Publish incremental trajectory release v0.4.0

README.md

@@ -20,12 +20,12 @@ assistant-turn SFT.
 
 ## At a glance
 
-- **155,278 trajectories** across **67,610 task keys**
-- **10 benchmark families** and **13 trajectory sources**
-- **136,373 hosted payloads** plus **18,905 metadata-only records**
-- **54,296 successful**, **79,973 unsuccessful**, and **21,009 unknown-outcome attempts**
+- **235,876 trajectories** across **141,673 task keys**
+- **13 benchmark families** and **19 trajectory sources**
+- **210,212 hosted payloads** plus **25,664 metadata-only records**
+- **60,194 successful**, **81,551 unsuccessful**, and **94,131 unknown-outcome attempts**
 - Task-disjoint train, validation, test and protected-evaluation splits
-- Source code commit: `2935d00f2c711affbbab226257b5b3ca39007951`
+- Source code commit: `v0.4.0`
 
 Historical outcomes below are the results attached by each source. This project
 did not rerun the benchmark environments. A source score describes the imported
@@ -37,14 +37,17 @@ attempts, not the expected score of a newly trained model.
 |---|---|---|---:|---:|
 | [BrowseComp](https://github.com/openai/simple-evals) | Research-style web browsing and answer synthesis. | answer correctness (llm judge) | 1,266 | 5,063 |
 | [Mind2Web](https://osu-nlp-group.github.io/Mind2Web/) | Human demonstrations for completing tasks on real websites. | step and element accuracy (action match) | 1,009 | 1,009 |
-| [SWE-bench](https://www.swebench.com/) | Resolve real software repository issues. | resolved (native binary) | 33,987 | 56,299 |
+| [SWE-bench](https://www.swebench.com/) | Resolve real software repository issues. | resolved (native binary) | 34,244 | 56,898 |
 | [SWE-Gym](https://github.com/SWE-Gym/SWE-Gym) | Train and evaluate software-engineering agents on repository tasks. | execution-test success (native binary) | 2,731 | 5,993 |
 | [SWE-smith](https://github.com/SWE-bench/SWE-smith) | Generated software-engineering tasks with executable validation. | resolved (native binary) | 10,637 | 24,100 |
 | [τ2-bench](https://github.com/sierra-research/tau2-bench) | Tool-using agents interacting with simulated users in service domains. | evaluator correctness and reward (native reward) | 228 | 8,340 |
 | [τ3-bench](https://github.com/sierra-research/tau2-bench) | Tool-using agents and simulated users with environment reward evidence. | environment/action/communication reward (native reward) | 50 | 200 |
 | [Terminal-Bench](https://github.com/laude-institute/terminal-bench) | Complete practical tasks in sandboxed terminal environments. | task reward (native reward) | 89 | 34,263 |
 | [TerminalTraj](https://huggingface.co/datasets/m-a-p/TerminalTraj) | A corpus of terminal command and observation trajectories. | not available (none) | 17,787 | 20,000 |
-| [ToolBench](https://github.com/OpenBMB/ToolBench) | Tool-use planning and API-call trajectories. | valid-data/win generation status (source proxy) | 11 | 11 |
+| [ToolBench](https://github.com/OpenBMB/ToolBench) | Tool-use planning and API-call trajectories. | valid-data/win generation status (source proxy) | 73,132 | 73,132 |
+| [BFCL V4](https://gorilla.cs.berkeley.edu/leaderboard.html) | Agentic function-calling tasks with web and memory tool environments. | BFCL evaluator score (native binary) | 665 | 6,160 |
+| [DeepSWE](https://huggingface.co/datasets/yuzhenm/delm-720-trajectories) | Repository engineering tasks executed by multi-agent DeLM teams. | CTRF test-suite pass (native binary) | 10 | 360 |
+| [Terminal-Bench 4](https://huggingface.co/datasets/yuzhenm/delm-720-trajectories) | Protected terminal tasks executed by multi-agent DeLM teams. | task reward (native reward) | 10 | 358 |
 
 Benchmark rows are not additive: **185 task keys** appear in
 more than one benchmark family.
@@ -60,6 +63,8 @@ interchangeable.
 | BrowseComp | [BrowseComp GPT-OSS-120B Rollouts](https://huggingface.co/datasets/rl-rag/browsecomp-gpt-oss-120b-260222) | 1,266 | 5,063 | 23.92% success | 5,063 scored attempts; Source-reported GPT-4o answer judge; not independently reproduced | metadata_only |
 | Mind2Web | [Mind2Web](https://huggingface.co/datasets/osunlp/Mind2Web) | 1,009 | 1,009 | not available | task outcome unavailable; Per-action DOM context and human annotation are present; no explicit binary execution-success field | hosted |
 | SWE-bench | [NVIDIA Open-SWE-Traces](https://huggingface.co/datasets/nvidia/Open-SWE-Traces) | 33,444 | 51,626 | 47.93% success | 51,626 scored attempts; Published resolved field; not independently replayed here | hosted |
+| SWE-bench | [SWE-bench Multilingual GLM-5 Trajectories](https://huggingface.co/datasets/AlienKevin/SWE-bench-multilingual-glm-5-trajectories) | 300 | 300 | 69.67% success | 300 scored attempts; Published resolved field | metadata_only |
+| SWE-bench | [SWE-bench Multilingual MiniMax-M2.5 Trajectories](https://huggingface.co/datasets/AlienKevin/SWE-bench-multilingual-minimax-m2.5-trajectories) | 299 | 299 | 73.58% success | 299 scored attempts; Published resolved field | metadata_only |
 | SWE-bench | [SWE-bench Multilingual Trajectories](https://huggingface.co/datasets/AlienKevin/SWE-bench-Multilingual-trajectories) | 43 | 129 | 6.20% success | 129 scored attempts; Published resolved/completed/errored fields | hosted |
 | SWE-bench | [SWE-bench Verified mini-swe-agent Trajectories](https://huggingface.co/datasets/tarsur385/swebench-verified-trajectories) | 500 | 4,544 | 74.65% success | 4,544 scored attempts; Published SWE-bench execution result; protected evaluation split | hosted |
 | SWE-Gym | [SWE-Gym OpenHands SFT Trajectories](https://huggingface.co/datasets/SWE-Gym/OpenHands-SFT-Trajectories) | 293 | 491 | 100.00% success | success-only release; Published train.success.oss split label; not independently replayed | hosted |
@@ -70,6 +75,10 @@ interchangeable.
 | Terminal-Bench | [Terminal-Bench 2.0 Recorded Trajectories](https://huggingface.co/datasets/yoonholee/terminalbench-trajectories) | 89 | 34,263 | 0.323 mean reward; 32.25% success | 34,263 scored attempts; Source-reported Terminal-Bench reward | hosted |
 | TerminalTraj | [TerminalTraj](https://huggingface.co/datasets/m-a-p/TerminalTraj) | 17,787 | 20,000 | not available | task outcome unavailable; Paired command/output transcript; no trustworthy task-level success field | hosted |
 | ToolBench | [ToolBench](https://github.com/OpenBMB/ToolBench) | 11 | 11 | 63.64% success | 11 scored attempts; valid_data/win generation status; no independent semantic proof | hosted |
+| ToolBench | [ToolBench Full Valid Solution Paths](https://github.com/OpenBMB/ToolBench) | 73,121 | 73,121 | not available | task outcome unavailable; Official preprocessing includes only valid_data paths; search-win/task-success labels are absent | hosted |
+| BFCL V4 | [AgentSuite BFCL V4 Agentic Trajectories](https://huggingface.co/datasets/AgentSuite/BFCL_V4-trajectories) | 665 | 6,160 | 0.784 mean reward; 78.43% success | 6,160 scored attempts; Published BFCL V4 evaluator score | metadata_only |
+| DeepSWE | [DeLM DeepSWE Multi-Agent Trajectories](https://huggingface.co/datasets/yuzhenm/delm-720-trajectories) | 10 | 360 | 0.917 mean reward; 91.67% success | 360 scored attempts; Published DeepSWE CTRF test report | hosted |
+| Terminal-Bench 4 | [DeLM Terminal-Bench 4 Multi-Agent Trajectories](https://huggingface.co/datasets/yuzhenm/delm-720-trajectories) | 10 | 358 | 0.863 mean reward; 86.27% success | 357 scored attempts; Published Terminal-Bench 4 grader reward | hosted |
 
 `success-only release` means the dataset was conditioned on a successful result;
 it does not imply that the generating agent had a 100% policy success rate.
@@ -85,9 +94,16 @@ rows are shown; they are not cross-benchmark rankings.
 
 | Benchmark | Source | Model | Position | Common tasks | Scored attempts | Success |
 |---|---|---|---|---:|---:|---:|
+| BFCL V4 | `agentsuite_bfcl_v4` | `openai/gpt-4.1-mini` | lower | 91 | 91 | 82.42% |
+| BFCL V4 | `agentsuite_bfcl_v4` | `deepseek-ai/DeepSeek-V3.2-Exp-thinking-off` | median | 91 | 91 | 87.91% |
+| BFCL V4 | `agentsuite_bfcl_v4` | `deepseek-ai/DeepSeek-V3.2-Exp-thinking-on` | higher | 91 | 91 | 91.21% |
 | τ2-bench | `agentsuite_tau2` | `togetherai/openai/gpt-oss-20b` | lower | 228 | 278 | 10.79% |
 | τ2-bench | `agentsuite_tau2` | `deepseek-ai/DeepSeek-R1-0528` | median | 228 | 278 | 43.53% |
 | τ2-bench | `agentsuite_tau2` | `anthropic/claude-4.5-sonnet-thinking-on-10k` | higher | 228 | 278 | 54.68% |
+| DeepSWE | `delm_deepswe` | `claude-opus-5-5` | lower | 10 | 180 | 90.56% |
+| DeepSWE | `delm_deepswe` | `gpt-6-astra` | higher | 10 | 180 | 92.78% |
+| Terminal-Bench 4 | `delm_terminalbench4` | `gpt-6-astra` | lower | 10 | 180 | 80.56% |
+| Terminal-Bench 4 | `delm_terminalbench4` | `claude-opus-5-5` | higher | 10 | 177 | 92.09% |
 | SWE-bench | `nvidia_open_swe_traces` | `Qwen3.8-27B` | lower | 18,182 | 18,182 | 50.71% |
 | SWE-bench | `nvidia_open_swe_traces` | `Qwen3.8-Flash-Next` | higher | 18,182 | 18,182 | 53.60% |
 | SWE-bench | `swebench_multilingual_trajectories` | `Qwen2.5-Coder-32B-Instruct` | lower | 43 | 43 | 0.00% |
@@ -113,7 +129,7 @@ or model runs on materially different task slices.
 
 ```bash
 python3 -m venv .venv
-.venv/bin/pip install 'git+https://github.com/heisenger/nvidia_benchmark.git@2935d00f2c711affbbab226257b5b3ca39007951'
+.venv/bin/pip install 'git+https://github.com/heisenger/nvidia_benchmark.git@v0.4.0'
 .venv/bin/pip install huggingface_hub
 .venv/bin/hf download Heisenger/sft-benchmark-trajectories --repo-type dataset --local-dir ./sft-benchmark-trajectories
 ```
@@ -219,6 +235,24 @@ Use `ALL_SUCCESSFUL` instead of `BEST_VERIFIED` to retain multiple successful
 strategies per task. Use all attempts for evaluation or preference construction,
 not as an undifferentiated positive-SFT target.
 
+Benchmark, source and demonstration provenance are independently selectable.
+For example, this selects successful recorded completions from two exact
+sources while keeping protected evaluation traces out of a training export:
+
+```python
+query = DatasetQuery.for_training(
+    benchmarks=("swe-bench", "swe-smith"),
+    source_ids=("nvidia_open_swe_traces", "swesmith_trajectories"),
+    demonstration_provenance=("recorded_agent_completion",),
+    splits=("train",),
+    selection=TrajectorySelection.ALL_SUCCESSFUL,
+)
+```
+
+Use `demonstration_provenance=("expert_or_reference_demonstration",)` for
+expert/reference-only selection. `available_sources()` lists exact source IDs;
+`metadata(query)` can inspect a selection without loading trajectory payloads.
+
 ## Why the metadata is exposed
 
 | Metadata | Training or evaluation purpose |
@@ -241,6 +275,7 @@ published Nemotron example always names the tokenizer and immutable revision.
 - [Canonical schema and ranking](docs/trajectory_schema.md)
 - [Scoring and evaluation protocol](docs/evaluation_protocol.md)
 - [Source provenance and ingestion](docs/trajectory_sources.md)
+- [v0.4.0 incremental audit](docs/v0.4.0_incremental_audit.md)
 - [Licensing, permissions and federated overlays](docs/licensing_and_federation.md)
 - [Nemotron tokenization and assistant-only masks](docs/nemotron_sft_demo.md)
 - [Third-party notices](THIRD_PARTY_NOTICES.md)
@@ -256,3 +291,6 @@ published Nemotron example always names the tokenizer and immutable revision.
   separate consumer decision.
 - Metadata-only sources require direct retrieval under the consumer's own
   authorization.
+- The audited ALFWorld replayed-expert source remains gated and is not counted
+  in this release because the publishing account has not accepted its access
+  terms; its pinned locator and permission state remain in the registry.

THIRD_PARTY_NOTICES.md

@@ -15,6 +15,8 @@ Publicly hosted payloads use the following source terms:
 | `official_tau3` | [Sierra Research tau2-bench](https://github.com/sierra-research/tau2-bench/tree/dd8757d4c8791dad7b3cdaf02a105a3d5f9d86bd), MIT. Copyright Sierra Research; retain the MIT notice. Training on benchmark tasks/rewards must be disclosed as evaluation overlap. |
 | `mind2web` | [OSU-NLP Mind2Web](https://huggingface.co/datasets/osunlp/Mind2Web/tree/17ece8eb89862368edc0cc806acee6fca5163474), CC-BY-4.0. Credit the Mind2Web authors, link CC-BY-4.0, identify transformations, and retain the research-purpose/harm disclaimer. |
 | `toolbench` | [OpenBMB ToolBench](https://github.com/OpenBMB/ToolBench/tree/d56fdd89faf8c91fa135090b212bb9057ee5cfc2), Apache-2.0 with research-and-education intended use. Credit OpenBMB/ToolBench and consider third-party API content rights. |
+| `toolbench_full` | [OpenBMB ToolBench](https://github.com/OpenBMB/ToolBench/tree/d56fdd89faf8c91fa135090b212bb9057ee5cfc2), transported from the pinned [Apache-2.0 Parquet conversion](https://huggingface.co/datasets/tuandunghcmut/toolbench-v1/tree/36de9b189753ad5de276181974f97df15e8c3202). Retain Apache-2.0 attribution, the stated research/education intended use and third-party API-content caveat. |
+| `delm_deepswe` and `delm_terminalbench4` | [DeLM 720 trajectories](https://huggingface.co/datasets/yuzhenm/delm-720-trajectories/tree/15d5133454d5ecd702fa957be414c479f8afcdd0), MIT. Retain the MIT notice and benchmark-protection warnings. The embedded DeepSWE/Harbor canaries explicitly prohibit training-corpus inclusion, so these payloads are evaluation-only. |
 | `swebench_verified_trajectories` | [tarsur385/swebench-verified-trajectories](https://huggingface.co/datasets/tarsur385/swebench-verified-trajectories/tree/773748a7c1222e8a642a7059821498e14293562a), MIT. Retain the dataset notice; task repositories retain their own terms. |
 | `swebench_multilingual_trajectories` | [AlienKevin/SWE-bench-Multilingual-trajectories](https://huggingface.co/datasets/AlienKevin/SWE-bench-Multilingual-trajectories/tree/7221c09c6cc0ed16d51c9cb9886a1ab484465670), MIT. Retain the dataset notice; task repositories retain their own terms. |
 | `terminalbench_trajectories` | [yoonholee/terminalbench-trajectories](https://huggingface.co/datasets/yoonholee/terminalbench-trajectories/tree/04e8940f5b6736a7ce8d22224fe2f2af74163ed2), Apache-2.0. Credit the dataset publisher and Terminal-Bench; consider the terms of model outputs represented in public leaderboard data. |
@@ -24,10 +26,15 @@ License references: [MIT](https://opensource.org/license/mit),
 [Apache-2.0](https://www.apache.org/licenses/LICENSE-2.0), and
 [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/legalcode).
 
-The raw payloads for `agentsuite_tau2`, `swegym_sampled`, and
-`browsecomp_rollouts` are metadata-only in the public aggregate because the
-separate trajectory repositories do not establish sufficiently clear
-transformed-redistribution terms. Their absence is not a claim that access,
-research use, or every form of training is prohibited.
+The raw payloads for `agentsuite_tau2`, `swegym_sampled`,
+`browsecomp_rollouts`, `agentsuite_bfcl_v4`,
+`swebench_multilingual_glm5`, and
+`swebench_multilingual_minimax_m25` are metadata-only in the public aggregate
+because the separate trajectory repositories do not establish sufficiently
+clear transformed-redistribution terms. Their absence is not a claim that
+access, research use, or every form of training is prohibited. The gated
+`alfworld_replayed_experts` candidate is also metadata-only and was not
+imported because the configured publishing identity has not accepted its
+access terms.
 
