Skip to content
Research Memex

Core References

AI Model Discovery Protocol

Model choice improves when you test candidates against your own research tasks.

The right model is the one that fits your materials, not the leaderboard.

The AI Model Reference Guide tells you what's available. This page is the protocol for finding what fits you: a structured experiment against your own research materials.

Getting started

Phase 1: Capability Discovery

Run a one- or two-day sampling pass.

  1. Choose one real task: theory synthesis, methods critique, extraction, or writing revision.
  2. Use the same prompt and evidence pack across current model families:
    • GPT-5.5 / GPT-5.4 mini
    • Claude Opus 4.8 / Claude Sonnet 5 / Claude Haiku 4.5
    • Gemini 3.5 Flash / Gemini 3.1 Pro preview
    • DeepSeek V4-era models
    • Kimi K2.7 Code / K2.6
    • GLM-5.2
    • Qwen3.5 / Grok 4.3 where available
  3. Record differences in style, depth, accuracy, and failure modes.
  4. Note the access path: web app, API key, Cherry Studio, OpenCode, Vox, Claude Code, or Antigravity CLI.

Phase 2: Control Surface Testing

Do not assume temperature is available or useful.

  1. Start with provider defaults.
  2. Test reasoning controls where supported:
    • OpenAI / xAI / GLM: reasoning effort.
    • Claude: effort, with provider sampling constraints.
    • Gemini 3.x: thinking_level, not temperature/top-p/top-k.
    • Kimi K2.7/K2.6: default temperature; thinking behavior depends on model.
  3. Use Sequential Thinking MCP when you want visible, revisable steps across models.
  4. Only test temperature on models whose current docs support it.

Phase 3: Task Matching

Build a small table for your own work:

TaskBest modelBackupWhy
Screening
Extraction
Theory synthesis
Methods critique
Draft revision
Code/data analysis

The "why" column matters. If you cannot explain why a model won, keep testing.

Phase 4: Personal Strategy

Decide the operating pattern you will actually use:

  • Primary model: the model you trust for high-stakes reasoning.
  • Volume model: the model cheap enough for iteration.
  • Challenge model: the model that sees different weaknesses.
  • Fallback model: the model you can use when quota, latency, or outages hit.
  • Privacy path: the local or constrained workflow for sensitive materials.

Evaluation notes

For every test, save:

  • The exact prompt.
  • The model and access path.
  • The evidence pack or files used.
  • The control settings, especially reasoning effort or thinking level.
  • The failure mode, not just the useful output.

Start with provider defaults, then test reasoning controls, and adjust sampling only where the provider supports it. Once you know which models fit your work, wire them into the workspace you actually use: Cherry Studio, OpenCode, Vox MCP, Claude Code, or Antigravity CLI.

cite this page

Lin, X. (2026). AI Model Discovery Protocol. Research Memex. https://research-memex.org/docs/implementation/core-references/ai-model-discovery-protocol

@misc{docs-implementation-core-references-ai-model-discovery-protocol-2026,
  author = {Xule Lin},
  title = {AI Model Discovery Protocol},
  year = {2026},
  howpublished = {\url{https://research-memex.org/docs/implementation/core-references/ai-model-discovery-protocol}},
  note = {ORCID: 0000-0001-7885-4194}
}

one renderingthe source remains