[PierreMardon]

On a blueprint drawn in a host, the developer found the page hard to read in three ways, and each came from the plan it is drawn from.

• Assumptions. The rules the plan settled on its own were marked in the middle of a criterion or a paragraph, five times in one sentence of the same form. The two that were the developer's to judge went by among three premises of fact, and none said what to judge it by.
• Copies. Enumerations chained their items with semicolons, and a notice the feature writes was copied onto the page: its sections, its inventory, its calls and its commands.
• References. The page worked by reference, as the plan does: 53 labels of steps taken from the plan, and references that stood in for what a sentence said ("Check: the one of criterion 3", "undone by step R5"). The rule against repeats had asked for them.

The developer decided on 2026-10-14 and on 2026-10-16. The rules hold for any blueprint, not for that page.

What changes

The assumptions

• A section after the opening, "Assumptions that approval confirms", written only when the plan marks such a rule. It opens on one callout, then one item per rule: the rule in a few words and where the page states it, what left it open, and the alternative set aside with what it would change. The mark where the rule is stated stays.
• Its title says what approving does, not "to confirm": the page is frozen at approval, and an agent that reads "to confirm" on it may take the rule for one still open. It is the reason the mark is already worded so, and a test holds it.
• Only the developer's rules. The plan's other assumptions, a matter of implementation or a premise of fact, are not shown as assumptions. One that touches the developer's control, work or time is said in the sensitive zones, as what the plan does.
• The plan gives for each such rule what leaves it open and the alternative it set aside, since the extractor reads only the plan. A premise of fact it could not check is no such assumption: the plan checks it in the slice that rests on it, or says what that slice does when it proves false.

Reading the page

• Each passage is understood where it stands. The plan is written for agents, which work by reference, and the page is read once by a person: the extractor translates the plan instead of copying its apparatus.
• A name, not a code. A step, a phase or an item is called by a short name in the words of the feature. Numbers stay on the criteria, which the review cites, and where they give an order.
• A reference adds to a sentence, never stands in for it: taken away, the sentence still says what it means. A fact stated elsewhere is named in a few words, and no sentence needs a file the developer does not read. The sensitive zones name the rule of a zone the same way.
• What follows from a rule is not written again. A symmetry, a procedure and its undoing for instance, is stated once, then what departs from it. An omission is what the developer cannot tell from the page.
• Lists. An enumeration of more than three items, or of items that each carry a clause, is a list, never a sentence that parts them with semicolons.
• A document the feature writes for a person is shown by what it does: what it is for, who reads it, what it makes its reader do and what that changes, in what order, under whose rights, and what it forbids. Its text stays in the plan. Nor is an inventory of what already exists copied: the page says what it covers and where it is taken from.

Elsewhere

• The evaluations read the section of the assumptions as part of the frame, not of the body, so the measures of the cut do not count it.
• The guide tells the developer where the assumptions are, why a premise of fact is not among them, and where the text of a document the feature writes is read.

ARCHITECTURE.md is left alone: its sentences on the frame and on the assumptions stay true, and the form of a document an agent writes is not its subject (AGENTS.md).

Tried

A fresh agent drew the blueprint of the host's plan again from the same plan folder, with the extractor and the template of each commit of this branch. The plan had been drafted under the old template of the plan.

	Before	First commit	Both commits
Size	20,498 bytes	21,265 bytes	16,263 bytes
Labels of steps, a letter and a number	41	53	0
References by the number of a criterion	11	10	5, each beside a sentence complete without it
References to a section, or to a file the developer does not read	1	13	0
Calls of an API written out	7	12	5, four of them in criteria copied from the plan
Paragraphs that chain items with semicolons	3	0	0
Rules shown as confirmed by approval	5, scattered	2, gathered	2, gathered

The undoing of the procedure is now said once, each step undone in the reverse order, followed by what it adds: the checks of what the sleep may have worn out, and a manual refresh.

Still there:

• The criteria carry how the plan proves them, and one lists the sections of the notice. The extractor copies the criteria of the plan whole, so that the two agree under the same numbers: the template of the plan is where that changes, An acceptance criterion carries how the plan proves it, and the blueprint copies it whole #125.
• The body still lists the fields the first step records, and some of the technique of the last check.

Checked

• scripts/gate.sh, green locally: 1,902 tests on 3.11 and on the newest Python.

Not measured: no campaign of the evaluations was run. The judge's method still gives a pointer as the way not to repeat, and does not know the section of the assumptions: #124, to land before the next campaign.