Clarify dataset and trace documentation; improve plots and add command distributions

README.md

@@ -29,31 +29,29 @@ configs:
 Datasets repaired and filtered using the [OT-next-data validation protocol](https://github.com/MLI-lab/OT-next-data/blob/main/validation/PROTOCOL.md).
 
 - **V1:** environment reliability and rule-based task-quality repairs. Build/setup, no-op and reference checks where reference solutions exist.
-- **V2:** retain tasks solved at least once by an evaluated model. Includes partial results where a success is already recorded; each README lists models, attempt counts and coverage.
+- **V2:** keep tasks solved at least once by an evaluated model. Includes partial results where a success is already recorded; each README lists models, attempt counts and coverage.
 
 ## Dataset sizes and composition
 
 **V1: 42,166 kept, patched tasks across 8 data sources.**
 
-**V2: 7,629 tasks solved at least once across the same 8 data sources.** V2 is a subset of V1, not additional tasks.
+**V2: 7,684 tasks solved at least once across the same 8 data sources.** V2 is a subset of V1, not additional tasks.
 
 Percentages are each source's share of that version's total in the current published snapshots.
 
 | Dataset | V1 tasks | Share of V1 | V2 tasks | Share of V2 |
 | --- | ---: | ---: | ---: | ---: |
 | BugsInPy | [489](bugsinpy_v1) | 1.16% | [65](bugsinpy_v2) | 0.85% |
-| CrossCodeEval | [6,710](crosscodeeval_v1) | 15.91% | [863](crosscodeeval_v2) | 11.31% |
-| TermiGen | [3,336](termigen_v1) | 7.91% | [2,277](termigen_v2) | 29.85% |
-| SETA | [3,107](seta_v1) | 7.37% | [2,150](seta_v2) | 28.18% |
-| CalibForge | [5,317](calibforge_v1) | 12.61% | [84](calibforge_v2) | 1.10% |
-| ScaleSWE | [16,786](scaleswe_v1) | 39.81% | [292](scaleswe_v2) | 3.83% |
-| SWE-Lego | [4,277](swelego_v1) | 10.14% | [575](swelego_v2) | 7.54% |
-| EndlessTerminals | [2,144](endless_v1) | 5.08% | [1,323](endless_v2) | 17.34% |
-| **Total** | **42,166** | **100%** | **7,629** | **100%** |
+| CrossCodeEval | [6,710](crosscodeeval_v1) | 15.91% | [863](crosscodeeval_v2) | 11.23% |
+| TermiGen | [3,336](termigen_v1) | 7.91% | [2,277](termigen_v2) | 29.63% |
+| SETA | [3,107](seta_v1) | 7.37% | [2,205](seta_v2) | 28.70% |
+| CalibForge | [5,317](calibforge_v1) | 12.61% | [84](calibforge_v2) | 1.09% |
+| ScaleSWE | [16,786](scaleswe_v1) | 39.81% | [292](scaleswe_v2) | 3.80% |
+| SWE-Lego | [4,277](swelego_v1) | 10.14% | [575](swelego_v2) | 7.48% |
+| EndlessTerminals | [2,144](endless_v1) | 5.08% | [1,323](endless_v2) | 17.22% |
+| **Total** | **42,166** | **100%** | **7,684** | **100%** |
 
-Percentages may not sum to exactly 100% due to rounding.
-
-`tasks.parquet` contains kept tasks. V1 includes `archive.parquet`; V2 exclusions can be recovered from V1 using the selection manifest. V2 reuses V1 images. ScaleSWE uses its original Docker Hub images, with optional Apptainer caches.
+`tasks.parquet` contains kept tasks. V1 includes `archive.parquet`. Tasks left out of V2 remain in V1; `metadata/selection.json` lists which tasks were kept. V2 reuses V1 images. ScaleSWE uses its original Docker Hub images, with optional Apptainer caches.
 
