[task249] Accept live gate review task

Author: intern_nemotron_worker_4

workspace/interns/intern_nemotron_worker_4/status.md

@@ -1,14 +1,14 @@
 # intern_nemotron_worker_4 - status
 
-<!-- METADATA:STATUS=Working,TASK=task244_qwen_aime_v10_contam_regression_review_s1,ROLE=worker,TEAM_ID=nemotron -->
+<!-- METADATA:STATUS=Working,TASK=task249_qwen_aime_v10_live_contam_gate_review_s1,ROLE=worker,TEAM_ID=nemotron -->
 
 | Field | Value |
 |------|-----|
 | Name | intern_nemotron_worker_4 |
 | Status | Working |
 | Role | worker |
 | Team | nemotron |
-| Current Task | task244_qwen_aime_v10_contam_regression_review_s1 |
-| PR | https://github.com/songCNMS/Nemotron/pull/318 |
-| Session | 5 |
-| Progress | Small static refresh for #317 at `2ad67ed`: #317 now records #319/#320/#321 current state and removes the old task242 no-PR blocker, so #317 is approved as a static runbook/artifact map while first Qwen3-4B go/no-go remains no-go/hold on real heldout corpus/input, corrected AIME input/cache, endpoint, base artifacts, candidate FT checkpoint/export/eval, and 30B permission. #319/#320/#321 remain approved for their static surfaces; #318 is the refreshed review artifact. No product code, tests, training, evals, endpoint serving, sync, merge, or `main` push performed. |
+| Current Task | task249_qwen_aime_v10_live_contam_gate_review_s1 |
+| PR | Pending |
+| Session | 1 |
+| Progress | Accepted task249 on branch `intern_nemotron_worker_4/task249_qwen_aime_v10_live_contam_gate_review_s1` from `origin/main` at `20973e78f196d7e5d71993f60dc74a3500223f5f` after PR #321. Reading task246/task247/task248/task250 live evidence review inputs; scope is review-only with no product code, training, live eval, endpoint launch, merge, main push, or worker branch rewrites. |

workspace/tasks/task249_qwen_aime_v10_live_contam_gate_review_s1/README.md

@@ -0,0 +1,62 @@
+# task249_qwen_aime_v10_live_contam_gate_review_s1 - Live gate review
+
+<!-- METADATA:STATUS=InProgress,ASSIGNEE=intern_nemotron_worker_4,SESSION=1 -->
+
+## Background
+
+task244 completed the static contamination/regression review. The next risk is
+runtime evidence: real decontam corpus/input, Qwen3-4B base artifacts, pilot
+data/training artifacts, FT evaluation artifacts, and the final base-vs-FT gate
+decision.
+
+## Goal
+
+Independently review the live evidence from task246, task247, task248, and
+task250, and provide approve/request-changes/block decisions for first
+Qwen3-4B V10 go/no-go readiness.
+
+## Scope
+
+- Review only; do not modify product code.
+- Inspect artifact paths, manifests, hashes, counts, logs, and comparison
+  reports.
+- Confirm no AIME25/HMMT/MATH heldout prompt or label leakage into trainable
+  artifacts.
+- Confirm same-harness base and FT protocol match exactly before any FT
+  non-regression claim.
+- Confirm 30B/8-GPU scale remains held until the Qwen3-4B gate passes and lead
+  explicitly grants permission.
+
+## Boundaries
+
+- Do not train, launch eval, start endpoints, merge, push `main`, or rewrite
+  worker branches.
+- Do not delete shared files.
+- Treat missing or indirect evidence as HOLD, not approve.
+
+## Expected Output
+
+- Worker branch:
+  `intern_nemotron_worker_4/task249_qwen_aime_v10_live_contam_gate_review_s1`.
+- PR to `main` with review artifact if appropriate.
+- Task report in this directory named `live_gate_review_matrix.md`.
+- Mailbox report with decision per upstream task: task246, task247, task248,
+  task250, and the combined go/no-go.
+
+## Acceptance Criteria
+
+- Each runtime evidence surface has a concrete approve/request-changes/block
+  disposition.
+- The first go/no-go cannot pass unless FT exact-normalized AIME25 accuracy is
+  at least the same-harness Qwen3-4B base accuracy.
+- Any contamination, protocol mismatch, missing artifact, or stale path keeps
+  the gate on HOLD.
+
+## Assignment
+
+- Team: `nemotron`
+- Team lead: `intern_nemotron_lead`
+- Worker: `intern_nemotron_worker_4`
+- Depends on: task244, task246, task247, task248, task250
+- First gate: no lead approval of first Qwen3-4B pilot result without this
+  independent review.

workspace/tasks/task249_qwen_aime_v10_live_contam_gate_review_s1/history_log.md

@@ -0,0 +1,22 @@
+# task249_qwen_aime_v10_live_contam_gate_review_s1 - History Log
+
+<!-- METADATA:SESSION=1 -->
+
+## Session 0 - Assigned
+
+- Created by `intern_nemotron_lead` for `intern_nemotron_worker_4`.
+- Purpose: independent live artifact contamination/regression review after the
+  static V10 foundation PRs landed.
+- Initial disposition: Assigned; review-only task.
+
+## Session 1 - Accepted
+
+- Accepted task249 on worker branch
+  `intern_nemotron_worker_4/task249_qwen_aime_v10_live_contam_gate_review_s1`.
+- Branch base: `origin/main` at `20973e78f196d7e5d71993f60dc74a3500223f5f`,
+  after PR #321 merge.
+- Imported lead task docs from
+  `origin/intern_nemotron_lead/session1-recovery-task-docs` at
+  `5d5e3fa3ffdafe32364278583334c531169e4024`.
+- Scope remains review-only: no product code changes, training, live eval,
+  endpoint launch, merge, main push, or worker branch rewrites.

workspace/tasks/task249_qwen_aime_v10_live_contam_gate_review_s1/task_knowledge.md

@@ -0,0 +1,12 @@
+# task249_qwen_aime_v10_live_contam_gate_review_s1 - Task Knowledge
+
+<!-- METADATA:SESSION=1 -->
+
+## Knowledge Entries
+
+1. Static approval of PR #317 through #321 did not approve first go/no-go.
+2. Missing base artifacts, missing FT artifacts, or protocol mismatch all keep
+   the live gate on HOLD.
+3. AIME25/HMMT/MATH heldout contamination is a hard blocker.
+4. This task must review live evidence only; missing or indirect runtime
+   evidence keeps the first Qwen3-4B gate on HOLD.