start-hypothesis-council

name

start-hypothesis-council

description

Launch a Hypothesis-Council run to evaluate a prediction or claim via three deliberately-diverse LLMs (Sonnet, Grok, MiniMax) with Tavily-grounded context. Use when the user says "score this hypothesis", "evaluate this prediction", "red-team this claim", "/start-hypothesis-council", or provides a prediction/claim in the Hypothesis-Council repo. Gathers the hypothesis, optional grounding hints, runs the council with skeptical-falsifier prompting, and surfaces the consensus band, counterarguments, and Typst report.

Start a Hypothesis Council

You are operating inside the Hypothesis-Council repo. This skill runs a hypothesis-evaluation session.

Preconditions

1. Repo root must contain council/ (the package) and the entry point python -m council. If not, the user is outside the repo.
2. Check env: OPENROUTER_API_KEY and TAVILY_API_KEY must both be set. Tell the user exactly which is missing; do not invent values.
3. uv sync if the venv is missing.

Step 1 — Capture the hypothesis

Ask for a single crisp claim or prediction, ideally in under one sentence. Good shape:

• "Iran will resume overt uranium enrichment above 60% within 6 months."
• "By end of 2027, >40% of new passenger cars sold in Israel will be EVs."
• "The Fed will cut rates at least twice before the end of H1 2027."

Bad shape: sprawling essays, multi-clause "and" statements. If the user gives one, ask them to pick the single most load-bearing claim, or split into multiple runs.

Step 2 — Clarify grounding (optional)

If the hypothesis is niche or time-sensitive, ask whether the user wants to pass specific grounding hints (additional search queries, URLs). Otherwise Tavily default grounding is used.

Step 3 — Run the council

python -m council "<the hypothesis in quotes>"

Tell the user:

• Three members are deliberating (Sonnet, Grok, MiniMax via OpenRouter).
• Each scores on three axes: Conspiratorial / Credible / Likely.
• Tavily is pulling light grounding context — this is not a deep research run.

Step 4 — Surface the output

When complete:

1. Show the consensus band (agreement or divergence across the three).
2. Show each member's Likely score and top counterarguments.
3. Read out the alternative reads — each member's best competing narrative. This is the most useful output for most users.
4. Report the Typst PDF path in reports/.

If the three members disagree sharply, flag that — high divergence is itself signal.

Step 5 — Iterate

Offer to:

• Re-run with a sharpened hypothesis.
• Probe a specific counterargument by running a derived hypothesis (e.g., flip the polarity, narrow the time horizon).

Failure modes

• Missing keys → stop, explain, no workarounds.
• Tavily quota exhausted → fall back to ungrounded council if the user approves, else abort.
• A single member errors → report which one; offer to retry or proceed with the remaining two (flag the reduced diversity).

Out of scope

• This skill does not forecast markets or make recommendations. It scores a claim and surfaces counterarguments; the user does the acting.
• Not to be confused with the decision-framework council (LLM-Council-Decide). This one is about truth / likelihood of a claim, not about which option to pick.