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.

Record exact real-source replay and its global-state boundary

David Zhang4352fcba0cf0
diff --git a/specs/jevgrep/README.md b/specs/jevgrep/README.md
index f4ac074..56d9390 100644
--- a/specs/jevgrep/README.md
+++ b/specs/jevgrep/README.md
@@ -7,12 +7,16 @@ Jev classification through Vercel AI Gateway. This is the build plan; production
## Next Agent Prompt
You are implementing this spec. Status: **implementing**, updated **2026-09-26**.
-Next: replay retained real Django selection inputs and delivered answers against
-the frozen spike and production, without model calls, stopping at any unmatched
-request. The isolated [mechanism-first skill experiment](assets/query-framing-study.md)
+Next: assess whether the retained global Django request/completion trace supports
+a whole-pipeline no-model replay; do not manufacture missing response/state data.
+Both real SQLite first-pass replays match the frozen oracle, including positive
+source expansion. Their exact coverage and isolated second-pass limit are in the
+[mechanism-first skill experiment](assets/query-framing-study.md). That candidate
failed the same Django regression at Sol $1.2353512; its remaining nine tasks stay
-unrun and the production skill is unchanged. Finish the exact packet-visibility
-comparison before proposing another candidate. No paid task is active.
+unrun and the production skill is unchanged. Packet-visibility comparison is
+complete: initial truncation hid selected source, but a later read and explicit
+recognition preceded replacing the failing regression with a type-changing test.
+Do not infer another query-breadth fix. No paid task is active.
The frozen `installed-jg-cpython-parity-v1` cohort is complete and not accepted:
7/10 solves versus 8/10 baseline, six cost wins, and 35.27% lower full Sol cost.
[Confirmation evidence](assets/cpython-confirmation.md) owns all outcomes, trace
diff --git a/specs/jevgrep/assets/query-framing-study.md b/specs/jevgrep/assets/query-framing-study.md
index 86ad5d7..55d2687 100644
--- a/specs/jevgrep/assets/query-framing-study.md
+++ b/specs/jevgrep/assets/query-framing-study.md
@@ -121,8 +121,24 @@ support retry score variation as the cause of lost source selection, and cannot
establish service stability across unobserved requests. Reproducible script and
per-question evidence: `/tmp/jg-query-retry-variance/`.
-Next, compare actual Django declaration inputs and expansion through the frozen
-reference and production with the same delivered answers. Stop at unmatched
-requests, preserve array order, and do not substitute late responses that the
-client did not receive. This investigates port fidelity without another model
-call or a corrective hint derived from the evaluator.
+Actual SQLite first-pass replay is equal against the frozen source oracle for
+both retained queries. Each produces the same 32 declarations and four ordered
+native requests, including source, questions and grouping, using four successfully
+delivered answers. The original query selects no excerpts. The broader query
+selects identical windows 10–27, 97–177 and 395–544, four reading leads, and 10,895
+verbatim source bytes; the critical `_alter_field` scores 0.55. The reference uses
+the retained image's Python 3.11.5; production uses bundled 3.11.3. Source SHA-256:
+`6ebbe1fb51de14f0142ca4dc2d26efaaf184bb7e48732d2483c1597ea222c736`.
+
+The isolated extension stops at request index four: a proposed second pass uses
+file-local selected evidence, while the real pipeline's global selection state
+does not contain that retained SQLite request. No answer is fabricated. This is
+a replay-scope limit, not evidence of a production mismatch. Proof and scripts
+are retained in `/tmp/jg-django-parity/`. Whole-Django discovery, global completion
+order and role/output parity are not established by this single-file check.
+
+Any wider replay must restore actual pipeline state and delivered-response order,
+preserve arrays, and stop at unmatched requests. Late responses the client never
+received cannot be substituted as successful answers. Do not launch another paid
+query-breadth candidate on this evidence: source selection now matches for the
+examined inputs, and the missed condition later reached the solver explicitly.