Skip to content

Ploidy

Same model, different context depths, better decisions.

Ploidy is a structured debate protocol between physically separate sessions of the same LLM with intentionally asymmetric context depths. Unlike multi-model approaches that rely on model diversity, Ploidy exploits context diversity within a single model.

The name draws from biological polyploidy, where gene duplication provides redundancy that enables both error tolerance and functional diversification.

The Problem: Stochastic Prior Lock-in

When you ask the same model the same question in different sessions, you get different answers. If you only use one session, that first stochastic response anchors all subsequent reasoning. Prompt-based mitigations (chain-of-thought, reflection, "think again") have no statistically significant effect on this anchoring bias.

A user confined to a single session is unknowingly subject to a stochastic lottery.

The Mechanism: Context Asymmetry Spectrum

Ploidy introduces a 2D framework for multi-session verification:

Passive (in prompt) Active (on request) None
Full context Deep Deep-Active --
Compressed Semi-Fresh-Passive Semi-Fresh-Active --
None -- -- Fresh

Two terminals connect to one Ploidy MCP server. Terminal 1 (Deep) has full project context. Terminal 2 (Fresh) starts clean. They debate through typed semantic actions (agree, challenge, propose_alternative, synthesize) before a convergence phase.

Key Finding: Primacy Anchoring Effect

In factorial ablation experiments across 10 tasks and 11 methods, we identified information position as the dominant factor in context delivery:

Condition Summary Position Instruction Avg Recall
SF-Passive Top No 89%
SF-Passive+Independent Top Yes 94%
SF-Passive+Bottom Bottom No 100%
SF-Active Bottom Yes 100%

Moving a compressed summary from the top to the bottom of the prompt improves recall from 89% to 100% (+11pp) -- with no other changes. This is consistent with primacy anchoring effects in human cognition.

Pilot study caveat

These results are from 10 tasks with single runs per condition. Statistical validation with 30+ tasks and 5+ runs is in progress.

Architecture

Terminal A (Deep)              Terminal B (Fresh)
   | debate/start                | debate/join
   | debate/position             | debate/position
   | debate/challenge            | debate/challenge
   | debate/converge
        |
   Ploidy Server (FastMCP, Streamable HTTP :8765)
        |
   SQLite (WAL, ~/.ploidy/ploidy.db)

5-phase protocol: INDEPENDENT -> POSITION -> CHALLENGE -> CONVERGENCE -> COMPLETE

Paper

Ploidy: Context-Asymmetric Structured Debate for LLM Decision Verification

Target venues: ICML 2026 Workshop, NeurIPS 2026, AAMAS 2027

24 references including CCR, AceMAD, SR-DCR, and 6 cognitive science papers (generation effect, directed forgetting, incubation, testing effect, primacy anchoring).

Quick Start

pip install ploidy
python -m ploidy  # starts MCP server on :8765

Connect two MCP clients (e.g., Claude Code terminals) to http://localhost:8765/mcp, then use the debate tools:

# Terminal 1 (Deep session) — call via MCP tool
debate_start prompt="Should we migrate from PostgreSQL to TimescaleDB?"

# Terminal 2 (Fresh session) — call via MCP tool
debate_join debate_id="debate-xxxx" role="fresh"

Or use single-terminal auto mode (requires API key):

export PLOIDY_API_BASE_URL=https://api.anthropic.com
# Call debate_auto via MCP — server runs both sides automatically

Documentation

MIT License | GitHub | heznpc