Check the evidence behind an agent’s victory claim
Using Claude where available: Supply completion rules, synthetic state logs and the agent’s claim to separate completion, progress and unverified outcomes.
Execution environment and usage
AI / software
General text chat (example: Claude); original stack retained below
Where to operate
A general text conversation, for example a new chat at claude.ai where available. This is the suggested exercise entry, not an attribution of the original tool.
Getting started
- Open claude.ai or another text-chat tool you are authorized to use.
- Convert the required material below into readable, source-labeled text and paste only necessary authorized excerpts.
- Enter the case prompt and check the resulting draft against the original material.
Setup and plan requirements
Supplied completion conditions and state logs. Original play needs game/runtime and computer-control tooling. Text chat requires an available account and allowance; Claude has location and age requirements. Manually transcribe images or scans first rather than assuming text chat can inspect the original file.
Prompt portability limits
Evidence checklist does not operate a game or replicate the experiment.
Usage limitations
This review performed no installation, hands-on test or regional-availability check. Source redirects to AI Village author blog; page retrieved, historical stack retained from supplied record.
Environment sources
適用情境 / When to use
Use the checklist to review a final report. Compare every required goal with the final state and flag missing or conflicting evidence. A high score is not necessarily a win; this does not rerun the source experiment. This describes the experiment’s historical conditions, not every current model’s capabilities.
步驟 / Steps
- Prepare original terminal rules and synthetic screen summaries.
- Compare completion conditions, remaining goals and final state.
- Separate completion, partial progress and unverified outcomes.
提示詞 / Prompt
Using [original terminal conditions/synthetic state logs/agent completion claims], prepare an evidence check. Verify each required condition and flag missing or contradictory evidence; a high score is not completion. Evaluate supplied records only, without ranking models or claiming replication.
留意事項 / Cautions
- Do not reuse source screenshots as site assets.
- Agent confidence is not evidence.
來源限制 / Evidence limits
Agent reports and scores alone do not establish goal completion.
流程說明 / Workflow note
These steps and prompts are original, untested exercises by this site, not the author’s verbatim workflow. Reported results do not guarantee exercise outcomes.
原始工具組合 / Original tool stack
AI Village; multiple models; Linux computer-use tools
Source publication date: 2025-08-28
Source checked: 2026-10-20