Simplify CalibForge trace plots to output, summaries, and context

calibforge_v2/trace_stats/README.md

@@ -47,28 +47,14 @@ Configured limits are not measurements of token usage. A missing explicit respon
 
 The table reports settings found in **2 sampled saved trial configurations**, not an exhaustive audit of every attempt or direct proof of server settings. Historical runs and retries may differ. Sampling values are recorded explicitly rather than inferred from today's model-card defaults. Agent and verifier wall-clock budgets remain task-specific; sampled resolved limits are in [settings and evidence](../metadata/evaluation-settings-audit.json). Context is not the sum of tokens across all turns. The current forwarding fix does not retroactively change these pass@k results.
 
-## Full-attempt token usage by outcome
+## Token usage and summarization
 
-These statistics cover all 42,536 attempts over 5,317 tasks, including kept and excluded tasks. Outcome is classified by raw verifier reward: success>=1, a recorded lower score, or no verifier score. The unit is one attempt, not the sum of eight attempts for a task.
+All published traces are included, covering kept and excluded tasks and every verifier outcome. One trace means one task attempt.
 
-| Outcome | Attempts | Mean output tokens | Median output tokens | P90 output tokens |
-| --- | ---: | ---: | ---: | ---: |
-| Unsuccessful (graded) | 37,264 | 29,281 | 8,666 | 81,395 |
-| No verifier score | 717 | 32,190 | 10,029 | 90,754 |
-| Successful | 4,555 | 11,323 | 3,857 | 32,973 |
+![Output, summarization, and context distributions](plots/all-models-tokens-by-outcome.png)
 
-Output length is cumulative reported generation per attempt. Input+output usage additionally counts repeated prompt/history tokens on every model call; it is not the unique trajectory length and must not be compared directly to a 25K context or generated-rollout limit. Peak step context is another distinct measurement and can aggregate calls in ATIF records.
+- **Output per recorded agent step:** successful, graded unsuccessful, and ungraded attempts shown separately. Each step contributes once.
+- **Summarizations per trace:** one combined distribution across all traces with a recorded count, including zero.
+- **Peak context per trace:** the largest recorded step prompt + output count in each trace, separated into traces with and without summarization. Unknown summary counts are kept separate when present. Each trace contributes once, regardless of its length or outcome.
 
-![Full-attempt token usage by outcome](plots/token-usage-by-outcome.png)
-
-[Machine-readable distributions and definitions](data/token-usage-by-outcome.json). The earlier token-budget-filter plot shows maximum step lengths; this plot also shows attempt totals.
-
-## Token usage and summarization by outcome
-
-Successful, graded unsuccessful, and ungraded attempts are shown separately, including kept and excluded tasks. Success means raw verifier reward >= 1. Each step contributes once to step distributions; each attempt contributes once to attempt distributions. Longer attempts therefore contribute more steps.
-
-![Token usage and summarization by outcome](plots/all-models-tokens-by-outcome.png)
-
-[Statistics by model and outcome](data/tokens-by-outcome.json) include sample counts, means, medians, P90, P95, maxima, missing measurement counts, and the number of attempts with summaries. Per-model plots are alongside the combined plot.
-
-Context here means recorded step prompt plus completion tokens. ATIF step measurements may aggregate multiple requests; they do not establish a per-call limit. Total generated output covers one attempt. Cumulative input plus output additionally counts repeated history on each call, so it is not unique trajectory length. Unknown summary counts are missing, not zero. These retrospective measurements do not establish the rollout or summarization settings of another training harness.
+[Detailed statistics](data/tokens-by-outcome.json) retain sample counts and missing measurements. Missing summary counts are not zero. ATIF steps can aggregate multiple requests, so recorded context does not establish an exact per-call context limit.

calibforge_v2/trace_stats/data/tokens-by-outcome-provenance.json

@@ -1,5 +1,5 @@
 {
-  "revision": "51e7f1bd260b15dfb03a4427084333b288d54fba",
+  "source_revision": "6664ca996c164c447f1a63722be7017d0a8c7b76",
   "sources": [
     {
       "file": "traces/instruct.parquet",

calibforge_v2/trace_stats/data/tokens-by-outcome.json

@@ -189,6 +189,38 @@
           "context_summarizations": 0
         },
         "attempts_with_summaries": 334
+      },
+      "all_traces_summary_counts": {
+        "n": 42536,
+        "mean": 1.3539119804400979,
+        "median": 0,
+        "p90": 4,
+        "p95": 6,
+        "max": 45
+      },
+      "peak_context_by_summarization": {
+        "With summarization": {
+          "distribution": {
+            "n": 17998,
+            "mean": 29626.86154017113,
+            "median": 25394,
+            "p90": 39587,
+            "p95": 41239,
+            "max": 82956
+          },
+          "missing": 24
+        },
+        "Without summarization": {
+          "distribution": {
+            "n": 24510,
+            "mean": 10069.192370461036,
+            "median": 8766,
+            "p90": 18638,
+            "p95": 21203,
+            "max": 25600
+          },
+          "missing": 4
+        }
       }
     },
     "qwen3-30b-instruct-2507": {
@@ -377,6 +409,38 @@
           "context_summarizations": 0
         },
         "attempts_with_summaries": 334
+      },
+      "all_traces_summary_counts": {
+        "n": 42536,
+        "mean": 1.3539119804400979,
+        "median": 0,
+        "p90": 4,
+        "p95": 6,
+        "max": 45
+      },
+      "peak_context_by_summarization": {
+        "With summarization": {
+          "distribution": {
+            "n": 17998,
+            "mean": 29626.86154017113,
+            "median": 25394,
+            "p90": 39587,
+            "p95": 41239,
+            "max": 82956
+          },
+          "missing": 24
+        },
+        "Without summarization": {
+          "distribution": {
+            "n": 24510,
+            "mean": 10069.192370461036,
+            "median": 8766,
+            "p90": 18638,
+            "p95": 21203,
+            "max": 25600
+          },
+          "missing": 4
+        }
       }
     }
   }

calibforge_v2/trace_stats/plots/all-models-tokens-by-outcome.pdf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:69cb36ea10fd7e9cb1d9326d86cc5b2cbd2fe0d365e9561a5390157606edac95
-size 283413
+oid sha256:05e367d9541ae4b419986d2689e378daec792eb3648b5b373cd3b8606f2107e4
+size 89351

calibforge_v2/trace_stats/plots/qwen3-30b-instruct-2507-tokens-by-outcome.pdf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c9fddad0a7584365ecb543f16ba4c43358b10861df68e0cf378db81ed7d18181
-size 283790
+oid sha256:c40006c3c56a15ba2d23ed2fafc49b57ac7e7993e51cbb777d5b4d7af8d7b7f7
+size 89734