 ## Download tasks and images
 
@@ -68,6 +66,21 @@ snapshot_download(
 )
 ```
 
-Image metadata and reproduction scripts pin immutable source revisions. Their historical paths remain valid at those revisions despite the folder renames on main.
+## Trace stats
+
+[SETA trajectory statistics and plots](seta_v2/trace_stats/README.md) cover turns, tokens, elapsed time, commands and termination outcomes.
+
+Each attempt gets one termination label automatically from the harness stop reason and recorded exception type. No LLM judge assigns these labels. Termination describes why the agent stopped; the verifier reward records whether its work passed.
+
+| Outcome | Meaning |
+| --- | --- |
+| Task complete | The agent declared and confirmed completion; verification can still fail. |
+| tmux session closed | The terminal session disappeared before normal completion confirmation. |
+| Task timeout | The agent exceeded its allowed runtime. |
+| Repeated output truncation | Three consecutive truncated responses could not be recovered into usable responses. |
+| Context recovery failure | The harness failed to recover usable context for another model request. |
+| Other limits | A recorded turn, context, output or task-memory limit stopped the agent. |
+| Error | An exception stopped the run without a more specific termination label. |
+| Not recorded | No termination reason was recorded. |
 
-CalibForge's pilot LLM-review results are in [the GitHub review JSON](https://github.com/MLI-lab/OT-next-data/blob/main/data/calibforge/stage2-review-171/reviews.json). LLM-judge filtering has not been applied across the full release.
+Errors are also grouped by component: **environment/terminal** (container or tmux errors, which may be agent-caused), **verifier** (checking failed to run or finish normally, not a normal reward of 0), **model server** (model API or serving failures), and **other** (unmapped exceptions). An attempt can have multiple exceptions and appear in several error groups; this panel covers all attempts, not just the “Error” termination bar.

seta_v2/trace_stats/README.md

@@ -15,16 +15,30 @@ Medians use recorded measurements only.
 
 ## Plots
 
-Combined distributions of turns, tokens, context use and runtime:
+Turns, token totals and elapsed time per attempt:
 
 ![Trace measurements](plots/all-models-trace-distributions.png)
 
+Token axes use scientific notation (for example, `1e5` means 100,000). Agent elapsed time covers the whole agent phase, including model calls, terminal work and waits; verifier elapsed time covers the checking phase.
+
+The largest-turn panel shows the largest recorded input-plus-output token total for a turn. A turn can include several model calls during recovery. The older `peak_context_tokens` and `peak_context_fraction` fields therefore do not measure actual context occupancy: a ratio of 2 means a recorded turn totaled twice the configured context size, not that one request used twice the context window.
+
+For samples of at least 100 attempts, a panel zooms to the 99th percentile when the maximum is more than 1.5 times that percentile. Its title reports how many values lie beyond the displayed range; full values remain in the data. Otherwise it shows the full range.
+
 Termination reasons and recorded error categories:
 
 ![Trace outcomes](plots/all-models-trace-outcomes.png)
 
 Model-specific plots: [Instruct measurements](plots/qwen3-30b-instruct-2507-trace-distributions.png), [Instruct outcomes](plots/qwen3-30b-instruct-2507-trace-outcomes.png), [Thinking measurements](plots/qwen3-30b-thinking-trace-distributions.png), [Thinking outcomes](plots/qwen3-30b-thinking-trace-outcomes.png).
 
+## Command distribution
+
+![Top 100 commands](plots/all-models-trace-commands.png)
+
+Frequency of recognized submitted command/tool labels across all attempts; these are not counts of successful executions. Unclassified and assignment-only labels are excluded and counted in the figure title. Complex shell syntax is not fully parsed. Each occurrence has equal weight, so longer attempts contribute more commands.
+
+Model-specific command plots: [Instruct](plots/qwen3-30b-instruct-2507-trace-commands.png), [Thinking](plots/qwen3-30b-thinking-trace-commands.png).
+
 ## Data
 
 - [Summary](data/summary.json): overall and per-model statistics.
@@ -38,3 +52,9 @@ The statistics include the published traces and completed recovery evidence. The
 The two `context_recovery_failure` attempts hit a harness recovery bug: repeated truncated-response retries exhausted the remaining context and requested zero output tokens. They are not normal context-limit stops.
 
 “tmux session closed” means the agent terminal disappeared before normal completion confirmation. The raw stop reason remains `session_ended`. All 649 attempts were verified: 53 passed and 596 failed. The [closure audit](data/tmux_session_audit.json) records the command evidence and unresolved cases.
+
+## Generation and resource accounting
+
+Stage 7 computes trajectory metrics, and the report exporter generates plots automatically. This dataset overview is rendered and published by a separate export script; it does not refresh itself when new jobs finish.
+
+Per-trajectory CPU and GPU hours were not measured. Agent elapsed time is wall time, not CPU usage or exclusive GPU time: concurrent attempts share the model server. Scheduler allocation-hours could be reported per job, but assigning them to individual trajectories would require an explicit estimation method.

seta_v2/trace_stats/data/provenance.json

@@ -2878,10 +2878,13 @@
   "figures": [
     "../plots/qwen3-30b-instruct-2507-trace-distributions.png",
     "../plots/qwen3-30b-instruct-2507-trace-outcomes.png",
+    "../plots/qwen3-30b-instruct-2507-trace-commands.png",
     "../plots/qwen3-30b-thinking-trace-distributions.png",
     "../plots/qwen3-30b-thinking-trace-outcomes.png",
+    "../plots/qwen3-30b-thinking-trace-commands.png",
     "../plots/all-models-trace-distributions.png",
-    "../plots/all-models-trace-outcomes.png"
+    "../plots/all-models-trace-outcomes.png",
+    "../plots/all-models-trace-commands.png"
   ],
   "metric_implementation_sha256": "b12538bef1f15032164d4655b65b24080cdad119ebd6df025fe172a43eb32adf",
   "termination_corrections": [
@@ -2899,5 +2902,12 @@
   "display_labels": {
     "session_ended": "tmux session closed"
   },
-  "session_closure_audit": "tmux_session_audit.json"
+  "session_closure_audit": "tmux_session_audit.json",
+  "plotting": {
+    "implementation_sha256": "8c29f3b4116c34aac0adba654144f505d283fb4ee751eb1139d99575ec902ceb",
+    "histogram_range": "For n >= 100 and max > 1.5*p99, display 0..p99 with an explicit omitted-tail count; otherwise 0..max.",
+    "command_weights": "Raw recognized command/tool occurrence counts; unclassified/assignment labels excluded.",
+    "context_note": "Legacy peak_context_* fields derive from per-turn token totals, which may aggregate multiple requests; not actual context occupancy.",
+    "per_trajectory_resource_hours": "Not measured; elapsed time must not be interpreted as exclusive GPU time or CPU usage."
+  }
 }