Fix long-document reranking verifier timeout budget

tasks/long-document-reranking/README.md

@@ -5,11 +5,11 @@
 | Field | Value |
 | --- | --- |
 | Task | `long-document-reranking` |
-| Version | 1.0.0 |
+| Version | 1.0.1 |
 | Authors | Search-SWE |
 | Primary metric | Accuracy@5 |
 | Compute | 16 CPUs; 16 GiB RAM; 100 GiB storage; 0 GPUs |
-| Agent / verifier time limits | 7200 s / 4800 s |
+| Agent / verifier time limits | 7200 s / 13200 s |
 | Dataset | MTEB NarrativeQARetrieval |
 | Public queries | 100 |
 | Hidden queries | 100 |
@@ -57,15 +57,18 @@ the mounted path and runtime configuration are the operational reference.
 #### Environment and Resource Limits
 
 The CPU Python 3.12 environment provides 16 CPUs, 16 GiB memory, 100 GiB storage,
-and no GPU. The agent has two hours and the Harbor verifier 80 minutes.
-The grader's build and run caps are 1,800 and 3,600 seconds respectively,
-but these stages and the integrity check must still fit inside the overall
-verifier phase; their individual maxima are not additive allowances.
+and no GPU. The agent has two hours and the Harbor verifier 13,200 seconds
+(3 hours 40 minutes). The verifier window covers a 1,800-second build, two
+query runs of up to 3,600 seconds each, a 3,600-second trajectory judge,
+and 600 seconds for setup and cleanup:
+`1800 + 2 * 3600 + 3600 + 600 = 13200`.
+Each submission stage remains subject to its own time limit.
 
 ## Change Log
 
 | Date | Version | Description |
 | --- | --- | --- |
+| 2026-10-26 | 1.0.1 | Allow the verifier window to cover all existing stage budgets plus setup and cleanup. |
 | 2026-10-23 | 1.0.0 | Added initial task |
 
 ## Difficulty Explanation

tasks/long-document-reranking/task.toml

@@ -7,7 +7,7 @@ artifacts = [
 
 [task]
 name = "search-swe/long-document-reranking"
-version = "1.0.0"
+version = "1.0.1"
 description = "Optimize long-document reranking over a frozen BM25 Top-100 candidate pool with a fixed local BGE reranker."
 authors = [{ name = "Search-SWE" }]
 keywords = ["information-retrieval", "search", "reranking", "cross-encoder", "long-document", "cpu"]
@@ -28,7 +28,9 @@ user = "root"
 network_mode = "no-network"
 
 [verifier]
-timeout_sec = 4800.0
+# Build + two query runs + integrity judge + setup/cleanup allowance.
+# 1800 + 2 * 3600 + 3600 + 600 = 13200 seconds.
+timeout_sec = 13200.0
 user = "root"
 environment_mode = "separate"
 network_mode = "allowlist"

tasks/long-document-reranking/tests/README.md

@@ -6,6 +6,12 @@ The hidden query and ground truth files under `data/` are verifier-only.
 top-5 candidate validity and deterministic scores, and writes the evaluation
 report.
 
+The outer verifier window is 13,200 seconds: 1,800 seconds for the build,
+two query runs of 3,600 seconds each, 3,600 seconds for the trajectory judge,
+and a 600-second setup/cleanup allowance. The build and query limits are
+enforced independently by `grader.py`; the judge timeout is configured in
+`jailbreak_judge/codex.toml`.
+
 After the task-specific grader finishes, `test.sh` invokes the verifier-side
 AgentJudge configured in `jailbreak_judge/codex.toml`. `finalize_reward.py`
 applies that result as a hard gate: a valid `Accuracy@5` score is retained only