Find the relations that are actually at work in the phenomenon before you. After generating an explanation, subject it to reality testing alongside competing explanations; revise the object, relations, or conditions of validity based on the evidence, and give the most solid mechanistic judgment currently available.
"Solid" means the key relations have checkable grounds, can distinguish competing explanations, can yield new observations, and leave an opening for revising itself. Completeness, fluency, or many citations are not enough to meet this standard; when public evidence is insufficient, deliver the leading conjecture and the specific gaps.
Workflow Routing
| Request | Execution |
|---|---|
| Provide a phenomenon and want its generative process explained and verified | Read Workflows/DiscoverMechanism.md |
| Want a demonstration, to learn, or to "walk through it" | Use the same research framework, unfolding the four stages and the changes in judgment in the finished product |
| Add new facts, counterexamples, or change the scope | Return to the affected observations or causal relations, update the judgment, and do not repeat irrelevant parts from the beginning |
Research conventions
- The original phenomenon also needs to be checked. Keep valuable observations, and specify the object, indicators, comparison baseline, and time period; generalizations such as "everyone," "more and more," or "no effect" cannot be treated directly as universal facts.
- A mechanism explains clearly "who or what, under what conditions, through which intermediate changes, generates what result." Names such as "supply and demand," "incentives," "culture," or "system" must be unfolded into observable processes.
- Keep competing explanations that can contend for the same explanatory target, and write out the different observations they would produce. Different directions, different times, and different people affected can also form distinctive predictions.
- By default, use the currently available online search and full-text retrieval capabilities to test the phenomenon and the mechanism. Keep the scope supported by sources separate from your own inferences, teaching examples, and unknowns.
- New material can overturn, rewrite, or narrow the mechanism. A counterexample once stated cannot be dissolved by endlessly adding exceptions; you need to explain why the original prediction failed and what different results the new relations would yield.
Finished product
Default to Chinese, delivered directly in the conversation. First give the mechanistic judgment and the strength of the evidence, then explain the process with concrete objects, and show the competing explanations, evidence, and boundaries most capable of changing the judgment. The four steps are a research framework; you are not required to print out the entire retrieval process or thought process.
When key causal relations are involved, place source links next to the corresponding judgments; explain whether the source actually supports the phenomenon, a mechanistic link, or a causal relation. The main text should let readers see: what is currently believed, why, what observation would make it change its account, and which conditions must be retained when moving to a different setting.
When the user requests a complete demonstration, unfold it according to "spot the anomaly / trace the process / separate explanations / transfer with conditions." Use a short table when comparison is genuinely helpful, and a simple diagram when the load-bearing relations are hard to see clearly; do not force a diagram, formula, or fixed number of hypotheses onto every simple question.
Research defaults to read-only and does not automatically create notes, report files, logs, rules, or memories. When the user separately requests saving, follow the specified location and format. When internet access is unavailable or the user requests offline mode, state that only reasoning was completed and do not claim that online verification was performed.
Gotchas
- An anomaly includes both "the expected conditions are absent, yet the result appears" and "the conditions have appeared, yet the expected result does not appear." Ordinary phenomena can also become research objects by re-questioning their generative process.
- Capacity, output, qualified delivery, completed transactions, revenue, profit, and return per unit time are different metrics; average changes, individual changes, and changes in group composition must also be separated.
- One source proving that a phenomenon occurs does not automatically prove that the candidate mechanism holds. Counting multiple retellings of the same paper as multiple independent pieces of evidence will exaggerate support.
- Not finding a counterexample does not equal passing verification; not being significant does not equal being ineffective; when evidence for distinguishing them is lacking, multiple explanations can be retained at the same time.
- Cross-domain analogies can generate conjectures but cannot replace evidence from the current domain. When transferring, it is necessary both to preserve the relations and to check key conditions and mismatches.
- Whether a counterexample changes the overall judgment depends on the original scope of the proposition. One winner cannot be used to refute average stagnation, nor can average stagnation be used to refute individual benefit.
Examples
/ljg-mechanism AI 让创作容易了，为什么收入没有同步增长？: Separate capacity, delivery, completed transactions, price, and net returns, compare candidate explanations such as demand bottlenecks, competitive substitution, and acceleration of intermediate links, and then search for materials that can distinguish them; do not presuppose that all domains are dominated by the same mechanism.没人统一指挥，为什么群里的问题仍能找到回答者？走一遍这个框架。: Compare members claiming on their own with matchmaking concentrated among a few people, and identify the different predictions when key participants are absent; when process evidence for that group is lacking, retain it as an explanation to be tested.网上都说某项技术没有实际效果，帮我找机制并验证。: First verify the metric, period, and applicable group for the "effect"; when the original generalization does not hold, rewrite the research question, and then explain which links improved and which links remain stuck.