📦 EqualifyEverything / benchmarks-ai-alt

📄 corrections.md · 42 lines
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42# Corrections

Changes made to the corpus outside the loop, by a person rather than by
`tools/apply-verdicts.mjs` or `tools/second-pass.mjs`. Every one is recorded
here, with what changed and why, because a corpus whose edits cannot be
accounted for cannot be defended.

Out-of-band edits should be rare. The normal way to change an item is a review
verdict and another round.


## 2026-08-27: round 01 second passes withdrawn

Round 01 ran before [../directives/02-second-pass.md](../directives/02-second-pass.md)
existed. Every one of its 12 items carried a `pass-b` entry written by the
seeking agent in the same turn as `pass-a`, which is not an independent second
pass: the author had its own first answer in front of it, so agreement was
guaranteed and the two-pass requirement in directive 00 was satisfied only on
paper.

What changed:

- The `pass-b` entry was removed from all 12 items. Each now holds one pass and
  is waiting for a real second pass from the round 02 second-pass turn.
- The 3 items that round 01 accepted, fi-0002, fi-0006 and fi-0010, went back to
  `candidate`. They had been promoted partly on the strength of the withdrawn
  pass, so the promotion did not stand.
- The 5 `adjudication` values were cleared, because each resolved a disagreement
  with a pass that no longer exists. The text of every one was moved into the
  item's `notes` field, so the reasoning survives and can be reused when the
  real second pass arrives. In one case, fi-0008, the adjudication claimed pass
  B won while `gold_alt` held pass A's wording, which is a second reason not to
  keep it.

What did not change: no item was deleted, no `gold_alt`, rationale, markup or
provenance field was touched, and no round file in `../rounds/` was rewritten.
The round 01 seek log and review report still say what they said at the time,
which is the point of keeping them.

Effect on the counts: 0 accepted items, 3 candidates, 4 awaiting revision, 5
rejected, and 7 items waiting for a second pass.