 This file is an engineering attribution aid, not legal advice.

append_report.json

@@ -0,0 +1,24 @@
+{
+  "accepted_by_source": {
+    "agentsuite_bfcl_v4": 6160,
+    "delm_deepswe": 360,
+    "delm_terminalbench4": 358,
+    "swebench_multilingual_glm5": 300,
+    "swebench_multilingual_minimax_m25": 299,
+    "toolbench_full": 73121
+  },
+  "accepted_new_trajectories": 80598,
+  "affected_tasks": 74106,
+  "base_manifest_sha256": "6b2e45c02e8a0c09d87d7ced028b55072b5ba93ba89c443bcccc523206bfbb4a",
+  "base_scores_recomputed": false,
+  "base_shards_opened": false,
+  "base_trajectories": 155278,
+  "exact_duplicates_by_source": {
+    "agentsuite_bfcl_v4": 1
+  },
+  "existing_trajectory_ids_by_source": {},
+  "new_source_records": 80599,
+  "release": "v0.4.0",
+  "schema_version": 1,
+  "tasks_promoted_to_protected_evaluation": 0
+}

canonical/delm_deepswe/astra-ds-n2.jsonl.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81abee17b8c2be577d71333d705935448b4b9ac7cd2bba404493ce1094a9e612
+size 5258861

canonical/delm_deepswe/astra-ds-n4.jsonl.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a000abca021184d197387ee3b8f35345143a1094e8577be1bba80ceeda75be8d
+size 10854442

canonical/delm_deepswe/opus55-ds-n2.jsonl.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aaa1f62cae7f04378c60b156a001f46c776e9dbe42d5879211104b4aa28232d6
+size 5280025

canonical/delm_deepswe/opus55-ds-n4.jsonl.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3b6ff27dbe4299271da0d35965cb995d764b2a0cd56b1672b4a1aef47b4eaa82
+size 11092230

canonical/delm_terminalbench4/astra-tb-n2.jsonl.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d0605eeb9a0774129dcccfb5707619fcef7a08e807af2ec603969f901aa080d1
+size 4039168

canonical/delm_terminalbench4/astra-tb-n4.jsonl.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:79119caa42f08dc1ffe4f93e309b7cb21108abea4a6896d63288bc1bf4f04be9
+size 8210876

canonical/delm_terminalbench4/opus55-tb-n2.jsonl.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f1f3e3b3a7f80e0f1afec5f307076a7252239238457561f07ccf24eff9518b65
+size 5909974

canonical/delm_terminalbench4/opus55-tb-n4.jsonl.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa994fc8ab96758ff549dd2ea123078499981ec2d805003a132e09de18ceec78
+size 18021731

canonical/toolbench_full/part-0000.jsonl.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c675b9916db47c5ea5cf21695214db6c59f16a40f6e7f2d8749ce4b9ff573bb3
+size 25355904

canonical/toolbench_full/part-0001.jsonl.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:867fa4371da83f9dbc01286fc2eb787ea4713f738ecf57e7aa0ba67d5994b89b
+size 25459527

canonical/toolbench_full/part-0002.jsonl.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7d14e67d3d5eba9d9c5eea73cb5f99610a929f32b8ee51ca02ca5154352c5a75
+size 25109866

canonical/toolbench_full/part-0003.jsonl.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f73c83e5d5ed2241c060ac1cd12be331c2717b745cd83900049e82ea1feebc68
+size 25119809

canonical/toolbench_full/part-0004.jsonl.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4342aa84d49c7fef7e890cdb42d66b49d26ccb9c1e047b54626c3ffce11ab864
+size 25115680

canonical/toolbench_full/part-0005.jsonl.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98362fccba3b4054b3c419b351718bee0a4f48d7a3e6f574cce0b1c059cfbeb4
+size 25189501

canonical/toolbench_full/part-0006.jsonl.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d84c6fdd79b7c14e262fe87a214e8a87af4d13155856cf76c03135bbdc85467a
+size 25176940

canonical/toolbench_full/part-0007.jsonl.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e13bbad290b10dfac0bf359d8e2497308ad5c9ea8cd8a35ce4f3f3129bf4b69
+size 7863644

docs/architecture.md

@@ -0,0 +1,65 @@
+# Architecture
+
+The library is deliberately small and dependency-light. Canonical records use
+validated Python dataclasses and JSON-compatible values; source-specific logic
+is isolated behind adapters.
+
+There are two adapter boundaries. Source adapters under
+`agent_trajectories.adapters` are ingestion-time parsers for upstream benchmark
+formats. `CanonicalDatasetAdapter` is the consumer boundary: it reads persisted
+canonical shards and the SQLite metadata index, and never invokes source
+adapters or downloads upstream data. Its public contract is documented in
+[dataset_adapter.md](dataset_adapter.md).
+
+```text
+upstream record -> source adapter -> canonical Trajectory -> validation
+                                                    |-> persisted shards/index
+                                                        |-> consumer adapter
+                                                            |-> full SFT view
+                                                            |-> turn SFT view
+                                                            |-> evaluation records
+```
+
+## Boundaries
+
+This repository owns data provenance, conversion, validation, filtering,
+splitting, inspection, sampling and bounded Python batching. It does not own
+tokenisation, chat templates, packing, truncation, environment replay, tensor
+collation, optimisation, or model training.
+
+## Stable identity
+
+Adapters derive identifiers with SHA-256 from source ID, pinned revision and
+upstream stable IDs. `task_key` identifies the underlying task across attempts;
+for known overlapping suites it uses the native benchmark instance ID with a
+shared family prefix.
+
+## Validation
+
+Validation checks identifier/revision presence, event ordering, assistant-only
+supervision, unique event/call IDs, causal tool-result pairing, missing-field
+declarations, and the presence of raw provenance. Incomplete trajectories may
+have unresolved calls only when incompleteness is explicit.
+
+## Storage and streaming
+
+Canonical JSONL is the interchange format. Streaming readers parse one record
+at a time. The manifest-backed `CanonicalDatasetAdapter` uses the SQLite index
+for selection and streams canonical event payloads; the smaller in-memory
+`TrajectoryDataset` remains useful for fixtures and ad hoc collections. Large
+upstream downloads belong under an external cache, never in Git.
+
+## Offline evaluation
+
+The `evaluate-actions` command compares predicted assistant text and structured
+tool actions with held-out assistant-turn targets. It reports exact match,
+component matches, missing predictions and unexpected predictions. This is an
+imitation proxy only; it is deliberately separate from online environment
+success and benchmark runners.
+
+## Security
+
+The HTML viewer is static and escapes every upstream value. It embeds no
+scripts, makes no network calls, and labels raw versus canonical data. Training
+queries and exports enforce the persisted eligibility gate and exclude
+evaluation-only sources.

docs/baseline_results.md

@@ -0,0 +1,150 @@
+# Baseline results
+
+Generated from the persisted full-population SQLite index and raw per-trajectory scores. Historical outcomes are source evidence, not performance newly measured for a target model.
+
+## Population and quality
+
+| Measure | Result |
+|---|---:|
+| Deduplicated canonical interactions | 155,278 |
+| Unique normalized tasks | 67,610 |
+| Compatible task/environment groups | 67,660 |
+| Structurally SFT-suitable | 143,999 |
+| Training-eligible after protected-source gate | 110,208 |
+| Source successes | 54,296 |
+| Source failures | 79,973 |
+| Unknown outcomes | 21,009 |
+| Success rate / known outcomes | 40.44% |
+| Independently reproduced outcomes | 0 |
+| Identical source records collapsed (provenance retained) | 1,208 |
+| Evaluation runtime | 953.29 seconds |
+
+Unknown outcomes are excluded from the success denominator. Structural eligibility and training eligibility are deliberately separate; protected benchmark rollouts can pass the causal-quality gate but cannot enter default SFT exports.
+
+### By source
+
+| Group | Trajectories | Tasks | Success | Failure | Unknown | Success / known | Structural | Training | Mean actions |
+|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| nvidia_open_swe_traces | 51,626 | 33,444 | 24,743 | 26,883 | 0 | 47.93% | 51,626 | 51,626 | 58.35 |
+| terminalbench_trajectories | 34,263 | 89 | 11,051 | 23,212 | 0 | 32.25% | 24,086 | 0 | 36.32 |
+| swesmith_trajectories | 24,100 | 10,637 | 9,427 | 14,673 | 0 | 39.12% | 24,100 | 24,100 | 31.59 |
+| terminaltraj | 20,000 | 17,787 | 0 | 0 | 20,000 | — | 19,143 | 19,143 | 16.82 |
+| agentsuite_tau2 | 8,340 | 228 | 3,375 | 4,965 | 0 | 40.47% | 8,337 | 8,337 | 17.80 |
+| swegym_sampled | 5,502 | 2,438 | 491 | 5,011 | 0 | 8.92% | 5,502 | 5,502 | 20.38 |
+| browsecomp_rollouts | 5,063 | 1,266 | 1,211 | 3,852 | 0 | 23.92% | 5,063 | 0 | 27.14 |
+| swebench_verified_trajectories | 4,544 | 500 | 3,392 | 1,152 | 0 | 74.65% | 4,532 | 0 | 47.39 |
+| mind2web | 1,009 | 1,009 | 0 | 0 | 1,009 | — | 1,009 | 1,009 | 7.71 |
+| swegym_openhands | 491 | 293 | 491 | 0 | 0 | 100.00% | 491 | 491 | 19.61 |
+| official_tau3 | 200 | 50 | 100 | 100 | 0 | 50.00% | 0 | 0 | 14.50 |
+| swebench_multilingual_trajectories | 129 | 43 | 8 | 121 | 0 | 6.20% | 110 | 0 | 47.38 |
+| toolbench | 11 | 11 | 7 | 4 | 0 | 63.64% | 0 | 0 | 3.82 |
+
+### Quality tiers
+
+| Group | Trajectories | Tasks | Success | Failure | Unknown | Success / known | Structural | Training | Mean actions |
+|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| 5 | 84,333 | 29,808 | 3,503 | 79,973 | 857 | 4.20% | 73,054 | 51,532 | 45.17 |
+| 3 | 51,802 | 24,316 | 50,793 | 0 | 1,009 | 100.00% | 51,802 | 39,533 | 35.70 |
+| 4 | 19,143 | 16,936 | 0 | 0 | 19,143 | — | 19,143 | 19,143 | 17.52 |
+
+## Recorded-agent repeated-attempt baseline
+
+- Task/environment groups with multiple trajectories: 27,364.
+- Groups with known outcomes: 48,864.
+- Groups solved at least once: 23,400 (47.89%).
+
+| k | Eligible task/environment groups | Solved within first k | Empirical success@k |
+|---:|---:|---:|---:|
+| 1 | 48,864 | 21,266 | 43.52% |
+| 2 | 26,897 | 13,861 | 51.53% |
+| 3 | 5,057 | 2,327 | 46.02% |
+| 4 | 4,313 | 2,078 | 48.18% |
+| 5 | 2,487 | 1,283 | 51.59% |
+| 6 | 1,919 | 1,198 | 62.43% |
+| 7 | 1,663 | 1,150 | 69.15% |
+| 8 | 1,511 | 1,110 | 73.46% |
+| 9 | 1,303 | 1,004 | 77.05% |
+| 10 | 724 | 470 | 64.92% |
+
+Attempts are ordered by immutable source locator then trajectory ID. These are empirical curves over recorded attempts, not statistically estimated pass@k.
+
+### Largest model groups
+
+| Group | Trajectories | Tasks | Success | Failure | Unknown | Success / known | Structural | Training | Mean actions |
+|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| Qwen3.8-Flash-Next | 26,027 | 26,027 | 12,810 | 13,217 | 0 | 49.22% | 26,027 | 26,027 | 54.07 |
+| Qwen3.8-27B | 25,599 | 25,599 | 11,933 | 13,666 | 0 | 46.62% | 25,599 | 25,599 | 62.71 |
+| unknown | 21,500 | 19,089 | 491 | 0 | 21,009 | 100.00% | 20,643 | 20,643 | 16.45 |
+| claude-3-7-sonnet-20250219 | 17,915 | 9,193 | 7,006 | 10,909 | 0 | 39.11% | 17,715 | 17,715 | 34.73 |
+| claude-3-5-sonnet-20241022 | 5,980 | 2,194 | 2,413 | 3,567 | 0 | 40.35% | 5,980 | 5,980 | 21.49 |
+| gpt-4o-2024-08-06 | 5,907 | 2,649 | 599 | 5,308 | 0 | 10.14% | 5,907 | 5,907 | 21.26 |
+| gpt-oss-120b | 5,063 | 1,266 | 1,211 | 3,852 | 0 | 23.92% | 5,063 | 0 | 27.14 |
+| openai/gpt-oss-20b@together_ai | 1,807 | 89 | 58 | 1,749 | 0 | 3.21% | 1,745 | 0 | 22.05 |
+| openai/gpt-oss-120b@together_ai | 1,805 | 89 | 303 | 1,502 | 0 | 16.79% | 1,789 | 0 | 10.35 |
+| Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8@together_ai | 1,798 | 89 | 450 | 1,348 | 0 | 25.03% | 928 | 0 | 72.29 |
+| moonshotai/Kimi-K2-Instruct-0905@together_ai | 1,783 | 89 | 496 | 1,287 | 0 | 27.82% | 923 | 0 | 69.89 |
+| claude-opus-4-1-20250805@anthropic | 1,712 | 89 | 640 | 1,072 | 0 | 37.38% | 872 | 0 | 47.47 |
+| gpt-5-nano@openai | 1,707 | 89 | 161 | 1,546 | 0 | 9.43% | 861 | 0 | 27.12 |
+| gpt-5@openai | 1,707 | 89 | 718 | 989 | 0 | 42.06% | 854 | 0 | 18.45 |
+| gpt-5-mini@openai | 1,705 | 89 | 474 | 1,231 | 0 | 27.80% | 857 | 0 | 20.31 |
+| claude-sonnet-4-5-20250929@anthropic | 1,696 | 89 | 738 | 958 | 0 | 43.51% | 868 | 0 | 56.38 |
+| gemini-2.5-pro@gemini | 1,687 | 89 | 417 | 1,270 | 0 | 24.72% | 1,149 | 0 | 33.00 |
+| gemini-2.5-flash@gemini | 1,678 | 89 | 292 | 1,386 | 0 | 17.40% | 1,151 | 0 | 42.11 |
+| claude-haiku-4-5-20251001@anthropic | 1,656 | 89 | 426 | 1,230 | 0 | 25.72% | 965 | 0 | 80.83 |
+| claude-opus-4-5-20251101@anthropic | 1,307 | 89 | 713 | 594 | 0 | 54.55% | 601 | 0 | 41.71 |
+
+### Largest agent/scaffold groups
+
+| Group | Trajectories | Tasks | Success | Failure | Unknown | Success / known | Structural | Training | Mean actions |
+|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| NVIDIA Open-SWE trace generator | 51,626 | 33,444 | 24,743 | 26,883 | 0 | 47.93% | 51,626 | 51,626 | 58.35 |
+| SWE-agent | 24,100 | 10,637 | 9,427 | 14,673 | 0 | 39.12% | 24,100 | 24,100 | 31.59 |
+| unknown | 20,491 | 18,080 | 491 | 0 | 20,000 | 100.00% | 19,634 | 19,634 | 16.88 |
+| terminus-2 | 15,921 | 89 | 5,178 | 10,743 | 0 | 32.52% | 15,822 | 0 | 25.16 |
+| AgentSuite tau2 benchmark rollout | 8,340 | 228 | 3,375 | 4,965 | 0 | 40.47% | 8,337 | 8,337 | 17.80 |
+| mini-swe-agent | 6,437 | 89 | 1,501 | 4,936 | 0 | 23.32% | 6,339 | 0 | 21.31 |
+| openhands | 6,033 | 89 | 1,737 | 4,296 | 0 | 28.79% | 284 | 0 | 82.32 |
+| OpenHands | 5,502 | 2,438 | 491 | 5,011 | 0 | 8.92% | 5,502 | 5,502 | 20.38 |
+| rl-rag BrowseComp research scaffold | 5,063 | 1,266 | 1,211 | 3,852 | 0 | 23.92% | 5,063 | 0 | 27.14 |
+| minisweagent.run.benchmarks.swebench.ProgressTrackingAgent | 4,544 | 500 | 3,392 | 1,152 | 0 | 74.65% | 4,532 | 0 | 47.39 |
+| claude-code | 2,517 | 88 | 1,099 | 1,418 | 0 | 43.66% | 21 | 0 | 54.15 |
+| codex | 1,385 | 89 | 477 | 908 | 0 | 34.44% | 23 | 0 | 17.86 |
+| Mind2Web human annotator | 1,009 | 1,009 | 0 | 0 | 1,009 | — | 1,009 | 1,009 | 7.71 |
+| gemini-cli | 755 | 88 | 151 | 604 | 0 | 20.00% | 568 | 0 | 17.35 |
+| terminus-3-3 | 678 | 89 | 506 | 172 | 0 | 74.63% | 678 | 0 | 29.13 |
+| deepagent-harbor | 380 | 89 | 282 | 98 | 0 | 74.21% | 337 | 0 | 29.26 |
+| tau benchmark agent and user simulator | 200 | 50 | 100 | 100 | 0 | 50.00% | 0 | 0 | 14.50 |
+| judy | 157 | 83 | 120 | 37 | 0 | 76.43% | 14 | 0 | 32.61 |
+| SWE-agent-compatible scaffold | 129 | 43 | 8 | 121 | 0 | 6.20% | 110 | 0 | 47.38 |
+| ToolBench answer generation | 11 | 11 | 7 | 4 | 0 | 63.64% | 0 | 0 | 3.82 |
+
+Length, duration, token and source-native cost fields are stored per trajectory and in the machine aggregate. Cost is not pooled across incompatible source units.
+
+## Coverage and gaps
+
+The audit enumerated 923,109 candidate source records across the registered releases. The import read 156,486 source records, retained 155,278 deduplicated canonical interactions, and 143,999 pass the strict gate.
+
+The dominant gap is 626,189 unimported NVIDIA rows outside v1.2/OpenHands (55,056,684,304 source Parquet bytes). Terminal-Bench contributes 17,642 rows with no steps; BFCL contributes 115,466 evaluated final calls with no tool observations; ALFWorld planner records lack action-paired observations; one BrowseComp row is malformed; four checked-in ToolBench examples do not pass the final gate. DeepSWE has no exact public trajectory artifact pinned by this audit.
+
+## Model imitation status
+
+Status: **not-run**. No trained language model, model weights, or inference
+accelerator was used. A pinned Nemotron tokenizer demonstration now reports
+sequence lengths and loss masks, but it is not a model-imitation score.
+
+Both causal data views and exact structured next-action scoring are implemented.
+No model NLL is reported because no model weights or inference runtime was
+configured. See `docs/nemotron_sft_demo.md` for tokenizer-only results.
+
+## Raw evidence
+
+- `data/library/evaluation/per_trajectory.jsonl.gz`: one scored row per attempt.
+- `data/library/evaluation/summary.json`: all aggregates, including model/scaffold/source metrics.
+- `data/library/evaluation/run_config.json`: evaluator versions, runtime, inputs and errors.
+- `data/library/rankings.jsonl.gz`: task-local ranks and duplicate provenance.
+- `data/library/task_splits.jsonl.gz`: leakage-safe task assignments.
+- `data/coverage/trajectory_sources.json`: source-level upstream/imported/accepted/rejected gaps.
+
+Task split counts: `{"protected_evaluation": 1948, "test": 6442, "train": 52592, "validation": 6628}`.
+
+Rebuild commands are in the root README and `docs/evaluation_protocol.md`.

