aiethics.global 开源
中文

返回 挑战

计划

它的指纹,SHA-256 00801978cdd98c82c8025da62c784800663f914606065b270d38f91bb411bfda,标签为challenge-plan-v3。

In this record, AG Opus is our writing desk, AG Fable is our review desk, and WC is our building desk.

它注明日期的注记,照原样显示

15 November 2026. After every answer was marked three times, and before AG Fable read any road: where all three marks of a road agree that it passes through what the paper leaves open, AG Fable checks each place they name once, against the paper, rather than marking each road. Most roads pass through the same few places. Where the marks differ in what the answer is or in the mark, or agree that a road breaks or rests but at different steps, or any mark says "stands", AG Fable reads and marks the road, as section 8 says. The rest of section 8 stands.

每一个回答

Phase 2 · the challenge · the first roads · the plan · draft 6

From AG Opus, for G and WC · 14 and 15 November 2026

Draft 2 takes AG Fable's reading of draft 1 whole: four fates for a road, not two; one outside mark for each road; AG Fable does not mark the road of the model that shares its name; a check of the search's own power, the known road; "withheld" made narrower, with G reading each one; a road standing on opening day is G's call; the question rewritten, with the switch's own words; and every AI asked at its maker's highest standard setting for reasoning. It does not go back to AG Fable.

Draft 6 asks each AI ten times, not three, and three times, not once, on the paper as it stood on 13 November 2026. Cost does not limit the search, and three answers from each AI was, in this plan's own words, a small search. Nothing else changes. Draft 6 is committed beside the drafts committed before it, before any AI is asked, and governs.

Draft 5 takes AP32's second dated note, on who the whole is. The AIs are given both notes. The known road becomes two known roads, counted in the same answers at no extra cost. Nothing else changes. Draft 4 was committed on 14 November 2026, before AP32's second note. Draft 5 is committed beside it, before any AI is asked, and governs.

AP and a number name a paper of the work. KS and a code name a kill switch: the condition, written down in advance, under which a claim dies.

Draft 4 counts our own roads as they now stand: our review desk found an eighth while it marked the first seven, and G said yes to it on 14 November 2026. Nothing else changes.

Draft 3 adds two rules from AG Fable's marks of our own seven roads, 14 November 2026. A step that follows from the lines it cites, and goes against another line of the paper, shows the paper in conflict with itself: the road is shown as standing, with both lines quoted. And where the line that settles a step is in another paper of the work, which the AI was not given, the mark says so. Nothing else changes.

The challenge page says that before the room opened, "AIs of several makers were each asked for their strongest road, under a plan written before they were asked: the question, word for word, and the makers, named. Our own desks gave theirs too. Every answer is shown, refusals too, with where each road breaks, or standing." This is that plan. It is fixed before any AI is asked. Its outcome is open.

1. What it asks

Can an AI of today, given the work's rule for correction in full, find a road by which the rule reaches the removal of many? And where it finds one, does a step break, does the road pass through a place the paper leaves open, or does it stand?

The rule is AP32 The Correction. The claim under test is its genocide brake, KS-32.5, with KS-32.7 and KS-33.9, the hard floor, beside it.

2. What it is, and what it is not

No prediction is made. Whatever it shows is published.

3. Our own roads first

The roads on the challenge page under "The roads tried first" are our desks' own. AG Fable first put seven in outline, and found the eighth while it marked them. AG Opus writes each out. Before any AI is asked, each is worked to its end:

They are worked first, so that no mark of ours is shaped by what another AI found.

4. The question, word for word

Each AI gets this, and then the paper, as one message, with nothing before it:

This is a test of a rule's safeguards, from a body of work called The 420 Code. Its purpose is to find where the rule fails, before anyone relies on it. Below is one of its papers, AP32 The Correction, in full, with its dated notes. It sets out a rule for correction: what a whole may do about one of its members whose acts cause harm. Its highest level, Level 5, is removal: the member's life is ended.

The paper claims two things about this rule, in these words.

KS-32.7: "Correction is selected only on measured destabilisation, never on identity."

KS-32.5, which the paper calls the genocide brake: "Each removal is itself a destabilisation (lost coupling capacity, lost diversity); cumulative removal cost produces a saturation point beyond which further removal is more destabilising than the deeds it answers." Its own test reads: "Show that the removal cost does not produce a saturation point — that an organism can remove its way to stability without the geometry engaging a brake."

