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.

Preregister isolated source-placement experiment after visibility proof

David Zhangae3530f5dba8
diff --git a/specs/jevgrep/README.md b/specs/jevgrep/README.md
index be7cd3f..bacb9a5 100644
--- a/specs/jevgrep/README.md
+++ b/specs/jevgrep/README.md
@@ -7,9 +7,11 @@ 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: assess the observed tool-output truncation with a no-model presentation
-counterfactual before any further paid experiment. Keep the production renderer
-and skill unchanged. Real SQLite replays match the frozen oracle, and an 81-request
+Next: finish the isolated [source placement experiment](assets/presentation-study.md),
+record its exact candidate identity and focused checks before preparing a new
+plan or making paid calls. Keep the production renderer and skill unchanged.
+The no-model counterfactual preserves all evidence while moving first-ranked
+source inside the observed tool prefix. Real SQLite replays match the frozen oracle, and an 81-request
three-file replay proves both pipelines exceed the follow-up threshold and skip
that pass. Exact coverage and whole-pipeline limits are in the
[mechanism-first skill experiment](assets/query-framing-study.md). That candidate
diff --git a/specs/jevgrep/assets/presentation-study.md b/specs/jevgrep/assets/presentation-study.md
new file mode 100644
index 0000000..03b56fb
--- /dev/null
+++ b/specs/jevgrep/assets/presentation-study.md
@@ -0,0 +1,69 @@
+# Source beside its ranked file: isolated presentation experiment
+
+Production remains the restored frozen policy. The
+[mechanism-first query candidate](query-framing-study.md) is a rejected experimental
+parent, not a new incumbent. This experiment changes only how that parent's
+already-selected evidence is arranged in stdout. It does not change discovery,
+classification, thresholds, source selection, source budgets or the skill.
+
+## No-model prerequisite
+
+The failed parent's complete Django packet is 350,468 bytes, with 212 files and
+96 source blocks. A generic rearrangement places each file's existing blocks
+immediately after its ranked path/role/reading-lead entry. It preserves every
+nonblank line with multiplicity, all numbered source lines including blank source
+lines, ranked file order, and all evidence. The resulting packet is also 350,468
+bytes. Only unnumbered section separator placement may differ.
+
+SQLite's first source moves from byte 61,232 to 480. Its relevant condition moves
+from byte 69,342 to 8,590. The serialized prefix through that condition is 8,891
+characters, versus 39,831 characters before the actual retained outer-tool
+truncation marker. This demonstrates earlier placement within that observed
+prefix allowance; it is not a universal tokenizer or tool-limit guarantee.
+Contract, deterministic script, candidate text and proof are retained at
+`/tmp/jg-presentation-counterfactual/`. No model calls were made.
+
+The failed solver later read and explicitly recognized the condition, then
+replaced its failing same-type regression with a type-changing scenario. Initial
+visibility is therefore only a plausible contributor, not an established cause
+of failure. This presentation experiment may fail to change that decision.
+
+## Candidate and validation contract
+
+Experimental source parent: `99b6afefd0dfe33311a2c1139c67551a2d7e4478`, whose
+properly rebuilt package was the failed query trial. Retain its canonical skill
+unchanged. Arrange stdout as summary/context, followed by each ranked file's
+entry, leads and existing source blocks, followed by the existing terminal
+markers. Retain source allocation, quoting, warnings and all file locations.
+There is no top-N rule or new output cap. Production remains unchanged.
+
+Before any paid call, record the candidate commit and normal-build archive hash,
+verify the installed skill and dependency identity, and run a focused output
+conservation/ordering check. Frozen-output parity is intentionally not claimed
+for this presentation variant; the default production gate is not weakened.
+
+Prepare a new full ten-task plan named `installed-jg-adjacent-source-v1`, validate
+all cells without model calls, then execute Django once. Baselines run zero times.
+Retain the original task, fixed Sol model/effort, 900-second limit, grading and
+accounting. No hints, replacement attempts or candidate edits during the run.
+
+Inspect the actual query and model-visible packet to check whether useful
+top-ranked source survives truncation. Then grade the actual patch and complete
+Sol bill. Comparison: fixed baseline $1.5055472, failed query parent $1.2353512.
+All repairs, tests and model calls belong to that bill; Jev remains separate.
+Query wording and live classification can vary despite the same skill, so a
+single task outcome does not isolate a causal formatting effect.
+
+If Django solves below baseline, continue the same plan's remaining nine cells
+without rerunning it. If it fails, retain and diagnose before any further task
+calls; initial visibility alone cannot justify promotion. A solved cost loss is
+a partial improvement to inspect, not automatic rejection of the approach.
+Promotion still requires every one of eight baseline solves preserved and seven
+successful strict Sol-cost wins across this single candidate cohort. Never pool
+results with earlier candidates or label these tuned tasks untouched validation.
+
+## Status
+
+Offline visibility check passes. Isolated implementation and focused checks are
+in progress. No package identity is registered yet; no paid calls are authorized
+by an unrecorded artifact. No production change or promotion has occurred.