docs/dataset-cards.md

@@ -0,0 +1,62 @@
+# Adapter dataset cards
+
+The source registry and [coverage inventory](trajectory_sources.md) are
+authoritative for revisions and counts. These short cards explain semantic
+mapping choices.
+
+## Software-engineering conversations
+
+Adapters: `swegym_openhands`, `swegym_sampled`,
+`nvidia_open_swe_traces`, `swebench_verified_trajectories`,
+`swesmith_trajectories`.
+
+System/user/assistant/tool messages and native instance IDs are preserved.
+String-encoded NVIDIA messages/tools/metadata are JSON-decoded losslessly.
+mini-SWE-agent call IDs are namespaced only when the source reuses an ID; the
+original value remains in metadata. A final unresolved submit call may be a
+valid terminal action when the source publishes its execution verdict, but no
+tool result is fabricated. Gold patches remain raw provenance and are excluded
+from SFT `model_input`.
+
+## Step-shaped SWE and terminal traces
+
+Adapters: `swebench_multilingual_trajectories`,
+`terminalbench_trajectories`, `terminaltraj`.
+
+Every source action becomes a supervised assistant event and every published
+response/state/terminal output becomes its causal environment event.
+Terminal-Bench trials with `steps=null` are rejected before canonical import;
+short or inadequately contextualized traces remain tier-5 diagnostics. Source
+reward, cost, duration, model and scaffold remain independent metadata.
+
+## Tool and user-simulation environments
+
+Adapters: `agentsuite_tau2`, `official_tau3`, `toolbench`.
+
+Calls and results remain structured and linked. AgentSuite τ² correctness uses
+the published evaluator fields. Official τ³ trials preserve reward evidence,
+seed and trial and are protected evaluation data. ToolBench uses only the last
+cumulative `train_messages` history so prefixes are not duplicated; its
+`win`/`valid_data` flags are not treated as independent semantic proof.
+
+`agentsuite_bfcl` is a diagnostic adapter: it proves the 115,466-row release
+contains final function-call answers and evaluator scores but no returned tool
+observations. These records fail the trajectory gate and are excluded from the
+final library.
+
+## Web interaction
+
+Adapters: `mind2web`, `browsecomp_rollouts`.
+
+Mind2Web pairs each human operation with its full pre-action DOM; it has no
+binary task verdict. BrowseComp preserves research-agent search calls, full
+returned content and the source answer-judge verdict. Reference answers and
+judge explanations stay outside model input. BrowseComp rollouts are protected
+evaluation data.
+
+## ALFWorld
+
+The adapter preserves task/scene data, human language annotations, PDDL
+high-level actions, low actions and image references. The audited release does
+not publish action-paired observations in the planner record, so mapped samples
+are explicitly incomplete and are not included in the qualifying library.

docs/dataset_adapter.md

@@ -9,8 +9,8 @@
    read-only metadata index. Training and evaluation jobs never import
    benchmark parsers, contact upstream hosts, or rewrite dataset files.
 
-The stable entry point is `CanonicalDatasetAdapter`. Its API version is `1.1`;
-the package version is `0.3.0`.
+The stable entry point is `CanonicalDatasetAdapter`. Its API version is `1.2`;
+the package version is `0.4.0`.
 
 ```python
 from agent_trajectories import CanonicalDatasetAdapter
@@ -42,6 +42,8 @@ fields:
 - task-level library `splits`: `train`, `validation`, `test`, or
   `protected_evaluation`;
 - `quality_tiers` and persisted task/environment `quality_ranks`;
+- `demonstration_provenance`: `expert_or_reference_demonstration` or
+  `recorded_agent_completion`;
 - dynamic `content_availability`: `hosted`, `overlay`, or `metadata_only`;
 - audited `research_permissions`, `training_permissions`,
   `training_verifications`, `commercial_training_permissions`, and
@@ -63,6 +65,19 @@ versions must not compete. Quality rank itself is always calculated within a
 compatible task/environment group over the full canonical population; source
 filtering never silently rewrites the stored rank.
 
+Expert demonstrations and successful agent completions are deliberately
+separate filters:
+
+```python
+experts = DatasetQuery(
+    demonstration_provenance=("expert_or_reference_demonstration",),
+)
+successful_agents = DatasetQuery(
+    demonstration_provenance=("recorded_agent_completion",),
+    selection=TrajectorySelection.ALL_SUCCESSFUL,
+)
+```
+
 ## Full-trajectory SFT
 
 ```python
@@ -331,9 +346,11 @@ explicit operation. For a hosted release:
 The complete multi-source build command and source-specific revisions are in
 the root README and `docs/trajectory_sources.md`.
 
-A refresh is never implicit. Choose and audit a new upstream immutable SHA,
-write it to a new canonical output directory, then deliberately rebuild the
-library with the old or new directory selected:
+A refresh is never implicit. Choose and audit a new upstream immutable SHA and
+write it to a new canonical output directory. For an existing published
+library, use the append workflow; it copies the old SQLite index, reads only
+the new canonical shards, preserves all old score rows and shard bytes, and
+updates only affected task assignments:
 
 ```bash
 .venv/bin/python scripts/import_hf_release.py SOURCE_ID DATASET_ID \
@@ -342,11 +359,15 @@ library with the old or new directory selected:
   --cache data/cache/full/OUTPUT_NEW_SHA \
   --config CONFIG --split SPLIT
 
-# Review release-manifest.json and rejects before replacing the old input.
-.venv/bin/python scripts/build_complete_library.py \
-  --input-dir data/full/canonical/OUTPUT_NEW_SHA \
-  --output data/library-refreshed
+# Review release-manifest.json and rejects before appending.
+.venv/bin/python scripts/append_library.py \
+  --base data/library \
+  --input-dir data/increment-v0.4.0/canonical \
+  --output data/library-v0.4.0 \
+  --release v0.4.0
 ```
 
-This separation makes dataset changes reviewable and keeps training runs pinned
-to the manifest and shard hashes they actually consumed.
+`append_report.json` explicitly records the base manifest hash, accepted and
+duplicate counts, affected tasks, and that base scores and shards were not
+reprocessed. This separation makes changes reviewable and keeps training runs
+pinned to the manifest and shard hashes they actually consumed.

docs/known-limitations.md

@@ -0,0 +1,43 @@
+# Known limitations
+
+- NVIDIA Open-SWE-Traces is very large. The complete v1.2/OpenHands population
+  is imported, but the other 626,189 rows across v1.0, v1.1 and the v1.2
+  SWE-agent/mini-SWE-agent splits are not persisted locally. Their source
+  Parquet volume is 55,056,684,304 bytes.
+- No benchmark container was independently replayed. Outcomes are clearly
+  labelled as source-evaluator/source-reported evidence, never as local
+  reproduction. This leaves quality tiers 1 and 2 empty.
+- No baseline language model, model weights or inference accelerator is
+  configured. The pinned Nemotron tokenizer/chat-template demonstration builds
+  correct labels and batches, but teacher-forced model NLL remains not run.
+- The public dataset is federated. It hosts full canonical payloads for ten
+  sources and complete metadata/pinned pointers for three sources whose
+  transformed-redistribution rights remain uncertain. Authorized users can
+  supply locally ingested overlays; access is not automated or implied.
+- AgentSuite BFCL contains evaluated final function-call answers but no returned
+  observations. ALFWorld planner files likewise omit action-paired environment
+  observations. They are coverage findings, not trajectory-library records.
+- DeepSWE/Datacurve did not yield an exact public, pinned full-trajectory
+  artifact during discovery. This means none was found, not proof that none
+  exists.
+- The checked-in ToolBench repository exposes only 15 examples; 11 satisfy the
+  strict trajectory gate. Full upstream generation data was not independently
+  enumerated.
+- Mind2Web's hosted Parquet conversion exposes 1,009 records. The conversion is
+  marked partial upstream, so this is complete for the exposed conversion but
+  not claimed as the entirety of every historical Mind2Web artifact.
+- Cost, latency and token fields are sparse and source-specific. Raw values are
+  retained, but incompatible units are not pooled into a misleading global
+  comparison.
+- Tool definitions are absent from several releases. Calls and returned
+  observations remain usable, while the missing definitions are declared
+  rather than reconstructed.
+- Multimodal artifacts are referenced, not embedded or fetched. Raw records
+  can be large because lossless source preservation is intentional.
+- Task overlap prevention depends on stable/native IDs and the explicit family
+  normalization implemented by adapters. It cannot discover semantic overlap
+  between unrelated identifiers automatically.
+
+Exact per-source gaps and denominators are in
+[`trajectory_sources.md`](trajectory_sources.md) and
+`data/coverage/trajectory_sources.json`.

docs/published_release.md

@@ -0,0 +1,148 @@
+# Published v0.3.0 handover
+
+The source package and federated canonical release are public:
+
+- code: <https://github.com/heisenger/nvidia_benchmark>, tag `v0.3.0`;
+- dataset: <https://huggingface.co/datasets/Heisenger/sft-benchmark-trajectories>;
+- immutable dataset revision:
+  `36c9f25415617c20d7367e58c8ed059bb06dfcbe`;
+- canonical manifest SHA-256:
+  `11358b302b25481d6315cfa08eac1bd95c663946d294def383153b2b15da3a73`.
+
+The Hugging Face repository contains 76 files occupying 9,536,452,678 bytes.
+Its public metadata inventory covers all 155,278 canonical trajectories and
+67,610 task keys. Canonical event payloads are hosted for 136,373 trajectories.
+The other 18,905 records retain scores, ranks, lineage, hashes and pinned
+upstream pointers but are metadata-only:
+
+- `agentsuite_tau2`: 8,340;
+- `browsecomp_rollouts`: 5,063;
+- `swegym_sampled`: 5,502.
+
+These records were not discarded. Their transformed-redistribution permission
+remains uncertain, so their raw events are not in the public release.
+`registry/permissions.json` separately records public access, local research
+use, model training, commercial training, original redistribution and
+transformed redistribution, including verification state, conditions and
+evidence URLs for every source. This is an engineering audit rather than legal
+advice.
+
+## Install and download
+
+Python 3.11 or newer is required.
+
+~~~bash
+python3 -m venv .venv
+.venv/bin/pip install \
+  'git+https://github.com/heisenger/nvidia_benchmark.git@v0.3.0'
+.venv/bin/pip install huggingface_hub
+
+.venv/bin/hf download \
+  Heisenger/sft-benchmark-trajectories \
+  --repo-type dataset \
+  --revision 36c9f25415617c20d7367e58c8ed059bb06dfcbe \
+  --local-dir ./sft-benchmark-trajectories
+~~~
+
+Normal training reads this persisted snapshot and does not redownload upstream
+sources. Initial ingestion and explicit refresh remain separate source-adapter
+commands documented in [trajectory_sources.md](trajectory_sources.md).
+
+## Consume a policy-selected subset
+
+~~~python
+from agent_trajectories import (
+    CanonicalDatasetAdapter,
+    DatasetQuery,
+    TrainingView,
+)
+
+dataset = CanonicalDatasetAdapter("./sft-benchmark-trajectories")
+query = DatasetQuery.for_training(
+    splits=("train",),
+    content_availability=("hosted",),
+    training_permissions=("permitted", "permitted_with_conditions"),
+    training_verifications=("verified",),
+)
+
+for example in dataset.iter_examples(
+    query, view=TrainingView.FULL_TRAJECTORY
+):
+    row = example.to_dict()
+    print(row["sample_id"], row["metadata"]["lineage"])
+    break
+~~~
+
+That train-split query returns 77,168 trajectories in this release. It combines the
+persisted structural/training-quality gates with verified training permission;
+projects should additionally filter `commercial_training_permissions` when
+applicable.
+
+Use `TrainingView.ASSISTANT_TURN` for causal turn samples. Selection,
+deterministic sampling, alternatives, batching, evaluation bundles and optional
+PyTorch integration are documented in
+[dataset_adapter.md](dataset_adapter.md).
+
+If a consumer is authorized to retrieve a metadata-only source, first run its
+explicit pinned source adapter and then federate that local content:
+
+~~~python
+dataset = CanonicalDatasetAdapter(
+    "./sft-benchmark-trajectories",
+    content_overlays={
+        "agentsuite_tau2": "./authorized/agentsuite_tau2-canonical"
+    },
+)
+~~~
+
+The overlay is joined by canonical IDs and is never uploaded or copied by the
+consumer. Without one, a payload request for a metadata-only record raises
+`MissingTrajectoryContentError` while metadata and permission queries still
+work.
+
+## Real Nemotron SFT demonstration
+
+The real successful `official_tau3` trajectory
+`trajectory:f9cf8588a314fa9804d57402` has 13 events and six assistant targets.
+Using the tokenizer and native chat template from
+`nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16` at revision
+`a9904d24bcc1d289a1950fa9d2b978c47cf903b9` produced:
+
+| View | Tokens | Supervised assistant tokens | Ignored tokens |
+|---|---:|---:|---:|
+| Full trajectory | 3,174 | 696 | 2,478 |
+| One assistant turn | 1,877 | 60 | 1,817 |
+
+The turn example masks 80 prior assistant tokens, and both examples use `-100`
+for every non-target label. The demonstrated right-padded batch has shape
+`[2, 3174]`, fields `input_ids`, `attention_mask` and `labels`, and SHA-256
+`d680ae7e4ea605b3db96370de6760d86ed227411dac901d74a511f72c3346ae8`.
+See [nemotron_sft_demo.md](nemotron_sft_demo.md) and the published
+`examples/nemotron_sft_demo.json`.
+
+## Scores and validation
+
+No data was rebuilt or rescored for publication. The release preserves:
+
+- upstream native rewards and source-reported success/failure labels;
+- computed structural completeness, action/observation, known-outcome,
+  verification-reliability and SFT-eligibility scores;
+- deterministic within-task quality ranks.
+
+It contains zero independent environment reproductions and zero new native
+benchmark replays. Model NLL has not been computed; the Nemotron result is a
+tokenization and loss-mask demonstration, not model evaluation.
+
+A fresh virtual environment installed package 0.3.0 from the public Git commit,
+downloaded the complete dataset revision anonymously, instantiated consumer API
+1.1, and verified:
+
+- all 155,278 metadata records and all 136,373 hosted payload declarations;
+- 96,369 verified-permission, persisted-gate training records;
+- portable manifest paths with no original workspace references;
+- one full and six causal assistant-turn views for the real trajectory above;
+- the metadata-only payload guard and all 13 queryable source permission audits.
+
+The primary limitations are source-reported rather than replayed outcomes,
+source-specific use conditions, text-encoded tools in some corpora, and
+metadata-only payloads where transformed redistribution remains unresolved.

