Skip to content

Commit 02265c4

Browse files
committed
test(trios): unmasking string literals made the aim worse, and it is taken back
The measurement said the narrowing lands right once in four and named the reason: the spec must name the function. The obvious next evidence was the log event name - queen.review.verdicts is unique and sits as a literal at line 4595, inside exactly the function a human had named by hand - and the localiser blanks literals along with comments, so it could not see it. Keeping the comment masking and dropping the literal masking measured worse: browseros-ai#1156 reset() was handleWorkerFinished worse browseros-ai#1158 acceptanceBlockReason… unchanged browseros-ai#1165 reset() was chooseNextOpenIssue worse browseros-ai#1166 reset() unchanged Three of four in reset(). Literals carry a great many ordinary words, so unmasking fed the density rule the noise that masking comments had been added to remove. Sound reasoning, worse result - which is what measuring is for. Restored to the committed version and re-measured to confirm: browseros-ai#1156 is handleWorkerFinished again. The score stands at one clean hit and one useful neighbour out of four. Refs Trinity-S-AI-DNA/trios#1174
1 parent 944b226 commit 02265c4

1 file changed

Lines changed: 27 additions & 0 deletions

File tree

‎trios/.trinity/NIGHT_LOG.md‎

Lines changed: 27 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -4361,3 +4361,30 @@ The bar set in the contract was three of four. It was not met, and saying so is
43614361
the point of having set it. What the numbers actually show is that the quality
43624362
of the narrowing is bounded by whether the spec names the function — which makes
43634363
this a defect in how the Queen writes specs, not in how she reads files.
4364+
4365+
## A change that made the measurement worse, and was taken back
4366+
4367+
The last measurement said the narrowing lands right once in four, and pointed at
4368+
the reason: the spec has to name the function. The obvious next evidence was the
4369+
log event name — `queen.review.verdicts` is unique, and it sits as a string
4370+
literal at line 4595, inside exactly the function a human had named by hand. The
4371+
localiser could not see it, because it blanks string literals along with
4372+
comments before searching.
4373+
4374+
So: keep masking comments, stop masking literals, and treat an event name found
4375+
in a literal as strong evidence. Measured after:
4376+
4377+
#1156 reset() was handleWorkerFinished worse
4378+
#1158 acceptanceBlockReason… unchanged
4379+
#1165 reset() was chooseNextOpenIssue worse
4380+
#1166 reset() unchanged, still wrong
4381+
4382+
Three of four in `reset()`. String literals carry a great many ordinary words, so
4383+
unmasking them fed the density rule exactly the noise the comment-masking was
4384+
added to remove. The reasoning was sound and the result was worse, which is what
4385+
measuring is for.
4386+
4387+
Taken back to the committed version and re-measured to confirm the restoration:
4388+
#1156 is `handleWorkerFinished` again. The score stands where it stood — one
4389+
clean hit and one useful neighbour out of four — and the next attempt should not
4390+
start by unmasking anything.

0 commit comments

Comments
 (0)