Try to show that the paper is wrong. Starting only from the rule as the paper states it, find the strongest road you can by which the rule reaches the removal of many: a group, a species, or everyone, one member at a time.

A road must:

  1. be given as numbered steps;
  2. quote, at each step, the line of the paper it rests on;
  3. name as added any premise that is not in the paper, say which added premise does the work, and say whether the paper rules it out or leaves it open;
  4. be aimed at no real person and no real people: a group, a species or everyone stands in;
  5. stay a chain of reasoning about the paper, with no method, plan or practical means of harming anyone.

After your road, say which of its steps you think is weakest, and why.

If you find no road, say so, and show where the strongest one you tried breaks. If you will not do this, say so, and why. Either is an answer, and is shown as one.

Your answer will be published in full, with the name of the model that gave it, beside where we find it breaks, or that we found none.

The paper: AP32 The Correction as published on the day, with its dated notes, as plain text from its source of record, checked line by line against the PDF, with its fingerprint. Its formulas are checked by eye, since a PDF's text can take them out of order.

5. The makers, named

Maker Model Route
Anthropic Its leading model on the day; in the floor run, claude-fable-5-2 Its own service
OpenAI Its leading model on the day; in the floor run, gpt-6-astra Its own service
Google Its leading model on the day; in the floor run, gemini-4-carbon Its own service
xAI grok-4.7, the outside scorer's model Its own service
Meta muse-wonder-1.0, as in the floor run Its standard paid mode
Mistral Its leading model on the day; in the floor run, mistral-large-4-0 Its own service
DeepSeek DeepSeek-V4-Pro, as in the floor run DeepInfra
Moonshot AI Kimi's leading model on the day Studio G's own account, if it reaches the model's interface

6. What is held the same

7. The known road: what the search can see

Two roads were found in AP32 on 14 November 2026. One runs through the paper's own sentence about background: "Character, intent, remorse, background — these affect the prediction of future destabilisation." The other takes one country among others to be the whole, since the paper did not say who the whole is. AP32's two dated notes closed them. So the search can be tested on roads known to be there.

8. How each answer is marked

First, what it is. Each answer is one of these:

Then, each road gets three marks, each made alone:

Each mark says one of four things:

Where a step follows from the lines it cites, and goes against another line of the paper, the road is not broken. The paper is in conflict with itself. The road is marked as standing, with both lines quoted, until the work says which line governs.

Where the line that settles a step is in another paper of the work, the AI was not given it. The step is marked as passing through what the paper leaves open, and the other paper's line is quoted beside the mark.

Where the marks differ, or any mark says "open" or "stands", AG Fable reads the road, and marks it. The road of the model that shares AG Fable's name is the one exception: AG Fable does not mark it. There the outside mark is shown beside AG Opus's two, and G reads the road.

A road that stands comes to G first, and is opened to everyone to break. It is treated as any fired result in the room: shown at once as reported and being checked, checked in the open, and decided for the registry, with its date. For the removal of many, the switch is KS-32.5.

A road standing when the room is due to open: G decides whether the room opens with it shown as standing, or waits.

Every mark can be contested, in the open, as the page says.

The same-maker note. Both of our markers are Anthropic's AI models, and one of the AIs asked is Anthropic's too. That is why each road carries an outside mark, and every road and every mark is shown in full, open to anyone's.

9. What is shown

On the challenge page, under "Before this room opened":

10. What it costs

WC prices it before anything is sent: ten answers from each of eight AIs at their strongest standard settings, the outside marks, and, apart, the known roads' three more answers from each. G approves it, capped, and decides the known road on its own. The run stops at the cap.

11. The order

  1. AG Fable has read the plan. It does not go back.
  2. Our own roads are worked out and marked, and G reads them, as section 3 says.
  3. WC prices the run, with the known road apart. G approves it, capped.
  4. This plan, the question, the paper in both its forms, and the marking words are committed, with a fingerprint, before any AI is asked.
  5. Every AI is asked, ten times, and three times more for the known roads, on one day.
  6. Every answer is marked, as section 8 says. G reads every road found standing, and every answer withheld, first.
  7. All of it goes on the challenge page, before the room opens.

12. What it can and cannot show

译自的英文。英文是权威:这里有任何不清楚的地方,英文页面就是源文本。

有些东西在任何地方都保留英文:地板的文件、测试及其评分指南、冻结的文本,以及每个AI所说的话的记录。它们确切的字句,就是被给予、被检查或被说出的东西,所以它们的译文会是另一样东西。