Preface

In the DeepSeek Harness (DSH) ecosystem, mainstream AI paradigms aim to maximize accuracy and systematically filter out low-probability outputs. The excluded content is precisely where potential new cognition may emerge. Collaborative mode changes this structure: algorithms no longer filter out low-probability directions, but present them to humans as candidates, which humans decide based on their own objective function.

What Is This

Collaborative Emergent Intelligence is a dynamic plugin for DSH that implements a human-AI collaborative cognitive method. Its structure is: algorithms handle divergence, and humans handle convergence.

  • Core formula: Collaborative Emergent Intelligence = engineered divergence (AI) + intuitive convergence (human)
  • Maintainer: hanshanxue83
  • Positioning: a cognitive-boundary probing machine; combines temporal resonance with three adjudication points and is suitable for scenarios where the direction is unclear and the search space needs to be expanded.

Core Features

The plugin includes 13 tools responsible for handling phases, states, fragments, resonance probing, adjudication, logical synthesis, and reflexive recursion.

Five-tier Resonance Probing

Algorithms provide candidate directions, and humans perform adjudication. The five tiers include:
* manual: human-led
* semi: web-aligned coarse filtering
* audit: audit-biased
* adversarial: adversarial retrieval
* auto: four-way parallel (aligned + adversarial + audit + temporal resonance), with LLM consolidation of candidates

Temporal Resonance

In the temporal dimension, it scans historical fragments and determines structural recurrence through bigram overlap, domain prefixes, and question anchors, thereby establishing cross-time connections.

Three Adjudication Points and Safety Valve

  • Three adjudication points: surge confirmation (pursue / do not pursue), surge-break convergence (direction adjudication), blind-spot development (no condensation without development).
  • Safety valve: code-level rejection. Includes no condensation without development, old archives must not override new memories (seq monotonic protection), and a breadth lower-bound gate.

Memory System

  • Archive structure: 9 double-copy archives (.bak disaster recovery), including phase, fragment, surge-break, blind spot, connection, convergence, surging, reflexivity, and archive.
  • Persistence: automatic recovery on startup; if the seven pools exceed their limits, entries are moved to the archive area (archive.json); entries are silenced only, not deleted.

UI and Engine

  • LLM engine: uses the real DeepSeek API (deepseek-chat), with automatic fallback to manual consolidation if it fails.
  • Interface: DeepSeek official style; the conversation bar uses a minimalist light bar to indicate the phase; the three adjudication points trigger pop-ups, one at a time and mutually exclusive.

Installation and Activation

Rebuild within a DSH session.

新对话中告知：启动协作模式

Typical Usage

Enter divergence, record XX as a fragment in domain YY. Probe frag-3 in the auto tier. Select candidate B. The blind spot is.... Condense.

Applicable Scenarios and Notes

Applicable Conditions

Suitable for scenarios where the direction is unclear, the search space needs to be expanded, decisions are needed between mutually exclusive directions, and premature closure must be avoided.

Cautions

  • Inapplicable scenarios: Not suitable for millisecond-level factual queries, absence of a sense of surging, or insufficient algorithmic divergence capability.
  • Operating mechanism: Algorithms do not make the final selection; humans do not perform exhaustive search.
  • Version: npm version 1.0.0 (machine version v8.3 / cei-5 pkg-19).
  • Permissions: The plugin runs with the current dsh process permissions; check the source code and license before installation.

Conclusion

Collaborative Emergent Intelligence combines engineered divergence with intuitive convergence, providing a concrete path for human-AI collaboration. It does not replace the user’s judgment; instead, it surfaces low-probability cognitive possibilities for human decision-making.