Repository navigation
docs(now): two mechanisms the number-audit taxonomy lacked (#2492) - #2498
Merged
Merged
Conversation
The skill sorts twenty-two cases into four ways a true sentence goes false. Tested against this campaign's own failures -- not re-read, tested -- two fit none of them. 8.5 the claim was TRUE and its meaning was not. "26,546 tokens never reach codegen": every word measured, two thirds of it a decision the compiler documents in the function that does it. A wrong number is caught by anyone who re-runs the command; a right one with the wrong meaning survives re-running because re-running reproduces it. 8.6 the observation was right and the REASON was invented. "No success path to break" was true; the crisp mechanical justification beside it was fabricated, and because it read as a considered distinction it stopped anyone measuring for a day. Both are mechanisms rather than variants by the section's own criterion: 8.4's remedy is to adversarially re-derive the accusation, and re-deriving either gives the same answer back. Both needed someone to read the code that produced the behaviour. The index went stale one line after the fix: MEMORY.md said four, and so did the skill's frontmatter description -- the text loaded into context to decide relevance. Both corrected. Mechanism 8.2 in miniature. Refs #2492
Contributor
|
📓 NotebookLM Notebook linked to this PR
This notebook contains session context, decisions, and artifacts for this work. |
Contributor
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The
number-auditskill sorts twenty-two cases into four ways a true sentence goes false. I have cited it all campaign without testing it.Tested — not re-read, tested, by taking each failure this campaign produced and asking which mechanism it belongs to — two had no home.
8.5 — the claim was TRUE and its meaning was not
Every word measured. Two thirds of that total is
forall-quantified statements the compiler documents, in the function that skips them, as "not runtime-checkable". The actionable number was 9,547.It fits none of the four: checked against the thing it describes, the world had not moved, the instrument was honest, the verification was right.
A wrong number is caught by anyone who re-runs the command. A right number with the wrong meaning survives re-running, because re-running reproduces it.
Check: grep the code that produces the behaviour for a comment about it, before publishing the count.
8.6 — the observation was right and the REASON was invented
A tool reported two gates as having "no success path to break". True. The explanation beside it — "forcing a ternary's whole line to 1 is the silent operator seen backwards" — was fabricated and wrong, and because it read as a considered distinction it stopped anyone measuring for a full day.
Check: for every limitation you write down, name the experiment that would show it is not one, and run it. A reason you cannot falsify in one command is a hypothesis wearing a fact's clothes.
Why these are mechanisms and not variants
The section's own criterion is that the fixes do not transfer. 8.4's remedy is to adversarially re-derive the accusation — and re-deriving 26,546 gives 26,546; re-deriving "two gates show no path" gives two gates. Both needed someone to read the code that produced the behaviour, which is a different action from checking the arithmetic.
The index was stale one line after the fix
MEMORY.mdsaid "4 механизма". The skill's own frontmatter description — the text loaded into context to decide whether the skill is relevant at all — said "Four failure mechanisms". Both corrected.That is mechanism 8.2 in miniature, found immediately after adding the section that names it.
Refs #2492