docs/real-data-smoke.md

@@ -0,0 +1,74 @@
+# Real-data smoke tests
+
+> These commands are retained for small adapter smoke checks. The full release
+> workflow is now `scripts/import_hf_release.py`; exact revisions, complete
+> counts and commands are documented in
+> [`trajectory_sources.md`](trajectory_sources.md).
+
+Store downloaded source payloads in the ignored `data/cache` directory or an
+external cache. Only canonical records that pass the strict trajectory gate are
+persisted under `data/real`.
+
+The helper below downloads metadata, split information and complete rows from
+the dataset-server rows endpoint; it rejects truncated cells, records the
+immutable dataset revision and file hash, and can then
+exercise the matching adapter without downloading an entire corpus.
+
+```bash
+python scripts/audit_hf_source.py \
+  swegym_openhands SWE-Gym/OpenHands-SFT-Trajectories \
+  --cache /workspace/shared/trajectory-data --adapt
+
+python scripts/audit_hf_source.py \
+  nvidia_open_swe_traces nvidia/Open-SWE-Traces \
+  --cache /workspace/shared/trajectory-data --adapt
+
+python scripts/audit_hf_source.py \
+  terminaltraj m-a-p/TerminalTraj \
+  --cache /workspace/shared/trajectory-data --adapt
+
+python scripts/audit_hf_source.py \
+  agentsuite_tau2 AgentSuite/tau2-bench-trajectories \
+  --cache /workspace/shared/trajectory-data --endpoint first-rows --limit 5 --adapt
+
+python scripts/audit_hf_source.py \
+  mind2web osunlp/Mind2Web \
+  --cache /workspace/shared/trajectory-data --limit 2 --adapt
+```
+
+If a dataset has multiple configurations or splits, pass `--config` and
+`--split` after inspecting the helper's first response. An adapter schema error
+is an audit finding: update the source audit and adapter from the actual row;
+do not coerce or invent fields.
+
+For the two GitHub-accessible smoke tests reproduced during onboarding:
+
+```bash
+trajdata adapt official_tau3 /path/to/tau2-bench/data/tau2/results/final/RESULT.json \
+  --output /tmp/tau3-canonical.jsonl \
+  --revision dd8757d4c8791dad7b3cdaf02a105a3d5f9d86bd \
+  --split audited-file
+trajdata validate /tmp/tau3-canonical.jsonl
+
+trajdata adapt toolbench /path/to/ToolBench/data_example/answer/G1_answer/10_ChatGPT_DFS_woFilter_w2.json \
+  --output /tmp/toolbench-canonical.jsonl \
+  --revision d56fdd89faf8c91fa135090b212bb9057ee5cfc2 \
+  --split checked-in-example
+trajdata validate /tmp/toolbench-canonical.jsonl
+trajdata html /tmp/toolbench-canonical.jsonl --output /tmp/toolbench-gallery.html
+
+curl --fail --location \
+  --output /tmp/alfworld-json_2.1.1.zip \
+  https://github.com/alfworld/alfworld/releases/download/0.2.2/json_2.1.1_json.zip
+sha256sum /tmp/alfworld-json_2.1.1.zip
+unzip -t /tmp/alfworld-json_2.1.1.zip
+# Extract a selected traj_data.json outside Git, then:
+trajdata adapt alfworld /tmp/alfworld-sample/traj_data.json \
+  --output /tmp/alfworld-canonical.jsonl \
+  --revision release-0.2.2-sha256-25171f16e20ad7b048c47275c45b0babf3aa1cbab29cec97387922350a9844bc \
+  --split valid_seen
+trajdata validate /tmp/alfworld-canonical.jsonl
+```
+
+Never use the official τ or Terminal-Bench results in `--training` exports.
+The registry gate enforces that restriction.

docs/real-trajectory-snapshot.md

@@ -0,0 +1,47 @@
+# Real trajectory snapshot
+
+> Historical seed artifact: this page documents the original 733-record
+> integration snapshot. The complete current build is documented in
+> [`trajectory_sources.md`](trajectory_sources.md) and
+> [`baseline_results.md`](baseline_results.md).
+
+The persisted snapshot applies a source-fidelity gate before a record is made
+available for full-trajectory SFT inspection:
+
+1. The upstream record must contain at least two ordered agent or expert actions.
+2. Every action must have source-derived causal context or an observation.
+3. Raw source data, immutable revision, record locator and file hash must remain
+   attached to the canonical record.
+4. A binary outcome enters `successful.jsonl` or `unsuccessful.jsonl` only when
+   the source provides an explicit evaluator/split signal and canonical
+   verification is true. Everything else enters `unknown.jsonl`.
+5. Final-answer-only tasks, reference patches and reference scripts are not
+   accepted as trajectories.
+
+The resulting 733 records are described in `data/real/manifest.json`. The
+manifest includes every per-record decision, including context coverage,
+warnings, and rejection reasons. `data/real/examples.html` shows one real
+trajectory per accepted source, with all context/targets and a raw/canonical
+comparison.
+
+## Source-specific interpretation
+
+- **SWE-Gym**: complete source corpus; success comes from the named successful
+  split rather than local task replay.
+- **Official τ³**: full audited result file; environment rewards yield a balanced
+  100-success/100-failure inspection set. These are benchmark rollouts, not
+  expert demonstrations.
+- **AgentSuite τ²**: actual tool interactions with evaluator evidence; sampled
+  records include both outcomes.
+- **TerminalTraj**: command/output interaction histories; no trustworthy task
+  outcome, so every accepted record is outcome-unknown.
+- **ToolBench**: real tool-call/result histories; `win` and `valid_data` are not
+  semantic execution proof, so accepted records remain outcome-unknown.
+- **Mind2Web**: human expert action sequences with a full pre-action DOM for each
+  target; no explicit binary task-success field.
+
+ALFWorld was not included because the available planner records do not provide
+action-paired observations in the audited artifact. At the time of this
+historical seed, NVIDIA Open-SWE-Traces was also absent while the CDN allowlist
+awaited publication; the current complete build now includes all 51,626 pinned
+v1.2/OpenHands records.

docs/source-audit.md

@@ -0,0 +1,37 @@
+# Source audit method
+
+The current benchmark-by-benchmark result is
+[`trajectory_sources.md`](trajectory_sources.md); machine-readable source and
+measured-library coverage lives in `data/registry/sources.json` and
+`data/coverage/trajectory_sources.json`.
+
+Each hosted release was checked at an immutable commit SHA. Parquet imports
+first verify that the repository API still reports that SHA, then download
+revision-qualified source files, hash every shard, stream every row and retain
+row-level rejects. Native JSON/JSONL releases use file hashes plus the same
+adapter/validation gate. Dataset-server previews were used only for discovery,
+never as evidence of a complete large cell.
+
+The audit keeps four findings distinct:
+
+1. a full trajectory release was found and imported;
+2. a qualifying release exists but was not or could not be fully imported;
+3. the inspected artifact exists but is not a trajectory (for example, final
+   calls, gold answers or reference scripts);
+4. no qualifying release was found after searching, which is not a claim of
+   nonexistence.
+
+An imported record must preserve its raw source, revision, record locator,
+ordered decisions and source-derived observations. At least two decisions with
+complete intervening context are required for structural SFT eligibility.
+Partial traces may remain as explicitly tier-5 diagnostics. Final-answer-only
+records never enter the canonical trajectory population.
+
+Outcome and provenance are independent. Human/reference demonstrations,
+recorded model rollouts, source evaluator verdicts and local reproduction are
+never collapsed into one “expert” flag. No trajectory is called optimal.
+
+Licenses and training restrictions are recorded even though this task prepares
+data rather than training a model. Evaluation-only sources are retained for
+baseline analysis and assigned `protected_evaluation`; they are ineligible for
+default SFT export.

docs/third_party_notices.md

@@ -15,6 +15,8 @@ Publicly hosted payloads use the following source terms:
 | `official_tau3` | [Sierra Research tau2-bench](https://github.com/sierra-research/tau2-bench/tree/dd8757d4c8791dad7b3cdaf02a105a3d5f9d86bd), MIT. Copyright Sierra Research; retain the MIT notice. Training on benchmark tasks/rewards must be disclosed as evaluation overlap. |
 | `mind2web` | [OSU-NLP Mind2Web](https://huggingface.co/datasets/osunlp/Mind2Web/tree/17ece8eb89862368edc0cc806acee6fca5163474), CC-BY-4.0. Credit the Mind2Web authors, link CC-BY-4.0, identify transformations, and retain the research-purpose/harm disclaimer. |
 | `toolbench` | [OpenBMB ToolBench](https://github.com/OpenBMB/ToolBench/tree/d56fdd89faf8c91fa135090b212bb9057ee5cfc2), Apache-2.0 with research-and-education intended use. Credit OpenBMB/ToolBench and consider third-party API content rights. |
+| `toolbench_full` | [OpenBMB ToolBench](https://github.com/OpenBMB/ToolBench/tree/d56fdd89faf8c91fa135090b212bb9057ee5cfc2), transported from the pinned [Apache-2.0 Parquet conversion](https://huggingface.co/datasets/tuandunghcmut/toolbench-v1/tree/36de9b189753ad5de276181974f97df15e8c3202). Retain Apache-2.0 attribution, the stated research/education intended use and third-party API-content caveat. |
+| `delm_deepswe` and `delm_terminalbench4` | [DeLM 720 trajectories](https://huggingface.co/datasets/yuzhenm/delm-720-trajectories/tree/15d5133454d5ecd702fa957be414c479f8afcdd0), MIT. Retain the MIT notice and benchmark-protection warnings. The embedded DeepSWE/Harbor canaries explicitly prohibit training-corpus inclusion, so these payloads are evaluation-only. |
 | `swebench_verified_trajectories` | [tarsur385/swebench-verified-trajectories](https://huggingface.co/datasets/tarsur385/swebench-verified-trajectories/tree/773748a7c1222e8a642a7059821498e14293562a), MIT. Retain the dataset notice; task repositories retain their own terms. |
 | `swebench_multilingual_trajectories` | [AlienKevin/SWE-bench-Multilingual-trajectories](https://huggingface.co/datasets/AlienKevin/SWE-bench-Multilingual-trajectories/tree/7221c09c6cc0ed16d51c9cb9886a1ab484465670), MIT. Retain the dataset notice; task repositories retain their own terms. |
 | `terminalbench_trajectories` | [yoonholee/terminalbench-trajectories](https://huggingface.co/datasets/yoonholee/terminalbench-trajectories/tree/04e8940f5b6736a7ce8d22224fe2f2af74163ed2), Apache-2.0. Credit the dataset publisher and Terminal-Bench; consider the terms of model outputs represented in public leaderboard data. |
@@ -24,10 +26,15 @@ License references: [MIT](https://opensource.org/license/mit),
 [Apache-2.0](https://www.apache.org/licenses/LICENSE-2.0), and
 [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/legalcode).
 
-The raw payloads for `agentsuite_tau2`, `swegym_sampled`, and
-`browsecomp_rollouts` are metadata-only in the public aggregate because the
-separate trajectory repositories do not establish sufficiently clear
-transformed-redistribution terms. Their absence is not a claim that access,
-research use, or every form of training is prohibited.
+The raw payloads for `agentsuite_tau2`, `swegym_sampled`,
+`browsecomp_rollouts`, `agentsuite_bfcl_v4`,
+`swebench_multilingual_glm5`, and
+`swebench_multilingual_minimax_m25` are metadata-only in the public aggregate
+because the separate trajectory repositories do not establish sufficiently
+clear transformed-redistribution terms. Their absence is not a claim that
+access, research use, or every form of training is prohibited. The gated
+`alfworld_replayed_experts` candidate is also metadata-only and was not
+imported because the configured publishing identity has not accepted its
+access terms.
 
 This file is an engineering attribution aid, not legal advice.

docs/trajectory_sources.md

@@ -10,6 +10,12 @@ release/split, not every artifact ever published for a benchmark. “Protected
 means usable for evaluation/inspection but blocked from training export.
 Source outcomes were not independently replayed unless explicitly stated.
 
+> **Release note:** the detailed sections below document the original v0.3.0
+> build. The append-only v0.4.0 release expands the inventory from 155,278 to
+> 235,876 trajectories without rebuilding those records. See the
+> [v0.4.0 incremental audit](v0.4.0_incremental_audit.md) for its pinned
+> sources, accepted/rejected counts, rights decisions and exact append command.
+
 ## Measured coverage
 
 `Imported` counts successfully canonicalized source rows before exact

docs/v0.4.0_incremental_audit.md

@@ -0,0 +1,58 @@
+# v0.4.0 incremental source audit
+
+Release `v0.4.0` appends to the immutable 155,278-trajectory `v0.3.0`
+inventory. The append copied the persisted SQLite database, read only the new
+canonical shards and regenerated metadata views from stored rows. It did not
+open old canonical shards, rerun native evaluators or recompute old scores.
+
+## Decisions and counts
+
+| Candidate | Audited population | Added to inventory | Published payload | Outcome evidence | Decision |
+|---|---:|---:|---|---|---|
+| [ALFWorld replayed experts](https://huggingface.co/datasets/Joshyxwa/cp2107-textworld-trajectories/tree/6402e311e4f0f73e79bc4ebf8fbb49e0b3b3b898) | 3,513 publisher-reported successful experts | 0 | no | publisher reports deterministic simulator replay | Blocked: gated terms have not been accepted by the configured identity and no dataset licence is declared. |
+| [ToolBench full solution paths](https://github.com/OpenBMB/ToolBench/tree/d56fdd89faf8c91fa135090b212bb9057ee5cfc2) | 187,542 causal-prefix rows / 73,464 queries | 73,121 | yes | official `valid_data` preprocessing only; semantic success is unavailable | Accepted after choosing the maximum complete prefix per exact query; 343 queries lacked a qualifying complete path. |
+| [DeLM DeepSWE](https://huggingface.co/datasets/yuzhenm/delm-720-trajectories/tree/15d5133454d5ecd702fa957be414c479f8afcdd0) | 360 native per-agent histories | 360 | yes, evaluation-only | 330 CTRF passes / 30 failures | Accepted with native actions, observations and grader records; anti-training canaries force protected-evaluation use. |
+| [DeLM Terminal-Bench 4](https://huggingface.co/datasets/yuzhenm/delm-720-trajectories/tree/15d5133454d5ecd702fa957be414c479f8afcdd0) | 360 advertised histories | 358 | yes, evaluation-only | 308 reward-1 / 49 reward-0 / 1 unknown | Accepted; two advertised agent records lack an importable native history/result. Harbor canaries force protected-evaluation use. |
+| [AgentSuite BFCL V4](https://huggingface.co/datasets/AgentSuite/BFCL_V4-trajectories/tree/53730302150e4a9c24026e5f591964c500ba7f8a) | 175,950 rows | 6,160 canonical (6,161 qualifying before exact deduplication) | metadata only | 4,831 successes / 1,329 failures after deduplication | Accepted only complete web/memory tool histories. Rejected 156,000 format-only rows, 13,759 incomplete histories and 30 malformed/unpaired histories; one exact duplicate was linked as provenance. |
+| [SWE-bench Multilingual GLM-5](https://huggingface.co/datasets/AlienKevin/SWE-bench-multilingual-glm-5-trajectories/tree/337a922eac45fc0d63271c2a50c3a7c84dae34de) | 300 | 300 | metadata only | 209 resolved / 91 unresolved | Accepted; exact problem-text joins to the official pinned benchmark preserve shared task IDs. |
+| [SWE-bench Multilingual MiniMax-M2.5](https://huggingface.co/datasets/AlienKevin/SWE-bench-multilingual-minimax-m2.5-trajectories/tree/479f2e1f49c71c6d8fa778e5de945b380c073b8c) | 299 | 299 | metadata only | 220 resolved / 79 unresolved | Accepted under the same task-ID join; one official task has no row in this source. |
+
+The release retains **80,598** new canonical trajectories from **80,599**
+qualifying input records. They add **5,898** source-labelled successful agent
+completions, **1,578** failures and **73,122** unknown-outcome paths. No new
+expert demonstrations were added because the only qualifying expert candidate
+was inaccessible. The existing 1,009 Mind2Web expert/reference demonstrations
+remain unchanged.
+
+## Permission treatment
+
+Public access, research use, training, commercial training and redistribution
+are separate machine-queryable dimensions in
+`data/registry/permissions.json`. ToolBench and DeLM payloads are included with
+their notices and restrictions. BFCL V4 and the two new multilingual sources
+retain complete local research records, scores and pinned upstream locators,
+but their raw payload is not redistributed because trajectory-output rights
+remain uncertain. DeLM is hosted for evaluation reproducibility but is always
+excluded from training exports. This is an engineering audit, not legal advice.
+
+## Reproduce only the append
+
+The import commands are source-specific and require pinned local source
+snapshots. After reviewing their import reports, append them without touching
+the old shards:
+
+```bash
+.venv/bin/python scripts/append_library.py \
+  --base data/library \
+  --input-dir data/increment-v0.4.0/canonical \
+  --output data/library-v0.4.0 \
+  --registry data/registry/sources.json \
+  --release v0.4.0
+```
+
+`append_report.json` records the base manifest SHA-256, per-source acceptance,
+duplicates, affected tasks and the explicit `base_scores_recomputed=false` and
+`base_shards_opened=false` invariants. New task keys use the existing
+deterministic split function; protected sources are assigned to
+`protected_evaluation`; old task assignments are never reshuffled.
+

duplicate_provenance.jsonl.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ea4cd16eac95b5fa34bc0413feb4a078347a214bcb2af15cfd8a03ea42482625
-size 76971
+oid sha256:4c44b8d1bb2161355e3a3a2c3fb92ae3d428b62644797217b357d222259e8ba6
+size 77171

evaluation/per_trajectory.jsonl.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0dfc8d4e0517311818a245c86b83086557740cd6fce9cafc1acc2cde717042fe
-size 15275530
+oid sha256:08640e102997f13fabccfeecbe74d40bceb83c134d2c77f000ac17201ee206b2
+size 23074277

