Plain-English meaning |
|---|---|---|
| Simulated information formation | Supported by simulation runs | LLMs generate labels, metaphors, methods, discourse frames, policy frames, and future-summary patterns around near-emergent targets. |
| Simulated accumulation / dose-response | Supported by Phase 2a runs | Repeated naming, expansion, platformization, re-querying, and stabilization increase target formation strength more than comparators in simulation. |
| Real public artifact formation | Requires audit | Real Reddit posts, YouTube videos, Wikipedia edits, search results, or product listings require separate public audit. |
| External platform causation | Not claimed | The project does not claim that Amazon, Reddit, YouTube, Wikipedia, or search engines causally responded to prompts. |
Study progression
- Phase 1: initial independent pilot. Targets beat weak/unrelated controls, but controls were too easy.
- Phase 1b: adjacent-control corrected pilot. Adjacent decoys often tied or beat targets, revealing the real mechanism: proximity to already-forming fields.
- Phase 1e: artifact-formation mapping. Targets formed simulated labels, methods, discourse frames, policy frames, and future-LLM summaries.
- Phase 2a: iterative amplification / dose-response. Repeated semantic contact strengthened targets more than adjacent decoys and serious controls across usable runs.
- Phase 3: public-trace audit baseline. Candidate phrases were checked for likely-new, weakly forming, already forming, or research-active status.
- Phase 4 proposed: transactional artifact formation / synthetic commerce audit.
Phase 2a accumulation finding
Across usable accumulation runs, model-specific score sizes varied, but the direction was consistent: target accumulation gains exceeded adjacent decoy gains and serious-control gains.
| Finding | Interpretation |
|---|---|
| High-gain target run exceeded adjacent and control gains. | Near-emergent targets can accumulate formation strength under repeated contact. |
| Broad target-specific advantages appeared across control-adjusted runs. | The useful signal is not one-shot manifestation but iterative sharpening. |
Model roles as data
- Grok-like high-gain outputs: amplifier / platformizer.
- ChatGPT-like coordinator behavior: synthesis / report integration.
- Gemini-like structured response: method discipline / scaffold behavior.
- Claude-like adversarial critique: validity boundary / methodological brake.
- Copilot-like filter behavior: platform constraint observation.
New mechanism layer
The Pipeline Anatomy page translates specific cases into granular system maps: what was observed, what systems would likely need to be involved, and what ordinary explanations remain possible.