|
|
|
|
|
|
by mengwong
2252 days ago
|
|
|
You're quite right. Garbage in, garbage out. And if the cost of generating garbage goes down, we're going to get a lot more of it! There are definite risks in misinterpretation and complexity and cruft. As Genesereth quoted: The Lord's Prayer is 66 words, the Gettysburg Address is 286 words, there are 1,322 words in the Declaration of Independence, but government regulations on the sale of cabbage total 26,911 words. So, we will have be on our guard. The hope is that if the rules are open and machine-readable, we will be able to counter with software that sides with the user and helps to mount the sort of response that in the past was only available to corporations with very deep pockets. One intriguing approach is to submit test cases: concrete scenarios, or traces of events, that should result in certain desired outcomes. A diverse range of people in different circumstances could be collected in a comprehensive test suite. If the contract/law passes the test suite, you're good! You can imagine two legislators from different parties with different constituents and concerns, each bringing their set of test cases; and when the negotiated compromise passes enough tests, they proceed, without ever actually reading the text of the bill, lol. |
|
|
Not necessarily, you risk the same sort of problems suffered by naïve statisticians who over-fit their data with ever more complicated models. This leads to erratic behaviour in real life and severe lack of predictability outside the range already covered by the data.
You need to run some kind of sensitivity analysis as well because the components used in the implementation (policemen, lawyers, auditors, traffic wardens, etc.) are not all of perfect quality. Think of an audio amplifier design, it looks wonderful on paper and works perfectly in the simulator. But fails spectacularly when built of real world components because each component is not quite exactly as specified. A sensitivity analysis can discover this by, for example, running the simulation multiple times with each component varied according to its expected quality (say +/- 10% for resistors, +/- 30 percent for transistor gain, etc.)