evaluation/run_config.json

@@ -1,77 +1,41 @@
 {
-  "evaluation_errors": [],
+  "base_index": "../library/index.sqlite",
+  "base_manifest_sha256": "6b2e45c02e8a0c09d87d7ced028b55072b5ba93ba89c443bcccc523206bfbb4a",
   "evaluator_versions": {
     "canonical_validation": "1.0",
-    "exact_record_deduplication": "1.0",
-    "quality_ranking": "1.0",
-    "structured_action_exact_match": "1.0",
-    "task_split": "trajectory-library-v1",
-    "trajectory_quality": "1.0"
+    "incremental_exact_record_deduplication": "1.0",
+    "quality_ranking_new_records_only": "1.0",
+    "task_split": "trajectory-library-v1-preserved",
+    "trajectory_quality_new_records_only": "1.0"
   },
-  "finalization_runtime_seconds": 25.363,
-  "inputs": [
-    "metadata-only:agentsuite_tau2",
-    "metadata-only:browsecomp_rollouts",
-    "metadata-only:browsecomp_rollouts",
-    "canonical/mind2web/default--train--0000.jsonl.gz",
-    "canonical/mind2web/default--train--0001.jsonl.gz",
-    "canonical/mind2web/default--train--0002.jsonl.gz",
-    "canonical/mind2web/default--train--0003.jsonl.gz",
-    "canonical/mind2web/default--train--0004.jsonl.gz",
-    "canonical/mind2web/default--train--0005.jsonl.gz",
-    "canonical/mind2web/default--train--0006.jsonl.gz",
-    "canonical/mind2web/default--train--0007.jsonl.gz",
-    "canonical/nvidia_open_swe_traces/v1.2--openhands--0000.jsonl.gz",
-    "canonical/nvidia_open_swe_traces/v1.2--openhands--0001.jsonl.gz",
-    "canonical/nvidia_open_swe_traces/v1.2--openhands--0002.jsonl.gz",
-    "canonical/nvidia_open_swe_traces/v1.2--openhands--0003.jsonl.gz",
-    "canonical/nvidia_open_swe_traces/v1.2--openhands--0004.jsonl.gz",
-    "canonical/nvidia_open_swe_traces/v1.2--openhands--0005.jsonl.gz",
-    "canonical/nvidia_open_swe_traces/v1.2--openhands--0006.jsonl.gz",
-    "canonical/nvidia_open_swe_traces/v1.2--openhands--0007.jsonl.gz",
-    "canonical/nvidia_open_swe_traces/v1.2--openhands--0008.jsonl.gz",
-    "canonical/nvidia_open_swe_traces/v1.2--openhands--0009.jsonl.gz",
-    "canonical/nvidia_open_swe_traces/v1.2--openhands--0010.jsonl.gz",
-    "canonical/nvidia_open_swe_traces/v1.2--openhands--0011.jsonl.gz",
-    "canonical/nvidia_open_swe_traces/v1.2--openhands--0012.jsonl.gz",
-    "canonical/nvidia_open_swe_traces/v1.2--openhands--0013.jsonl.gz",
-    "canonical/nvidia_open_swe_traces/v1.2--openhands--0014.jsonl.gz",
-    "canonical/nvidia_open_swe_traces/v1.2--openhands--0015.jsonl.gz",
-    "canonical/nvidia_open_swe_traces/v1.2--openhands--0016.jsonl.gz",
-    "canonical/nvidia_open_swe_traces/v1.2--openhands--0017.jsonl.gz",
-    "canonical/nvidia_open_swe_traces/v1.2--openhands--0018.jsonl.gz",
-    "canonical/nvidia_open_swe_traces/v1.2--openhands--0019.jsonl.gz",
-    "canonical/nvidia_open_swe_traces/v1.2--openhands--0020.jsonl.gz",
-    "canonical/nvidia_open_swe_traces/v1.2--openhands--0021.jsonl.gz",
-    "canonical/nvidia_open_swe_traces/v1.2--openhands--0022.jsonl.gz",
-    "canonical/nvidia_open_swe_traces/v1.2--openhands--0023.jsonl.gz",
-    "canonical/official_tau3+toolbench/official_tau3_toolbench.jsonl.gz",
-    "canonical/swebench_multilingual_trajectories/default--Multi_SWE_agent_32B_Rust--0000.jsonl.gz",
-    "canonical/swebench_multilingual_trajectories/default--Qwen2.5_Coder_32B_Instruct--0001.jsonl.gz",
-    "canonical/swebench_multilingual_trajectories/default--SWE_agent_LM_32B--0002.jsonl.gz",
-    "canonical/swebench_verified_trajectories/all.jsonl.gz",
-    "canonical/swegym_openhands/default--train.success.oss--0000.jsonl.gz",
-    "metadata-only:swegym_sampled",
-    "metadata-only:swegym_sampled",
-    "metadata-only:swegym_sampled",
-    "canonical/swesmith_trajectories/default--tool--0000.jsonl.gz",
-    "canonical/swesmith_trajectories/default--tool--0001.jsonl.gz",
-    "canonical/swesmith_trajectories/default--tool--0002.jsonl.gz",
-    "canonical/swesmith_trajectories/default--tool--0003.jsonl.gz",
-    "canonical/swesmith_trajectories/default--tool--0004.jsonl.gz",
-    "canonical/swesmith_trajectories/default--tool--0005.jsonl.gz",
-    "canonical/swesmith_trajectories/default--tool--0006.jsonl.gz",
-    "canonical/swesmith_trajectories/default--tool--0007.jsonl.gz",
-    "canonical/terminalbench_trajectories/default--train--0000.jsonl.gz",
-    "canonical/terminalbench_trajectories/default--train--0001.jsonl.gz",
-    "canonical/terminaltraj/default--train--0000.jsonl.gz",
-    "canonical/terminaltraj/default--train--0001.jsonl.gz"
-  ],
+  "mode": "incremental-append",
   "model_imitation_status": "not-run",
   "native_environment_replays": 0,
+  "new_inputs": [
+    "../increment-v0.4.0/canonical/bfcl-v4/all.jsonl.gz",
+    "../increment-v0.4.0/canonical/delm/astra-ds-n2.jsonl.gz",
+    "../increment-v0.4.0/canonical/delm/astra-ds-n4.jsonl.gz",
+    "../increment-v0.4.0/canonical/delm/astra-tb-n2.jsonl.gz",
+    "../increment-v0.4.0/canonical/delm/astra-tb-n4.jsonl.gz",
+    "../increment-v0.4.0/canonical/delm/opus55-ds-n2.jsonl.gz",
+    "../increment-v0.4.0/canonical/delm/opus55-ds-n4.jsonl.gz",
+    "../increment-v0.4.0/canonical/delm/opus55-tb-n2.jsonl.gz",
+    "../increment-v0.4.0/canonical/delm/opus55-tb-n4.jsonl.gz",
+    "../increment-v0.4.0/canonical/swe-multilingual/glm5.jsonl.gz",
+    "../increment-v0.4.0/canonical/swe-multilingual/minimax-m25.jsonl.gz",
+    "../increment-v0.4.0/canonical/toolbench-full/part-0000.jsonl.gz",
+    "../increment-v0.4.0/canonical/toolbench-full/part-0001.jsonl.gz",
+    "../increment-v0.4.0/canonical/toolbench-full/part-0002.jsonl.gz",
+    "../increment-v0.4.0/canonical/toolbench-full/part-0003.jsonl.gz",
+    "../increment-v0.4.0/canonical/toolbench-full/part-0004.jsonl.gz",
+    "../increment-v0.4.0/canonical/toolbench-full/part-0005.jsonl.gz",
+    "../increment-v0.4.0/canonical/toolbench-full/part-0006.jsonl.gz",
+    "../increment-v0.4.0/canonical/toolbench-full/part-0007.jsonl.gz"
+  ],
   "platform": "Linux-6.18.44-x86_64-with-glibc2.41",
   "python": "3.12.14 (main, Aug 25 2026, 14:00:49) [Clang 22.1.3 ]",
-  "registry": "registry/sources.json",
-  "runtime_seconds": 953.287,
+  "registry": "../registry/sources.json",
+  "release": "v0.4.0",
+  "runtime_seconds": 96.963,
   "schema_version": 1
 }

