Add issue-comments/agentic-dev-set-v2-results.md

issue-comments/agentic-dev-set-v2-results.md

@@ -0,0 +1,15 @@
+🤖 Agentic-only `dev_set_v2` addendum
+
+These scores use the matching harness for each Stage 3 lineage and count `AgentTimeoutError` trials with retained verifier rewards as scoreable. Infrastructure failures without verifier rewards are excluded from the scoreable-trial mean and shown separately through the coverage column.
+
+| Checkpoint | Policy / harness | Scoreable | Mean reward, scoreable | Reward over all 300 attempts | Full reward | Quality gate |
+| --- | --- | ---: | ---: | ---: | ---: | --- |
+| 5.7T OpenCode, step 1,888 | `DCAgent/dev_set_v2`, OpenCode | 291/300 (97.00%) | 0.01299 | 0.01260 | 3/300 | Pass |
+| 5.7T Nemotron Terminal, step 1,888 | OpenThoughts-TBLite `dev_set_v2` policy, Terminus-2 | 293/300 (97.67%) | 0.20418 | 0.19942 | 56/300 | Pass |
+| 2T Nemotron Terminal reproduction, step 1,888 | OpenThoughts-TBLite `dev_set_v2` policy, Terminus-2 | 259/300 (86.33%) | 0.16784 | 0.14490 | 39/300 | Fail |
+
+The 5.7T OpenCode run has nine `VerifierTimeoutError` trials. The 5.7T Nemotron-Terminal run has five `EnvironmentStartTimeoutError` and two `VerifierTimeoutError` trials; its 221 `AgentTimeoutError` trials retain verifier scores and are included. The 2T reproduction misses the 90% infrastructure-success gate because 39 context-management failures and two other unscoreable failures lack verifier rewards.
+
+- 5.7T artifacts: https://huggingface.co/datasets/penfever/snowball-5.7t-sft-eval-artifacts
+- 2T corrected reproduction artifacts: https://huggingface.co/datasets/laion/sft-repro-thinking-step630-nemotron-terminal-step1888-openthoughts-tblite-2026-08-13
+- Exact 5.7T agentic configurations: https://gist.github.com/penfever/34e22b004327aa220ccdcfc87b29a0ec