tetraframe

name	tetraframe
description	Run the TetraFrame tetralemmatic reasoning pipeline on a question or problem. Use when the user wants to decompose a question into four independent positions (P, not-P, both, neither) and synthesize a non-averaging result. Triggers on /tetraframe, 'run tetraframe', 'tetralemma', or 'four corners analysis'.
metadata	{"author":"Hunter Bown","version":"0.1.0","style":"instruction-only","compatibility":"Agent Skills compatible clients including Claude Code, OpenAI Codex, and OpenCode"}

TetraFrame — Tetralemmatic Reasoning Pipeline

Run TetraFrame's 6-stage DSPy pipeline to decompose a question into four independent positions and synthesize a transformed output.

What it does

Takes a seed question → extracts a predicate → generates four isolated corners (P, not-P, both, neither) → maps their contradictions and complementarities → synthesizes a non-averaging result → verifies quality.

Quick start

The TetraFrame repo is at /Volumes/VIXinSSD/tetralemma.

Run with auto-discovery (simplest)

cd /Volumes/VIXinSSD/tetralemma
venv/bin/tetraframe discover  # see available backends
venv/bin/tetraframe run "<seed>" --out runs/latest/run.json

Run with a specific config

cd /Volumes/VIXinSSD/tetralemma
venv/bin/tetraframe run "<seed>" --config configs/openai.yaml --out runs/latest/run.json

Available configs

Config	Backend
configs/openai.yaml	OpenAI API (OPENAI_API_KEY)
configs/anthropic.yaml	Anthropic API (ANTHROPIC_API_KEY)
configs/openrouter.yaml	OpenRouter API (OPENROUTER_API_KEY)
configs/claude_code_cli.yaml	Claude Code CLI
configs/codex_cli.yaml	Codex CLI
configs/opencode_cli.yaml	OpenCode CLI
configs/mimo_v2_pro.yaml	Mimo v2 Pro via OpenCode

Reading results

After a run, read the output:

import json
from pathlib import Path

run = json.loads(Path("/Volumes/VIXinSSD/tetralemma/runs/latest/run.json").read_text())

# Key sections
print(run["predicate_selection"]["primary_predicate"]["text"])
for mode in ["P", "not-P", "both", "neither"]:
    print(f"\n{mode}: {run['corners'][mode]['patched_claim']}")
print(f"\nP*: {run['transformed_frame']['transformed_predicate']}")
print(f"\nVerification: {run['verification']['aggregate_score']}")

Writing good seeds

A seed should contain a genuine tension or decision. Good seeds:

• Frame a real tradeoff: "We keep framing X as A-or-B, but the real choice may be C."
• Surface a hidden assumption: "The team assumes X, but Y suggests otherwise."
• Identify a stuck decision: "We can't decide between X and Y because Z."

Bad seeds are vague ("How should we improve?") or already resolved ("Should we use React?").

Pipeline stages

1. Seed Distill — extracts stakes, constraints, unknowns, hidden assumptions
2. Predicate Selection — picks the operational predicate for four-corner reasoning
3. Four Corners — generates P, not-P, both, neither in isolation (each self-critiqued)
4. Cartography — maps cross-corner relations, fair reconstructions, arbiter notes
5. Transform — synthesizes P* (non-averaging output from all four corners)
6. Verify — quality checks (branch independence, divergence, rigor, slop, fake novelty)

Verification scores to watch

Metric	What it catches
Branch independence	Corners leaked information to each other
Rigor of Both	"Both" is actually a compromise, not a genuine co-holding
Rigor of Neither	"Neither" is evasion, not a real frame diagnosis
Transformation quality	P* is just an average of the corners
Fake novelty risk	P* introduced unsupported terminology
Slop risk	Mushy language ("nuanced", "balanced", "consider")