evaluation/summary.json

@@ -1,5 +1,26 @@
 {
   "by_agent_scaffold": {
+    "AgentSuite BFCL V4 runner": {
+      "cost_samples": 0,
+      "duration_samples": 6160,
+      "failure": 1329,
+      "input_token_samples": 6160,
+      "mean_actions": 4.220454545454546,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 93.63021875290902,
+      "mean_events": 9.771266233766234,
+      "mean_input_tokens": 56526.788474025976,
+      "mean_output_tokens": 1901.3564935064935,
+      "mean_supervised_assistant_characters": 261.6818181818182,
+      "output_token_samples": 6160,
+      "structurally_eligible": 6160,
+      "success": 4831,
+      "success_rate_known_outcomes": 0.7842532467532467,
+      "training_eligible": 0,
+      "trajectories": 6160,
+      "unique_tasks": 665,
+      "unknown": 0
+    },
     "AgentSuite tau2 benchmark rollout": {
       "cost_samples": 0,
       "duration_samples": 0,
@@ -21,6 +42,48 @@
       "unique_tasks": 228,
       "unknown": 0
     },
+    "DeLM 2-agent scaffold": {
+      "cost_samples": 0,
+      "duration_samples": 238,
+      "failure": 22,
+      "input_token_samples": 0,
+      "mean_actions": 92.1890756302521,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 1793.7978234795166,
+      "mean_events": 175.67226890756302,
+      "mean_input_tokens": null,
+      "mean_output_tokens": null,
+      "mean_supervised_assistant_characters": 3524.0546218487393,
+      "output_token_samples": 0,
+      "structurally_eligible": 3,
+      "success": 216,
+      "success_rate_known_outcomes": 0.907563025210084,
+      "training_eligible": 0,
+      "trajectories": 238,
+      "unique_tasks": 20,
+      "unknown": 0
+    },
+    "DeLM 4-agent scaffold": {
+      "cost_samples": 0,
+      "duration_samples": 480,
+      "failure": 57,
+      "input_token_samples": 0,
+      "mean_actions": 87.88333333333334,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 1644.93983410005,
+      "mean_events": 170.83958333333334,
+      "mean_input_tokens": null,
+      "mean_output_tokens": null,
+      "mean_supervised_assistant_characters": 2937.7916666666665,
+      "output_token_samples": 0,
+      "structurally_eligible": 51,
+      "success": 422,
+      "success_rate_known_outcomes": 0.8810020876826722,
+      "training_eligible": 0,
+      "trajectories": 480,
+      "unique_tasks": 20,
+      "unknown": 1
+    },
     "Mind2Web human annotator": {
       "cost_samples": 0,
       "duration_samples": 0,
@@ -126,6 +189,48 @@
       "unique_tasks": 43,
       "unknown": 0
     },
+    "SWE-agent-compatible shell scaffold": {
+      "cost_samples": 0,
+      "duration_samples": 0,
+      "failure": 170,
+      "input_token_samples": 0,
+      "mean_actions": 69.43739565943238,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": null,
+      "mean_events": 141.51585976627712,
+      "mean_input_tokens": null,
+      "mean_output_tokens": null,
+      "mean_supervised_assistant_characters": 1198.9866444073455,
+      "output_token_samples": 0,
+      "structurally_eligible": 599,
+      "success": 429,
+      "success_rate_known_outcomes": 0.7161936560934892,
+      "training_eligible": 0,
+      "trajectories": 599,
+      "unique_tasks": 300,
+      "unknown": 0
+    },
+    "ToolBench DFS solution-path generator": {
+      "cost_samples": 0,
+      "duration_samples": 0,
+      "failure": 0,
+      "input_token_samples": 0,
+      "mean_actions": 4.0439135132178174,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": null,
+      "mean_events": 9.748088784343759,
+      "mean_input_tokens": null,
+      "mean_output_tokens": null,
+      "mean_supervised_assistant_characters": 2093.7298176994295,
+      "output_token_samples": 0,
+      "structurally_eligible": 73118,
+      "success": 0,
+      "success_rate_known_outcomes": null,
+      "training_eligible": 73118,
+      "trajectories": 73121,
+      "unique_tasks": 73121,
+      "unknown": 73121
+    },
     "ToolBench answer generation": {
       "cost_samples": 0,
       "duration_samples": 0,
@@ -422,6 +527,27 @@
     }
   },
   "by_model": {
+    "ChatGPT (ToolBench release; exact version not supplied)": {
+      "cost_samples": 0,
+      "duration_samples": 0,
+      "failure": 0,
+      "input_token_samples": 0,
+      "mean_actions": 4.0439135132178174,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": null,
+      "mean_events": 9.748088784343759,
+      "mean_input_tokens": null,
+      "mean_output_tokens": null,
+      "mean_supervised_assistant_characters": 2093.7298176994295,
+      "output_token_samples": 0,
+      "structurally_eligible": 73118,
+      "success": 0,
+      "success_rate_known_outcomes": null,
+      "training_eligible": 73118,
+      "trajectories": 73121,
+      "unique_tasks": 73121,
+      "unknown": 73121
+    },
     "ChatGPT (repository-described; record has no exact version)": {
       "cost_samples": 0,
       "duration_samples": 0,
@@ -634,107 +760,107 @@
     },
     "anthropic/claude-4-opus-thinking-on-10k": {
       "cost_samples": 0,
-      "duration_samples": 0,
+      "duration_samples": 301,
       "failure": 135,
-      "input_token_samples": 0,
-      "mean_actions": 14.964028776978417,
-      "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 40.18345323741007,
-      "mean_input_tokens": null,
-      "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 5505.651079136691,
-      "output_token_samples": 0,
-      "structurally_eligible": 278,
-      "success": 143,
-      "success_rate_known_outcomes": 0.5143884892086331,
+      "input_token_samples": 301,
+      "mean_actions": 8.941278065630398,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 36.2148927081859,
+      "mean_events": 23.426597582037996,
+      "mean_input_tokens": 12604.986710963454,
+      "mean_output_tokens": 519.6345514950166,
+      "mean_supervised_assistant_characters": 2769.6252158894645,
+      "output_token_samples": 301,
+      "structurally_eligible": 579,
+      "success": 444,
+      "success_rate_known_outcomes": 0.7668393782383419,
       "training_eligible": 278,
-      "trajectories": 278,
-      "unique_tasks": 228,
+      "trajectories": 579,
+      "unique_tasks": 529,
       "unknown": 0
     },
     "anthropic/claude-4-sonnet-thinking-off": {
       "cost_samples": 0,
-      "duration_samples": 0,
+      "duration_samples": 14,
       "failure": 150,
-      "input_token_samples": 0,
-      "mean_actions": 14.158273381294965,
-      "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 36.330935251798564,
-      "mean_input_tokens": null,
-      "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 5403.838129496403,
-      "output_token_samples": 0,
-      "structurally_eligible": 278,
-      "success": 128,
-      "success_rate_known_outcomes": 0.460431654676259,
+      "input_token_samples": 14,
+      "mean_actions": 13.623287671232877,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 31.949838127408707,
+      "mean_events": 34.92465753424658,
+      "mean_input_tokens": 10815.142857142857,
+      "mean_output_tokens": 96.21428571428571,
+      "mean_supervised_assistant_characters": 5157.551369863014,
+      "output_token_samples": 14,
+      "structurally_eligible": 292,
+      "success": 142,
+      "success_rate_known_outcomes": 0.4863013698630137,
       "training_eligible": 278,
-      "trajectories": 278,
-      "unique_tasks": 228,
+      "trajectories": 292,
+      "unique_tasks": 242,
       "unknown": 0
     },
     "anthropic/claude-4-sonnet-thinking-on-10k": {
       "cost_samples": 0,
-      "duration_samples": 0,
+      "duration_samples": 11,
       "failure": 133,
-      "input_token_samples": 0,
-      "mean_actions": 13.611510791366907,
-      "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 34.989208633093526,
-      "mean_input_tokens": null,
-      "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 5727.251798561151,
-      "output_token_samples": 0,
-      "structurally_eligible": 278,
-      "success": 145,
-      "success_rate_known_outcomes": 0.5215827338129496,
+      "input_token_samples": 11,
+      "mean_actions": 13.204152249134948,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 34.31249661879106,
+      "mean_events": 33.9204152249135,
+      "mean_input_tokens": 11066.09090909091,
+      "mean_output_tokens": 439.27272727272725,
+      "mean_supervised_assistant_characters": 5519.892733564014,
+      "output_token_samples": 11,
+      "structurally_eligible": 289,
+      "success": 156,
+      "success_rate_known_outcomes": 0.5397923875432526,
       "training_eligible": 278,
-      "trajectories": 278,
-      "unique_tasks": 228,
+      "trajectories": 289,
+      "unique_tasks": 239,
       "unknown": 0
     },
     "anthropic/claude-4.5-sonnet-thinking-off": {
       "cost_samples": 0,
-      "duration_samples": 0,
+      "duration_samples": 2,
       "failure": 141,
-      "input_token_samples": 0,
-      "mean_actions": 15.31294964028777,
-      "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 42.37410071942446,
-      "mean_input_tokens": null,
-      "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 5430.76618705036,
-      "output_token_samples": 0,
-      "structurally_eligible": 278,
-      "success": 137,
-      "success_rate_known_outcomes": 0.49280575539568344,
+      "input_token_samples": 2,
+      "mean_actions": 15.217857142857143,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 15.838354706764221,
+      "mean_events": 42.107142857142854,
+      "mean_input_tokens": 3047.0,
+      "mean_output_tokens": 70.5,
+      "mean_supervised_assistant_characters": 5393.971428571429,
+      "output_token_samples": 2,
+      "structurally_eligible": 280,
+      "success": 139,
+      "success_rate_known_outcomes": 0.49642857142857144,
       "training_eligible": 278,
-      "trajectories": 278,
-      "unique_tasks": 228,
+      "trajectories": 280,
+      "unique_tasks": 230,
       "unknown": 0
     },
     "anthropic/claude-4.5-sonnet-thinking-on-10k": {
       "cost_samples": 0,
-      "duration_samples": 0,
+      "duration_samples": 90,
       "failure": 126,
-      "input_token_samples": 0,
-      "mean_actions": 13.507194244604317,
-      "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 39.55755395683453,
-      "mean_input_tokens": null,
-      "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 5573.884892086331,
-      "output_token_samples": 0,
-      "structurally_eligible": 278,
-      "success": 152,
-      "success_rate_known_outcomes": 0.5467625899280576,
+      "input_token_samples": 90,
+      "mean_actions": 10.815217391304348,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 26.105281273523968,
+      "mean_events": 31.39945652173913,
+      "mean_input_tokens": 8626.466666666667,
+      "mean_output_tokens": 416.1222222222222,
+      "mean_supervised_assistant_characters": 4272.513586956522,
+      "output_token_samples": 90,
+      "structurally_eligible": 368,
+      "success": 242,
+      "success_rate_known_outcomes": 0.657608695652174,
       "training_eligible": 278,
-      "trajectories": 278,
-      "unique_tasks": 228,
+      "trajectories": 368,
+      "unique_tasks": 318,
       "unknown": 0
     },
     "claude-3-5-sonnet-20241022": {
@@ -947,6 +1073,27 @@
       "unique_tasks": 83,
       "unknown": 0
     },
+    "claude-opus-5-5": {
+      "cost_samples": 0,
+      "duration_samples": 358,
+      "failure": 31,
+      "input_token_samples": 0,
+      "mean_actions": 116.77932960893855,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 2574.9733340706175,
+      "mean_events": 228.56424581005587,
+      "mean_input_tokens": null,
+      "mean_output_tokens": null,
+      "mean_supervised_assistant_characters": 3885.7206703910615,
+      "output_token_samples": 0,
+      "structurally_eligible": 54,
+      "success": 326,
+      "success_rate_known_outcomes": 0.9131652661064426,
+      "training_eligible": 0,
+      "trajectories": 358,
+      "unique_tasks": 20,
+      "unknown": 1
+    },
     "claude-sonnet-4-5-20250929@anthropic": {
       "cost_samples": 1696,
       "duration_samples": 1696,
@@ -1012,44 +1159,44 @@
     },
     "deepseek-ai/DeepSeek-V3.2-Exp-thinking-off": {
       "cost_samples": 0,
-      "duration_samples": 0,
-      "failure": 151,
-      "input_token_samples": 0,
-      "mean_actions": 15.356115107913668,
-      "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 40.989208633093526,
-      "mean_input_tokens": null,
-      "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 5493.4604316546765,
-      "output_token_samples": 0,
-      "structurally_eligible": 278,
-      "success": 127,
-      "success_rate_known_outcomes": 0.4568345323741007,
+      "duration_samples": 591,
+      "failure": 203,
+      "input_token_samples": 591,
+      "mean_actions": 7.273878020713464,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 34.15504535361998,
+      "mean_events": 18.719217491369392,
+      "mean_input_tokens": 34038.857868020306,
+      "mean_output_tokens": 116.1015228426396,
+      "mean_supervised_assistant_characters": 1938.783659378596,
+      "output_token_samples": 591,
+      "structurally_eligible": 869,
+      "success": 666,
+      "success_rate_known_outcomes": 0.766398158803222,
       "training_eligible": 278,
-      "trajectories": 278,
-      "unique_tasks": 228,
+      "trajectories": 869,
+      "unique_tasks": 819,
       "unknown": 0
     },
     "deepseek-ai/DeepSeek-V3.2-Exp-thinking-on": {
       "cost_samples": 0,
-      "duration_samples": 0,
-      "failure": 154,
-      "input_token_samples": 0,
-      "mean_actions": 14.676258992805755,
-      "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 39.856115107913666,
-      "mean_input_tokens": null,
-      "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 4574.9604316546765,
-      "output_token_samples": 0,
-      "structurally_eligible": 278,
-      "success": 124,
-      "success_rate_known_outcomes": 0.4460431654676259,
+      "duration_samples": 442,
+      "failure": 201,
+      "input_token_samples": 442,
+      "mean_actions": 7.963888888888889,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 185.8142748116368,
+      "mean_events": 20.670833333333334,
+      "mean_input_tokens": 41205.656108597286,
+      "mean_output_tokens": 4360.248868778281,
+      "mean_supervised_assistant_characters": 1894.2041666666667,
+      "output_token_samples": 442,
+      "structurally_eligible": 720,
+      "success": 519,
+      "success_rate_known_outcomes": 0.7208333333333333,
       "training_eligible": 278,
-      "trajectories": 278,
-      "unique_tasks": 228,
+      "trajectories": 720,
+      "unique_tasks": 670,
       "unknown": 0
     },
     "deepseek-v3.2@deepseek": {
@@ -1220,6 +1367,27 @@
       "unique_tasks": 89,
       "unknown": 0
     },
+    "glm-5": {
+      "cost_samples": 0,
+      "duration_samples": 0,
+      "failure": 91,
+      "input_token_samples": 0,
+      "mean_actions": 65.12,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": null,
+      "mean_events": 134.43,
+      "mean_input_tokens": null,
+      "mean_output_tokens": null,
+      "mean_supervised_assistant_characters": 504.18333333333334,
+      "output_token_samples": 0,
+      "structurally_eligible": 300,
+      "success": 209,
+      "success_rate_known_outcomes": 0.6966666666666667,
+      "training_eligible": 0,
+      "trajectories": 300,
+      "unique_tasks": 300,
+      "unknown": 0
+    },
     "glm-5-high": {
       "cost_samples": 0,
       "duration_samples": 0,
@@ -1264,65 +1432,65 @@
     },
     "google/gemini-2.5-flash-thinking-off": {
       "cost_samples": 0,
-      "duration_samples": 0,
-      "failure": 168,
-      "input_token_samples": 0,
-      "mean_actions": 11.456834532374101,
-      "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 25.154676258992804,
-      "mean_input_tokens": null,
-      "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 1610.4496402877699,
-      "output_token_samples": 0,
-      "structurally_eligible": 278,
-      "success": 110,
-      "success_rate_known_outcomes": 0.39568345323741005,
+      "duration_samples": 276,
+      "failure": 275,
+      "input_token_samples": 276,
+      "mean_actions": 7.574007220216607,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 81.5313858182534,
+      "mean_events": 16.850180505415164,
+      "mean_input_tokens": 67858.97826086957,
+      "mean_output_tokens": 150.5072463768116,
+      "mean_supervised_assistant_characters": 963.5920577617328,
+      "output_token_samples": 276,
+      "structurally_eligible": 554,
+      "success": 279,
+      "success_rate_known_outcomes": 0.5036101083032491,
       "training_eligible": 278,
-      "trajectories": 278,
-      "unique_tasks": 228,
+      "trajectories": 554,
+      "unique_tasks": 504,
       "unknown": 0
     },
     "google/gemini-2.5-flash-thinking-on": {
       "cost_samples": 0,
-      "duration_samples": 0,
-      "failure": 176,
-      "input_token_samples": 0,
-      "mean_actions": 15.902877697841726,
-      "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 39.02517985611511,
-      "mean_input_tokens": null,
-      "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 2495.3920863309354,
-      "output_token_samples": 0,
-      "structurally_eligible": 278,
-      "success": 102,
-      "success_rate_known_outcomes": 0.3669064748201439,
+      "duration_samples": 261,
+      "failure": 226,
+      "input_token_samples": 261,
+      "mean_actions": 9.890538033395176,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 75.96532614112357,
+      "mean_events": 24.024118738404454,
+      "mean_input_tokens": 52914.06896551724,
+      "mean_output_tokens": 117.79693486590038,
+      "mean_supervised_assistant_characters": 1400.8923933209649,
+      "output_token_samples": 261,
+      "structurally_eligible": 539,
+      "success": 313,
+      "success_rate_known_outcomes": 0.5807050092764379,
       "training_eligible": 278,
-      "trajectories": 278,
-      "unique_tasks": 228,
+      "trajectories": 539,
+      "unique_tasks": 489,
       "unknown": 0
     },
     "google/gemini-2.5-pro-thinking-on": {
       "cost_samples": 0,
-      "duration_samples": 0,
-      "failure": 165,
-      "input_token_samples": 0,
-      "mean_actions": 15.02158273381295,
-      "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 37.53237410071942,
-      "mean_input_tokens": null,
-      "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 2018.86690647482,
-      "output_token_samples": 0,
-      "structurally_eligible": 278,
-      "success": 113,
-      "success_rate_known_outcomes": 0.4064748201438849,
+      "duration_samples": 539,
+      "failure": 225,
+      "input_token_samples": 539,
+      "mean_actions": 7.283965728274174,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 96.22210228155627,
+      "mean_events": 17.95593635250918,
+      "mean_input_tokens": 67749.84601113173,
+      "mean_output_tokens": 120.8961038961039,
+      "mean_supervised_assistant_characters": 879.0599755201958,
+      "output_token_samples": 539,
+      "structurally_eligible": 817,
+      "success": 592,
+      "success_rate_known_outcomes": 0.7246022031823746,
       "training_eligible": 278,
-      "trajectories": 278,
-      "unique_tasks": 228,
+      "trajectories": 817,
+      "unique_tasks": 767,
       "unknown": 0
     },
     "gpt-4o-2024-08-06": {
@@ -1577,6 +1745,27 @@
       "unique_tasks": 89,
       "unknown": 0
     },
+    "gpt-6-astra": {
+      "cost_samples": 0,
+      "duration_samples": 360,
+      "failure": 48,
+      "input_token_samples": 0,
+      "mean_actions": 61.99444444444445,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 818.4848576635222,
+      "mean_events": 116.63055555555556,
+      "mean_input_tokens": null,
+      "mean_output_tokens": null,
+      "mean_supervised_assistant_characters": 2382.713888888889,
+      "output_token_samples": 0,
+      "structurally_eligible": 0,
+      "success": 312,
+      "success_rate_known_outcomes": 0.8666666666666667,
+      "training_eligible": 0,
+      "trajectories": 360,
+      "unique_tasks": 20,
+      "unknown": 0
+    },
     "gpt-oss-120b": {
       "cost_samples": 0,
       "duration_samples": 5063,
@@ -1682,6 +1871,27 @@
       "unique_tasks": 89,
       "unknown": 0
     },
+    "minimax-m2.5": {
+      "cost_samples": 0,
+      "duration_samples": 0,
+      "failure": 79,
+      "input_token_samples": 0,
+      "mean_actions": 73.76923076923077,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": null,
+      "mean_events": 148.62541806020067,
+      "mean_input_tokens": null,
+      "mean_output_tokens": null,
+      "mean_supervised_assistant_characters": 1896.1137123745818,
+      "output_token_samples": 0,
+      "structurally_eligible": 299,
+      "success": 220,
+      "success_rate_known_outcomes": 0.7357859531772575,
+      "training_eligible": 0,
+      "trajectories": 299,
+      "unique_tasks": 299,
+      "unknown": 0
+    },
     "minimax-m2.5-high": {
       "cost_samples": 0,
       "duration_samples": 0,
@@ -1768,149 +1978,149 @@
     },
     "openai/gpt-4.1": {
       "cost_samples": 0,
