← ClaudeAtlas

conciliumlisted

Adversarial cross-model review for hard, load-bearing tasks — combining frontier models: the Claude session (Opus 5 or Fable 5.1 as the intended orchestrator) hands a claim, diff, or result to an OpenAI model (gpt-6-astra / gpt-5.6-sol / gpt-5.6-terra / gpt-5.5, via the codex CLI on ChatGPT-subscription auth, no API key), which probes it with falsification attempts and PROPOSES a verdict; the orchestrator checks the probe and RATIFIES. Two EXPERIMENTAL, opt-in extra-family seats can be added when asked for — Kimi (Moonshot, via the Kimi Code CLI or the local Kimi Desktop runner) and Grok (xAI, via the Cursor Agent CLI on Cursor-subscription auth); neither is part of a default round. Use whenever the user wants a second opinion from a different model, a cross-model or concilium review, adversarial verification of a research claim, benchmark number, or diff, says "have GPT/codex check this", explicitly asks to add the kimi or grok seat, wants codex set up as a reviewer, needs to switch codex models mid-session
raichominev/concilium · ★ 2 · AI & Automation · score 70
Install: claude install-skill raichominev/concilium
# Concilium — cross-model adversarial review A second, *different* model reviews your (or the user's) claims adversarially. Different model lineage means different blind spots — that's the value. The reviewer PROPOSES; the calling session RATIFIES. Never let either side's confidence substitute for evidence. Designed to be orchestrated from Claude Code — **Opus 5 and Fable 5.1 are both first-class ratification seats** (Fable 5.1 is the standard Fable seat since 2026-09-08; the chair-parity measurement on the blind outcome-prediction benchmark was made on Opus 5 and Fable 5, and Fable 5.1's weaker *reviewer* calibration below says nothing about the orchestrator role; any Claude model can drive the loop, but the ratifier should be one of the two). The GPT side (astra/sol/terra/5.5 via codex) does the independent probing and mechanical execution — and that cross-family seat is load-bearing: it measurably catches what same-family chairs jointly miss. ## Prerequisites (check once per environment) 1. `codex login status` → must say "Logged in using ChatGPT" (subscription OAuth). An API key is NOT needed. It is not forbidden either — the wrappers inherit whatever session the CLI already has, so `codex login --with-api-key` works identically and simply bills per token. What does NOT work is the reverse: a subscription can NOT be used as an API key, so don't attempt proxy/router bridges. 2. Discover available models: `codex debug models` or `~/.codex/models_cache.json`.