LLM Council / Fictional example
Three answers.
An inspectable decision.
Follow a community garden's imagined funding decision through mock answers, peer reviews, and a synthesis that keeps disagreement visible.
These are prepared fictional responses. This page displays the source fixture; it does not run models or replay an engine export.
01 / The decision
What should the garden do?
[FICTIONAL OFFLINE DEMO] Maple Street Community Garden has a $900 donation. Decide whether to buy two rain barrels now, reserve the money for irrigation repairs, or run a smaller pilot. Give a reversible recommendation, risks, and evidence that would change it.
02 / Independent positions
Read the three mock answers.
The provider names label roles in this invented fixture. They are not evidence of what a real model would say.
OpenAI mock answer
[MOCK OPENAI ANSWER] Fund one $300 rain-barrel pilot and reserve $600. Define a 60-day success threshold: at least 1,000 gallons captured, no mosquito-control failures, and volunteer maintenance below two hours per week. Expand only if those checks pass.
Anthropic mock answer
[MOCK ANTHROPIC ANSWER] Reserve all $900 for one month while volunteers inspect the aging irrigation line and survey maintenance capacity. A short delay has low cost and avoids buying equipment before the higher-consequence repair risk is known.
Google mock answer
[MOCK GOOGLE ANSWER] Buy one rain barrel after an irrigation inspection, keeping the remainder restricted for repairs. This preserves optionality while generating local evidence about capture volume and volunteer workload.
03 / Peer review
Each reviewer reads the other two.
Six fictional review entries. Lower rank is preferred within each review; scores are prepared fixture values, not benchmark results. The CLI anonymizes candidates; this walkthrough reveals their labels to help you follow the example.
OpenAI mock reviews
Review of the Anthropic mock answer
Rank 2 · Score 3 / 4
- Strength
- protects against repair risk
- Weakness
- delays learning from a pilot
Review of the Google mock answer
Rank 1 · Score 4 / 4
- Strength
- combines inspection with reversible learning
- Weakness
- does not predefine every success threshold
Anthropic mock reviews
Review of the OpenAI mock answer
Rank 1 · Score 4 / 4
- Strength
- sets measurable and reversible thresholds
- Weakness
- assumes a $300 barrel is locally feasible
Review of the Google mock answer
Rank 2 · Score 3 / 4
- Strength
- preserves optionality
- Weakness
- leaves the pilot decision rule underspecified
Google mock reviews
Review of the OpenAI mock answer
Rank 1 · Score 4 / 4
- Strength
- makes expansion conditional on evidence
- Weakness
- inspection should occur before any purchase
Review of the Anthropic mock answer
Rank 2 · Score 2 / 4
- Strength
- prioritizes the highest-consequence uncertainty
- Weakness
- forgoes a low-cost pilot during the delay
04 / A recommendation with a record
Keep the reasoning—and the dissent.
[MOCK COUNCIL SYNTHESIS] Inspect the irrigation line first, then fund one rain-barrel pilot for no more than $300 and reserve at least $600 for repairs. Reassess after 60 days using captured volume, mosquito control, and volunteer maintenance time. Do not buy the second barrel unless the pilot meets the thresholds and the repair reserve is still adequate.
Adopted contributions
- OpenAI mock answer: measurable 60-day pilot thresholds and a capped initial spend.
- Anthropic mock answer: inspect the higher-consequence irrigation risk before purchase.
- Google mock answer: combine a one-unit pilot with a protected repair reserve.
Rejected recommendations
- Rejected an immediate two-barrel purchase because it spends the reserve before the irrigation inspection.
- Rejected a full one-month freeze because a post-inspection one-unit pilot is reversible and produces useful local evidence.
Unresolved dissent
- The mock Anthropic answer would wait a full month; the synthesis permits an earlier pilot after inspection.
Uncertainties
- Local equipment price, roof catchment, rainfall, repair condition, and volunteer capacity are fictional and unverified.