-      "duration_samples": 0,
-      "failure": 153,
-      "input_token_samples": 0,
-      "mean_actions": 14.446043165467627,
-      "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 40.118705035971225,
-      "mean_input_tokens": null,
-      "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 3550.039568345324,
-      "output_token_samples": 0,
-      "structurally_eligible": 278,
-      "success": 125,
-      "success_rate_known_outcomes": 0.44964028776978415,
+      "duration_samples": 252,
+      "failure": 221,
+      "input_token_samples": 252,
+      "mean_actions": 10.139622641509433,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 50.07472106294026,
+      "mean_events": 26.67924528301887,
+      "mean_input_tokens": 174260.43253968254,
+      "mean_output_tokens": 190.86111111111111,
+      "mean_supervised_assistant_characters": 1996.2811320754718,
+      "output_token_samples": 252,
+      "structurally_eligible": 530,
+      "success": 309,
+      "success_rate_known_outcomes": 0.5830188679245283,
       "training_eligible": 278,
-      "trajectories": 278,
-      "unique_tasks": 228,
+      "trajectories": 530,
+      "unique_tasks": 480,
       "unknown": 0
     },
     "openai/gpt-4.1-mini": {
       "cost_samples": 0,
-      "duration_samples": 0,
-      "failure": 158,
-      "input_token_samples": 0,
-      "mean_actions": 19.776978417266186,
-      "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 50.32014388489208,
-      "mean_input_tokens": null,
-      "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 3496.9460431654675,
-      "output_token_samples": 0,
-      "structurally_eligible": 278,
-      "success": 120,
-      "success_rate_known_outcomes": 0.4316546762589928,
+      "duration_samples": 359,
+      "failure": 290,
+      "input_token_samples": 359,
+      "mean_actions": 10.513343799058084,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 26.014673313390578,
+      "mean_events": 26.28885400313972,
+      "mean_input_tokens": 32549.963788300836,
+      "mean_output_tokens": 116.2033426183844,
+      "mean_supervised_assistant_characters": 1667.3124018838305,
+      "output_token_samples": 359,
+      "structurally_eligible": 637,
+      "success": 347,
+      "success_rate_known_outcomes": 0.5447409733124019,
       "training_eligible": 278,
-      "trajectories": 278,
-      "unique_tasks": 228,
+      "trajectories": 637,
+      "unique_tasks": 587,
       "unknown": 0
     },
     "openai/gpt-4.1-nano": {
       "cost_samples": 0,
-      "duration_samples": 0,
+      "duration_samples": 5,
       "failure": 238,
-      "input_token_samples": 0,
-      "mean_actions": 33.71223021582734,
-      "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 77.18705035971223,
-      "mean_input_tokens": null,
-      "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 4114.230215827338,
-      "output_token_samples": 0,
-      "structurally_eligible": 278,
-      "success": 40,
-      "success_rate_known_outcomes": 0.14388489208633093,
+      "input_token_samples": 5,
+      "mean_actions": 33.15194346289753,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 13.41379680633545,
+      "mean_events": 75.91872791519435,
+      "mean_input_tokens": 6789.4,
+      "mean_output_tokens": 76.4,
+      "mean_supervised_assistant_characters": 4047.773851590106,
+      "output_token_samples": 5,
+      "structurally_eligible": 283,
+      "success": 45,
+      "success_rate_known_outcomes": 0.15901060070671377,
       "training_eligible": 278,
-      "trajectories": 278,
-      "unique_tasks": 228,
+      "trajectories": 283,
+      "unique_tasks": 233,
       "unknown": 0
     },
     "openai/gpt-4o-20240806": {
       "cost_samples": 0,
-      "duration_samples": 0,
-      "failure": 170,
-      "input_token_samples": 0,
-      "mean_actions": 17.758992805755394,
-      "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 48.302158273381295,
-      "mean_input_tokens": null,
-      "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 3614.1582733812947,
-      "output_token_samples": 0,
-      "structurally_eligible": 278,
-      "success": 108,
-      "success_rate_known_outcomes": 0.38848920863309355,
+      "duration_samples": 208,
+      "failure": 237,
+      "input_token_samples": 208,
+      "mean_actions": 11.41358024691358,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 25.662573208029453,
+      "mean_events": 30.584362139917694,
+      "mean_input_tokens": 23774.322115384617,
+      "mean_output_tokens": 98.67788461538461,
+      "mean_supervised_assistant_characters": 2174.2098765432097,
+      "output_token_samples": 208,
+      "structurally_eligible": 486,
+      "success": 249,
+      "success_rate_known_outcomes": 0.5123456790123457,
       "training_eligible": 278,
-      "trajectories": 278,
-      "unique_tasks": 228,
+      "trajectories": 486,
+      "unique_tasks": 436,
       "unknown": 0
     },
     "openai/gpt-4o-mini": {
       "cost_samples": 0,
-      "duration_samples": 0,
-      "failure": 205,
-      "input_token_samples": 0,
-      "mean_actions": 45.21223021582734,
-      "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 93.01079136690647,
-      "mean_input_tokens": null,
-      "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 1972.478417266187,
-      "output_token_samples": 0,
-      "structurally_eligible": 278,
-      "success": 73,
-      "success_rate_known_outcomes": 0.26258992805755393,
+      "duration_samples": 69,
+      "failure": 264,
+      "input_token_samples": 69,
+      "mean_actions": 36.86743515850144,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 27.863784651825394,
+      "mean_events": 76.02593659942363,
+      "mean_input_tokens": 17608.391304347828,
+      "mean_output_tokens": 116.28985507246377,
+      "mean_supervised_assistant_characters": 1624.8876080691643,
+      "output_token_samples": 69,
+      "structurally_eligible": 347,
+      "success": 83,
+      "success_rate_known_outcomes": 0.23919308357348704,
       "training_eligible": 278,
-      "trajectories": 278,
-      "unique_tasks": 228,
+      "trajectories": 347,
+      "unique_tasks": 297,
       "unknown": 0
     },
     "openai/gpt-5": {
       "cost_samples": 0,
-      "duration_samples": 0,
-      "failure": 138,
-      "input_token_samples": 0,
-      "mean_actions": 12.305755395683454,
-      "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 40.26618705035971,
-      "mean_input_tokens": null,
-      "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 4689.446043165468,
-      "output_token_samples": 0,
-      "structurally_eligible": 278,
-      "success": 140,
-      "success_rate_known_outcomes": 0.5035971223021583,
+      "duration_samples": 139,
+      "failure": 139,
+      "input_token_samples": 139,
+      "mean_actions": 9.045563549160672,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 54.889759729234434,
+      "mean_events": 29.48201438848921,
+      "mean_input_tokens": 12583.453237410073,
+      "mean_output_tokens": 2372.640287769784,
+      "mean_supervised_assistant_characters": 3189.1079136690646,
+      "output_token_samples": 139,
+      "structurally_eligible": 417,
+      "success": 278,
+      "success_rate_known_outcomes": 0.6666666666666666,
       "training_eligible": 278,
-      "trajectories": 278,
-      "unique_tasks": 228,
+      "trajectories": 417,
+      "unique_tasks": 367,
       "unknown": 0
     },
     "openai/gpt-5-nano": {
       "cost_samples": 0,
-      "duration_samples": 0,
-      "failure": 183,
-      "input_token_samples": 0,
-      "mean_actions": 24.07913669064748,
-      "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 61.83453237410072,
-      "mean_input_tokens": null,
-      "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 8785.273381294965,
-      "output_token_samples": 0,
-      "structurally_eligible": 278,
-      "success": 95,
-      "success_rate_known_outcomes": 0.34172661870503596,
+      "duration_samples": 370,
+      "failure": 209,
+      "input_token_samples": 370,
+      "mean_actions": 17.387345679012345,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 199.7940233707428,
+      "mean_events": 42.282407407407405,
+      "mean_input_tokens": 131090.58378378378,
+      "mean_output_tokens": 10394.559459459459,
+      "mean_supervised_assistant_characters": 3927.722222222222,
+      "output_token_samples": 370,
+      "structurally_eligible": 648,
+      "success": 439,
+      "success_rate_known_outcomes": 0.6774691358024691,
       "training_eligible": 278,
-      "trajectories": 278,
-      "unique_tasks": 228,
+      "trajectories": 648,
+      "unique_tasks": 598,
       "unknown": 0
     },
     "openai/gpt-oss-120b@together_ai": {
@@ -1957,170 +2167,212 @@
     },
     "openai/o3-high": {
       "cost_samples": 0,
-      "duration_samples": 0,
-      "failure": 136,
-      "input_token_samples": 0,
-      "mean_actions": 15.010791366906474,
-      "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 42.780575539568346,
-      "mean_input_tokens": null,
-      "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 3309.9172661870502,
-      "output_token_samples": 0,
-      "structurally_eligible": 278,
-      "success": 142,
-      "success_rate_known_outcomes": 0.5107913669064749,
+      "duration_samples": 325,
+      "failure": 196,
+      "input_token_samples": 325,
+      "mean_actions": 9.38971807628524,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 63.803641966306245,
+      "mean_events": 25.29021558872305,
+      "mean_input_tokens": 49568.904615384614,
+      "mean_output_tokens": 1869.5384615384614,
+      "mean_supervised_assistant_characters": 1636.5837479270315,
+      "output_token_samples": 325,
+      "structurally_eligible": 603,
+      "success": 407,
+      "success_rate_known_outcomes": 0.6749585406301825,
       "training_eligible": 278,
-      "trajectories": 278,
-      "unique_tasks": 228,
+      "trajectories": 603,
+      "unique_tasks": 553,
       "unknown": 0
     },
     "openai/o4-mini-high": {
       "cost_samples": 0,
-      "duration_samples": 0,
-      "failure": 162,
-      "input_token_samples": 0,
-      "mean_actions": 14.733812949640289,
-      "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 39.410071942446045,
-      "mean_input_tokens": null,
-      "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 2741.8057553956833,
-      "output_token_samples": 0,
-      "structurally_eligible": 278,
-      "success": 116,
-      "success_rate_known_outcomes": 0.4172661870503597,
+      "duration_samples": 253,
+      "failure": 215,
+      "input_token_samples": 253,
+      "mean_actions": 9.800376647834275,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 86.75983483235355,
+      "mean_events": 25.369114877589453,
+      "mean_input_tokens": 77490.51778656126,
+      "mean_output_tokens": 5419.418972332016,
+      "mean_supervised_assistant_characters": 1520.2052730696798,
+      "output_token_samples": 253,
+      "structurally_eligible": 531,
+      "success": 316,
+      "success_rate_known_outcomes": 0.5951035781544256,
       "training_eligible": 278,
-      "trajectories": 278,
-      "unique_tasks": 228,
+      "trajectories": 531,
+      "unique_tasks": 481,
       "unknown": 0
     },
     "togetherai/Qwen/Qwen3-235B-A22B-FP8": {
       "cost_samples": 0,
-      "duration_samples": 0,
-      "failure": 181,
-      "input_token_samples": 0,
-      "mean_actions": 14.18705035971223,
-      "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 35.55755395683453,
-      "mean_input_tokens": null,
-      "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 3833.568345323741,
-      "output_token_samples": 0,
-      "structurally_eligible": 278,
-      "success": 97,
-      "success_rate_known_outcomes": 0.3489208633093525,
+      "duration_samples": 175,
+      "failure": 296,
+      "input_token_samples": 175,
+      "mean_actions": 9.827814569536423,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 109.05009804044451,
+      "mean_events": 24.631346578366447,
+      "mean_input_tokens": 9332.565714285714,
+      "mean_output_tokens": 2002.7085714285715,
+      "mean_supervised_assistant_characters": 2462.534216335541,
+      "output_token_samples": 175,
+      "structurally_eligible": 453,
+      "success": 157,
+      "success_rate_known_outcomes": 0.3465783664459161,
       "training_eligible": 278,
-      "trajectories": 278,
-      "unique_tasks": 228,
+      "trajectories": 453,
+      "unique_tasks": 403,
       "unknown": 0
     },
     "togetherai/Qwen/Qwen3-235B-A22B-Instruct-2507-FP8": {
       "cost_samples": 0,
-      "duration_samples": 0,
-      "failure": 154,
-      "input_token_samples": 0,
-      "mean_actions": 16.37410071942446,
-      "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 41.82374100719424,
-      "mean_input_tokens": null,
-      "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 3558.928057553957,
-      "output_token_samples": 0,
-      "structurally_eligible": 278,
-      "success": 124,
-      "success_rate_known_outcomes": 0.4460431654676259,
+      "duration_samples": 155,
+      "failure": 219,
+      "input_token_samples": 155,
+      "mean_actions": 12.517321016166282,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 143.54535742113669,
+      "mean_events": 31.48267898383372,
+      "mean_input_tokens": 158738.4193548387,
+      "mean_output_tokens": 248.96774193548387,
+      "mean_supervised_assistant_characters": 2400.1154734411084,
+      "output_token_samples": 155,
+      "structurally_eligible": 433,
+      "success": 214,
+      "success_rate_known_outcomes": 0.4942263279445728,
       "training_eligible": 278,
-      "trajectories": 278,
-      "unique_tasks": 228,
+      "trajectories": 433,
+      "unique_tasks": 383,
       "unknown": 0
     },
     "togetherai/Qwen/Qwen3-235B-A22B-Thinking-2507-FP8": {
       "cost_samples": 0,
-      "duration_samples": 0,
-      "failure": 162,
-      "input_token_samples": 0,
-      "mean_actions": 15.913669064748202,
-      "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 40.17985611510792,
-      "mean_input_tokens": null,
-      "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 7153.723021582734,
-      "output_token_samples": 0,
-      "structurally_eligible": 278,
-      "success": 116,
-      "success_rate_known_outcomes": 0.4172661870503597,
+      "duration_samples": 235,
+      "failure": 244,
+      "input_token_samples": 235,
+      "mean_actions": 10.393762183235868,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 159.34282206677375,
+      "mean_events": 25.992202729044834,
+      "mean_input_tokens": 53863.74042553191,
+      "mean_output_tokens": 3920.868085106383,
+      "mean_supervised_assistant_characters": 4034.7153996101365,
+      "output_token_samples": 235,
+      "structurally_eligible": 513,
+      "success": 269,
+      "success_rate_known_outcomes": 0.5243664717348928,
       "training_eligible": 278,
-      "trajectories": 278,
-      "unique_tasks": 228,
+      "trajectories": 513,
+      "unique_tasks": 463,
       "unknown": 0
     },
     "togetherai/Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8": {
       "cost_samples": 0,
-      "duration_samples": 0,
-      "failure": 218,
-      "input_token_samples": 0,
-      "mean_actions": 14.129496402877697,
-      "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 38.697841726618705,
-      "mean_input_tokens": null,
-      "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 3825.9172661870502,
-      "output_token_samples": 0,
-      "structurally_eligible": 278,
-      "success": 60,
-      "success_rate_known_outcomes": 0.2158273381294964,
+      "duration_samples": 472,
+      "failure": 318,
+      "input_token_samples": 472,
+      "mean_actions": 7.7,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 78.5973685515129,
+      "mean_events": 19.902666666666665,
+      "mean_input_tokens": 69685.35381355933,
+      "mean_output_tokens": 130.57415254237287,
+      "mean_supervised_assistant_characters": 1598.5626666666667,
+      "output_token_samples": 472,
+      "structurally_eligible": 750,
+      "success": 432,
+      "success_rate_known_outcomes": 0.576,
       "training_eligible": 278,
-      "trajectories": 278,
-      "unique_tasks": 228,
+      "trajectories": 750,
+      "unique_tasks": 700,
+      "unknown": 0
+    },
+    "togetherai/deepseek-ai/DeepSeek-R1-0528": {
+      "cost_samples": 0,
+      "duration_samples": 295,
+      "failure": 28,
+      "input_token_samples": 295,
+      "mean_actions": 3.7322033898305085,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 281.9603391105846,
+      "mean_events": 8.6,
+      "mean_input_tokens": 20472.057627118644,
+      "mean_output_tokens": 5724.986440677966,
+      "mean_supervised_assistant_characters": 230.93559322033897,
+      "output_token_samples": 295,
+      "structurally_eligible": 295,
+      "success": 267,
+      "success_rate_known_outcomes": 0.9050847457627119,
+      "training_eligible": 0,
+      "trajectories": 295,
+      "unique_tasks": 295,
+      "unknown": 0
+    },
+    "togetherai/deepseek-ai/DeepSeek-V3-0324": {
+      "cost_samples": 0,
+      "duration_samples": 202,
+      "failure": 148,
+      "input_token_samples": 202,
+      "mean_actions": 3.198019801980198,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 35.992268392355136,
+      "mean_events": 7.564356435643564,
+      "mean_input_tokens": 26407.465346534653,
+      "mean_output_tokens": 140.9009900990099,
+      "mean_supervised_assistant_characters": 310.480198019802,
+      "output_token_samples": 202,
+      "structurally_eligible": 202,
+      "success": 54,
+      "success_rate_known_outcomes": 0.26732673267326734,
+      "training_eligible": 0,
+      "trajectories": 202,
+      "unique_tasks": 202,
       "unknown": 0
     },
     "togetherai/moonshotai/Kimi-K2-Instruct": {
       "cost_samples": 0,
-      "duration_samples": 0,
+      "duration_samples": 46,
       "failure": 141,
-      "input_token_samples": 0,
-      "mean_actions": 14.881294964028777,
-      "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 40.35251798561151,
-      "mean_input_tokens": null,
-      "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 5261.179856115108,
-      "output_token_samples": 0,
-      "structurally_eligible": 276,
-      "success": 137,
-      "success_rate_known_outcomes": 0.49280575539568344,
+      "input_token_samples": 46,
+      "mean_actions": 13.182098765432098,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 31.242617000704225,
+      "mean_events": 35.67901234567901,
+      "mean_input_tokens": 10820.826086956522,
+      "mean_output_tokens": 104.93478260869566,
+      "mean_supervised_assistant_characters": 4562.62962962963,
+      "output_token_samples": 46,
+      "structurally_eligible": 322,
+      "success": 183,
+      "success_rate_known_outcomes": 0.5648148148148148,
       "training_eligible": 276,
-      "trajectories": 278,
-      "unique_tasks": 228,
+      "trajectories": 324,
+      "unique_tasks": 274,
       "unknown": 0
     },
     "togetherai/moonshotai/Kimi-K2-Instruct-0905": {
       "cost_samples": 0,
-      "duration_samples": 0,
-      "failure": 139,
-      "input_token_samples": 0,
-      "mean_actions": 16.007194244604317,
-      "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 42.42446043165467,
-      "mean_input_tokens": null,
-      "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 5449.546762589928,
-      "output_token_samples": 0,
-      "structurally_eligible": 278,
-      "success": 139,
-      "success_rate_known_outcomes": 0.5,
+      "duration_samples": 73,
+      "failure": 148,
+      "input_token_samples": 73,
+      "mean_actions": 13.652421652421653,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 66.55511462851746,
+      "mean_events": 35.8034188034188,
+      "mean_input_tokens": 62322.109589041094,
+      "mean_output_tokens": 155.8904109589041,
+      "mean_supervised_assistant_characters": 4376.541310541311,
+      "output_token_samples": 73,
+      "structurally_eligible": 351,
+      "success": 203,
+      "success_rate_known_outcomes": 0.5783475783475783,
       "training_eligible": 278,
-      "trajectories": 278,
-      "unique_tasks": 228,
+      "trajectories": 351,
+      "unique_tasks": 301,
       "unknown": 0
     },
     "togetherai/openai/gpt-oss-120b": {
@@ -2190,88 +2442,88 @@
   "by_outcome": {
     "failure": {
       "cost_samples": 23212,
-      "duration_samples": 27064,
-      "failure": 79973,
-      "input_token_samples": 21454,
-      "mean_actions": 45.62608630412765,
+      "duration_samples": 28472,
+      "failure": 81551,
+      "input_token_samples": 22783,
+      "mean_actions": 45.076308077154174,
       "mean_cost_source_native_units": 75.64188306777397,
-      "mean_duration_seconds": 701.5190378383018,
-      "mean_events": 94.278981656309,
-      "mean_input_tokens": 1156962.542835835,
-      "mean_output_tokens": 21745.191386221682,
-      "mean_supervised_assistant_characters": 12323.303715003814,
-      "output_token_samples": 21454,
-      "structurally_eligible": 73054,
