Skip to content

Find code by asking what it does. A CLI for coding agents that uses Jev to discover relevant files and source context.

docs: retain initial paired observations from authorized variance repeat

David Zhange48a517f10d3
diff --git a/specs/done/jevgrep/assets/variance-repeat.md b/specs/done/jevgrep/assets/variance-repeat.md
index 6e1f029..381388f 100644
--- a/specs/done/jevgrep/assets/variance-repeat.md
+++ b/specs/done/jevgrep/assets/variance-repeat.md
@@ -25,3 +25,33 @@ Live execution and sequential grading/accounting belong to sessions `67618` and
`27936`; logs use `/tmp/jg-freshness-repeat-v1b-{run,observer}.log`. Revalidate
their handles before any resume. Final results must be shown alongside the first
cohort, without pooling selected cells or replacing its failed verdict.
+
+## Completed observations
+
+Scikit-learn solved in both corrected-package runs. Repeat Sol cost is $0.3745434
+versus first-run $0.3434830 and baseline $0.2944360, so neither corrected run is a
+cost win. The repeat used 13 generations versus 14, not more generations. Both
+passed the same three test scopes (2, 61 and 172 tests). Its two returned source
+blocks match the first run, while the query, file selection and reading leads
+differ. The repeat additionally corrected RNG initialization to preserve ignored
+invalid seeds when shuffling is disabled, a compatibility issue the first patch
+retained. Official solve equality therefore does not imply identical patch quality.
+Evidence: `/tmp/jg-repeat-sklearn-comparison.md`.
+
+Django failed the same official collation regression in the repeat. Its complete
+Sol bill is $0.9526992 versus $1.0695890 in the first run and $1.5055472 baseline.
+No registered pass-to-pass test failed and no infrastructure failure was reported.
+The repeat therefore also fails the original requirement to preserve every
+baseline solve. This does not replace the user's separate product-acceptance
+decision or authorize another repeat. Jev's known subtotal for this task is
+$0.151341876; its full total is unknown and excluded.
+Unlike the first run, no retained repeat command/output reads the critical SQLite
+guard. The packet returned 42 files without excerpts and ranked SQLite 24th rather
+than ninth. Its narrower regression covered a PK type change with one foreign key,
+not a collation-only change. Final local tests ran 192 with 28 skips (164 passed).
+These are observed differences, not proof that ranking caused the miss. Evidence:
+`/tmp/jg-repeat-django-comparison.md`.
+
+Astropy solved again. Repeat Sol cost is $0.3507346 versus first-run $0.3914126
+and baseline $0.4286310, retaining a successful cost win. Its observed Jev total
+is complete at $0.076535382 and remains excluded from scored Sol cost.