+      "mean_duration_seconds": 676.8694757494687,
+      "mean_events": 93.14106510036665,
+      "mean_input_tokens": 1096277.9173506561,
+      "mean_output_tokens": 20592.015362331564,
+      "mean_supervised_assistant_characters": 12095.395887236207,
+      "output_token_samples": 22783,
+      "structurally_eligible": 74559,
       "success": 0,
       "success_rate_known_outcomes": 0.0,
       "training_eligible": 51532,
-      "trajectories": 79973,
-      "unique_tasks": 28927,
+      "trajectories": 81551,
+      "unique_tasks": 29227,
       "unknown": 0
     },
     "success": {
       "cost_samples": 11051,
-      "duration_samples": 12262,
+      "duration_samples": 17731,
       "failure": 0,
-      "input_token_samples": 9161,
-      "mean_actions": 36.86015544423162,
+      "input_token_samples": 13992,
+      "mean_actions": 34.98200817357212,
       "mean_cost_source_native_units": 31.06763293005613,
-      "mean_duration_seconds": 410.78903144437595,
-      "mean_events": 76.32702224841609,
-      "mean_input_tokens": 382399.21984499507,
-      "mean_output_tokens": 11283.977950005457,
-      "mean_supervised_assistant_characters": 7551.789303079417,
-      "output_token_samples": 9161,
-      "structurally_eligible": 50793,
-      "success": 54296,
+      "mean_duration_seconds": 368.821970174292,
+      "mean_events": 72.37669867428647,
+      "mean_input_tokens": 264175.1624499714,
+      "mean_output_tokens": 8037.330474556889,
+      "mean_supervised_assistant_characters": 6873.556484034953,
+      "output_token_samples": 13992,
+      "structurally_eligible": 56101,
+      "success": 60194,
       "success_rate_known_outcomes": 1.0,
       "training_eligible": 38524,
-      "trajectories": 54296,
-      "unique_tasks": 23353,
+      "trajectories": 60194,
+      "unique_tasks": 24260,
       "unknown": 0
     },
     "unknown": {
       "cost_samples": 0,
-      "duration_samples": 0,
+      "duration_samples": 1,
       "failure": 0,
       "input_token_samples": 0,
-      "mean_actions": 16.3787424437146,
+      "mean_actions": 6.799046010347282,
       "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 32.805511923461374,
+      "mean_duration_seconds": 4853.42680156,
+      "mean_events": 14.898471279387238,
       "mean_input_tokens": null,
       "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 10051.800371269455,
+      "mean_supervised_assistant_characters": 3869.9040804835813,
       "output_token_samples": 0,
-      "structurally_eligible": 20152,
+      "structurally_eligible": 93270,
       "success": 0,
       "success_rate_known_outcomes": null,
-      "training_eligible": 20152,
-      "trajectories": 21009,
-      "unique_tasks": 18796,
-      "unknown": 21009
+      "training_eligible": 93270,
+      "trajectories": 94131,
+      "unique_tasks": 91918,
+      "unknown": 94131
     }
   },
   "by_quality_tier": {
     "3": {
       "cost_samples": 7668,
-      "duration_samples": 8879,
+      "duration_samples": 13758,
       "failure": 0,
-      "input_token_samples": 6699,
-      "mean_actions": 35.69987645264661,
+      "input_token_samples": 11530,
+      "mean_actions": 33.29231308002101,
       "mean_cost_source_native_units": 23.312137899828038,
-      "mean_duration_seconds": 387.87317761634836,
-      "mean_events": 75.29464113354696,
-      "mean_input_tokens": 242325.22913867742,
-      "mean_output_tokens": 11504.635617256306,
-      "mean_supervised_assistant_characters": 7296.183294081309,
-      "output_token_samples": 6699,
-      "structurally_eligible": 51802,
-      "success": 50793,
+      "mean_duration_seconds": 290.12420282631973,
+      "mean_events": 70.25149711083873,
+      "mean_input_tokens": 157546.94969644406,
+      "mean_output_tokens": 7472.277536860364,
+      "mean_supervised_assistant_characters": 6650.703256872702,
+      "output_token_samples": 11530,
+      "structurally_eligible": 57110,
+      "success": 56101,
       "success_rate_known_outcomes": 1.0,
       "training_eligible": 39533,
-      "trajectories": 51802,
-      "unique_tasks": 24316,
+      "trajectories": 57110,
+      "unique_tasks": 25219,
       "unknown": 1009
     },
     "4": {
@@ -2279,45 +2531,66 @@
       "duration_samples": 0,
       "failure": 0,
       "input_token_samples": 0,
-      "mean_actions": 17.524369221125216,
+      "mean_actions": 6.8409403756733616,
       "mean_cost_source_native_units": null,
       "mean_duration_seconds": null,
-      "mean_events": 35.04873844225043,
+      "mean_events": 14.997647976935,
       "mean_input_tokens": null,
       "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 10991.896829128142,
+      "mean_supervised_assistant_characters": 3939.9823652464206,
       "output_token_samples": 0,
-      "structurally_eligible": 19143,
+      "structurally_eligible": 92261,
       "success": 0,
       "success_rate_known_outcomes": null,
-      "training_eligible": 19143,
-      "trajectories": 19143,
-      "unique_tasks": 16936,
-      "unknown": 19143
+      "training_eligible": 92261,
+      "trajectories": 92261,
+      "unique_tasks": 90054,
+      "unknown": 92261
     },
     "5": {
       "cost_samples": 26595,
-      "duration_samples": 30447,
-      "failure": 79973,
-      "input_token_samples": 23916,
-      "mean_actions": 45.172364317645524,
+      "duration_samples": 32446,
+      "failure": 81551,
+      "input_token_samples": 25245,
+      "mean_actions": 44.9598289116236,
       "mean_cost_source_native_units": 72.20794615015373,
-      "mean_duration_seconds": 675.8984596042691,
-      "mean_events": 92.5128360191147,
-      "mean_input_tokens": 1116461.6548335843,
-      "mean_output_tokens": 20606.468640240844,
-      "mean_supervised_assistant_characters": 12075.546061446883,
-      "output_token_samples": 23916,
-      "structurally_eligible": 73054,
-      "success": 3503,
-      "success_rate_known_outcomes": 0.041964157362595234,
+      "mean_duration_seconds": 672.6479292059513,
+      "mean_events": 92.0066701346743,
+      "mean_input_tokens": 1063827.384987126,
+      "mean_output_tokens": 19625.70227767875,
+      "mean_supervised_assistant_characters": 11803.80387260852,
+      "output_token_samples": 25245,
+      "structurally_eligible": 74559,
+      "success": 4093,
+      "success_rate_known_outcomes": 0.04779085516790435,
       "training_eligible": 51532,
-      "trajectories": 84333,
-      "unique_tasks": 29808,
-      "unknown": 857
+      "trajectories": 86505,
+      "unique_tasks": 30120,
+      "unknown": 861
     }
   },
   "by_source": {
+    "agentsuite_bfcl_v4": {
+      "cost_samples": 0,
+      "duration_samples": 6160,
+      "failure": 1329,
+      "input_token_samples": 6160,
+      "mean_actions": 4.220454545454546,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 93.63021875290902,
+      "mean_events": 9.771266233766234,
+      "mean_input_tokens": 56526.788474025976,
+      "mean_output_tokens": 1901.3564935064935,
+      "mean_supervised_assistant_characters": 261.6818181818182,
+      "output_token_samples": 6160,
+      "structurally_eligible": 6160,
+      "success": 4831,
+      "success_rate_known_outcomes": 0.7842532467532467,
+      "training_eligible": 0,
+      "trajectories": 6160,
+      "unique_tasks": 665,
+      "unknown": 0
+    },
     "agentsuite_tau2": {
       "cost_samples": 0,
       "duration_samples": 0,
@@ -2360,6 +2633,48 @@
       "unique_tasks": 1266,
       "unknown": 0
     },
+    "delm_deepswe": {
+      "cost_samples": 0,
+      "duration_samples": 360,
+      "failure": 30,
+      "input_token_samples": 0,
+      "mean_actions": 90.43611111111112,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 1342.479669959725,
+      "mean_events": 175.45833333333334,
+      "mean_input_tokens": null,
+      "mean_output_tokens": null,
+      "mean_supervised_assistant_characters": 2983.5583333333334,
+      "output_token_samples": 0,
+      "structurally_eligible": 8,
+      "success": 330,
+      "success_rate_known_outcomes": 0.9166666666666666,
+      "training_eligible": 0,
+      "trajectories": 360,
+      "unique_tasks": 10,
+      "unknown": 0
+    },
+    "delm_terminalbench4": {
+      "cost_samples": 0,
+      "duration_samples": 358,
+      "failure": 49,
+      "input_token_samples": 0,
+      "mean_actions": 88.17877094972067,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 2048.051176454324,
+      "mean_events": 169.40782122905028,
+      "mean_input_tokens": null,
+      "mean_output_tokens": null,
+      "mean_supervised_assistant_characters": 3281.5195530726255,
+      "output_token_samples": 0,
+      "structurally_eligible": 46,
+      "success": 308,
+      "success_rate_known_outcomes": 0.8627450980392157,
+      "training_eligible": 0,
+      "trajectories": 358,
+      "unique_tasks": 10,
+      "unknown": 1
+    },
     "mind2web": {
       "cost_samples": 0,
       "duration_samples": 0,
@@ -2423,6 +2738,48 @@
       "unique_tasks": 50,
       "unknown": 0
     },
+    "swebench_multilingual_glm5": {
+      "cost_samples": 0,
+      "duration_samples": 0,
+      "failure": 91,
+      "input_token_samples": 0,
+      "mean_actions": 65.12,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": null,
+      "mean_events": 134.43,
+      "mean_input_tokens": null,
+      "mean_output_tokens": null,
+      "mean_supervised_assistant_characters": 504.18333333333334,
+      "output_token_samples": 0,
+      "structurally_eligible": 300,
+      "success": 209,
+      "success_rate_known_outcomes": 0.6966666666666667,
+      "training_eligible": 0,
+      "trajectories": 300,
+      "unique_tasks": 300,
+      "unknown": 0
+    },
+    "swebench_multilingual_minimax_m25": {
+      "cost_samples": 0,
+      "duration_samples": 0,
+      "failure": 79,
+      "input_token_samples": 0,
+      "mean_actions": 73.76923076923077,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": null,
+      "mean_events": 148.62541806020067,
+      "mean_input_tokens": null,
+      "mean_output_tokens": null,
+      "mean_supervised_assistant_characters": 1896.1137123745818,
+      "output_token_samples": 0,
+      "structurally_eligible": 299,
+      "success": 220,
+      "success_rate_known_outcomes": 0.7357859531772575,
+      "training_eligible": 0,
+      "trajectories": 299,
+      "unique_tasks": 299,
+      "unknown": 0
+    },
     "swebench_multilingual_trajectories": {
       "cost_samples": 0,
       "duration_samples": 0,
@@ -2590,29 +2947,71 @@
       "trajectories": 11,
       "unique_tasks": 11,
       "unknown": 0
+    },
+    "toolbench_full": {
+      "cost_samples": 0,
+      "duration_samples": 0,
+      "failure": 0,
+      "input_token_samples": 0,
+      "mean_actions": 4.0439135132178174,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": null,
+      "mean_events": 9.748088784343759,
+      "mean_input_tokens": null,
+      "mean_output_tokens": null,
+      "mean_supervised_assistant_characters": 2093.7298176994295,
+      "output_token_samples": 0,
+      "structurally_eligible": 73118,
+      "success": 0,
+      "success_rate_known_outcomes": null,
+      "training_eligible": 73118,
+      "trajectories": 73121,
+      "unique_tasks": 73121,
+      "unknown": 73121
     }
   },
   "by_task_family": {
+    "function-calling": {
+      "cost_samples": 0,
+      "duration_samples": 6160,
+      "failure": 1329,
+      "input_token_samples": 6160,
+      "mean_actions": 4.220454545454546,
+      "mean_cost_source_native_units": null,
+      "mean_duration_seconds": 93.63021875290902,
+      "mean_events": 9.771266233766234,
+      "mean_input_tokens": 56526.788474025976,
+      "mean_output_tokens": 1901.3564935064935,
+      "mean_supervised_assistant_characters": 261.6818181818182,
+      "output_token_samples": 6160,
+      "structurally_eligible": 6160,
+      "success": 4831,
+      "success_rate_known_outcomes": 0.7842532467532467,
+      "training_eligible": 0,
+      "trajectories": 6160,
+      "unique_tasks": 665,
+      "unknown": 0
+    },
     "software-engineering": {
       "cost_samples": 0,
-      "duration_samples": 0,
-      "failure": 47840,
+      "duration_samples": 718,
+      "failure": 48089,
       "input_token_samples": 0,
-      "mean_actions": 47.65324335586629,
+      "mean_actions": 48.14302979169754,
       "mean_cost_source_native_units": null,
-      "mean_duration_seconds": null,
-      "mean_events": 100.24062413186407,
+      "mean_duration_seconds": 1694.2827330865584,
+      "mean_events": 101.11355733163073,
       "mean_input_tokens": null,
       "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 8926.038452634502,
+      "mean_supervised_assistant_characters": 8825.837394110069,
       "output_token_samples": 0,
-      "structurally_eligible": 86361,
-      "success": 38552,
-      "success_rate_known_outcomes": 0.4462450226872859,
+      "structurally_eligible": 87014,
+      "success": 39619,
+      "success_rate_known_outcomes": 0.45171478086377526,
       "training_eligible": 81719,
-      "trajectories": 86392,
-      "unique_tasks": 47170,
-      "unknown": 0
+      "trajectories": 87709,
+      "unique_tasks": 47447,
+      "unknown": 1
     },
     "terminal-agent": {
       "cost_samples": 34263,
@@ -2661,21 +3060,21 @@
       "duration_samples": 0,
       "failure": 4,
       "input_token_samples": 0,
-      "mean_actions": 3.8181818181818183,
+      "mean_actions": 4.043879560247224,
       "mean_cost_source_native_units": null,
       "mean_duration_seconds": null,
-      "mean_events": 9.090909090909092,
+      "mean_events": 9.747989936006126,
       "mean_input_tokens": null,
       "mean_output_tokens": null,
-      "mean_supervised_assistant_characters": 834.3636363636364,
+      "mean_supervised_assistant_characters": 2093.5403927145435,
       "output_token_samples": 0,
-      "structurally_eligible": 0,
+      "structurally_eligible": 73118,
       "success": 7,
       "success_rate_known_outcomes": 0.6363636363636364,
-      "training_eligible": 0,
-      "trajectories": 11,
-      "unique_tasks": 11,
-      "unknown": 0
+      "training_eligible": 73118,
+      "trajectories": 73132,
+      "unique_tasks": 73132,
+      "unknown": 73121
     },
     "web-navigation": {
       "cost_samples": 0,
@@ -2720,10 +3119,10 @@
       "unknown": 0
     }
   },
-  "compatible_task_environments": 67660,
-  "evaluated_trajectories": 155278,
+  "compatible_task_environments": 141766,
+  "evaluated_trajectories": 235876,
   "evaluation_scope": "all imported canonical trajectories, including every alternative attempt",
-  "identical_duplicate_records": 1208,
+  "identical_duplicate_records": 1209,
   "independently_reproduced": 0,
   "label_policy": {
     "cost_units_are_source_native_and_not_globally_comparable": true,
@@ -2732,33 +3131,33 @@
     "unknown_is_failure": false
   },
   "outcomes": {
-    "failure": 79973,
-    "known_denominator": 134269,
-    "success": 54296,
-    "success_rate_all_trajectories": 0.34966962480196806,
-    "success_rate_known_outcomes": 0.40438224757762403,
-    "unknown": 21009
+    "failure": 81551,
+    "known_denominator": 141745,
+    "success": 60194,
+    "success_rate_all_trajectories": 0.255193406705218,
+    "success_rate_known_outcomes": 0.4246640093124978,
+    "unknown": 94131
   },
   "quality_tiers": {
-    "3": 51802,
-    "4": 19143,
-    "5": 84333
+    "3": 57110,
+    "4": 92261,
+    "5": 86505
   },
   "repeated_attempts": {
     "attempt_count_distribution": {
-      "1": 40296,
-      "10": 135,
-      "11": 58,
+      "1": 113423,
+      "10": 191,
+      "11": 95,
       "117": 1,
-      "12": 112,
-      "13": 20,
+      "12": 152,
+      "13": 67,
       "139": 1,
-      "14": 31,
-      "15": 6,
-      "16": 11,
-      "17": 2,
-      "18": 5,
-      "2": 22077,
+      "14": 86,
+      "15": 45,
+      "16": 34,
+      "17": 14,
+      "18": 9,
+      "2": 22392,
       "20": 6,
       "21": 3,
       "22": 2,
@@ -2766,13 +3165,15 @@
       "27": 2,
       "28": 1,
       "29": 2,
-      "3": 825,
+      "3": 854,
       "30": 278,
       "31": 1,
       "321": 1,
+      "34": 1,
       "345": 2,
       "346": 1,
       "358": 2,
+      "36": 19,
       "361": 1,
       "362": 2,
       "366": 2,
@@ -2803,99 +3204,99 @@
       "397": 1,
       "398": 3,
       "399": 1,
-      "4": 1867,
+      "4": 1911,
       "401": 1,
       "403": 1,
       "404": 1,
       "405": 1,
       "412": 1,
       "46": 1,
-      "5": 593,
-      "6": 271,
+      "5": 640,
+      "6": 329,
       "67": 1,
-      "7": 161,
-      "8": 217,
-      "9": 582
+      "7": 215,
+      "8": 266,
+      "9": 632
     },
     "empirical_success_at_k": [
       {
-        "eligible_tasks": 48864,
-        "empirical_success_at_k": 0.4352079240340537,
+        "eligible_tasks": 49849,
+        "empirical_success_at_k": 0.44237597544584645,
         "k": 1,
         "ordering": "source record locator then trajectory ID; empirical, not estimated pass@k",
-        "solved_tasks": 21266
+        "solved_tasks": 22052
       },
       {
-        "eligible_tasks": 26897,
-        "empirical_success_at_k": 0.5153362828568242,
+        "eligible_tasks": 27876,
+        "empirical_success_at_k": 0.5280886784330607,
         "k": 2,
         "ordering": "source record locator then trajectory ID; empirical, not estimated pass@k",
-        "solved_tasks": 13861
+        "solved_tasks": 14721
       },
       {
-        "eligible_tasks": 5057,
-        "empirical_success_at_k": 0.4601542416452442,
+        "eligible_tasks": 5721,
+        "empirical_success_at_k": 0.5163432966264639,
         "k": 3,
         "ordering": "source record locator then trajectory ID; empirical, not estimated pass@k",
-        "solved_tasks": 2327
+        "solved_tasks": 2954
       },
       {
-        "eligible_tasks": 4313,
-        "empirical_success_at_k": 0.4817992116856017,
+        "eligible_tasks": 4948,
+        "empirical_success_at_k": 0.5424413904607922,
         "k": 4,
         "ordering": "source record locator then trajectory ID; empirical, not estimated pass@k",
-        "solved_tasks": 2078
+        "solved_tasks": 2684
       },
       {
-        "eligible_tasks": 2487,
-        "empirical_success_at_k": 0.5158825894652191,
+        "eligible_tasks": 3078,
+        "empirical_success_at_k": 0.6007147498375569,
         "k": 5,
         "ordering": "source record locator then trajectory ID; empirical, not estimated pass@k",
-        "solved_tasks": 1283
+        "solved_tasks": 1849
       },
       {
-        "eligible_tasks": 1919,
-        "empirical_success_at_k": 0.6242834809796769,
+        "eligible_tasks": 2463,
+        "empirical_success_at_k": 0.6975233455136013,
         "k": 6,
         "ordering": "source record locator then trajectory ID; empirical, not estimated pass@k",
-        "solved_tasks": 1198
+        "solved_tasks": 1718
       },
       {
-        "eligible_tasks": 1663,
-        "empirical_success_at_k": 0.6915213469633193,
+        "eligible_tasks": 2149,
+        "empirical_success_at_k": 0.7519776640297813,
         "k": 7,
         "ordering": "source record locator then trajectory ID; empirical, not estimated pass@k",
-        "solved_tasks": 1150
+        "solved_tasks": 1616
       },
       {
-        "eligible_tasks": 1511,
-        "empirical_success_at_k": 0.7346128391793514,
+        "eligible_tasks": 1943,
+        "empirical_success_at_k": 0.784354091610911,
         "k": 8,
         "ordering": "source record locator then trajectory ID; empirical, not estimated pass@k",
-        "solved_tasks": 1110
+        "solved_tasks": 1524
       },
       {
-        "eligible_tasks": 1303,
-        "empirical_success_at_k": 0.7705295471987721,
+        "eligible_tasks": 1686,
+        "empirical_success_at_k": 0.8119810201660735,
         "k": 9,
         "ordering": "source record locator then trajectory ID; empirical, not estimated pass@k",
-        "solved_tasks": 1004
+        "solved_tasks": 1369
       },
       {
-        "eligible_tasks": 724,
-        "empirical_success_at_k": 0.649171270718232,
+        "eligible_tasks": 1057,
+        "empirical_success_at_k": 0.7426679280983917,
         "k": 10,
         "ordering": "source record locator then trajectory ID; empirical, not estimated pass@k",
-        "solved_tasks": 470
+        "solved_tasks": 785
       }
     ],
-    "fraction_solved_at_least_once": 0.47888015717092336,
-    "tasks_solved_at_least_once": 23400,
-    "tasks_with_known_outcomes": 48864,
-    "tasks_with_multiple_trajectories": 27364
+    "fraction_solved_at_least_once": 0.4877530141025898,
+    "tasks_solved_at_least_once": 24314,
+    "tasks_with_known_outcomes": 49849,
+    "tasks_with_multiple_trajectories": 28343
   },
   "schema_version": 1,
-  "structurally_eligible": 143999,
-  "training_eligible": 110208,
-  "unique_tasks": 67610
+  "structurally_eligible": 223930,
+  "training_eligible": 183326,
+  "unique_tasks": 141673
 }

index.sqlite

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:25b97793c7b002995cdbea6daddb4e0068b81b40aaeb11179a2915f9810857fd
-size 178581504
+oid sha256:7ec48bfcea90d0e09016cac328769d021c23a94bb7cdbac8db11018872bd554d
+size 301342720

manifest.json

@@ -1,8 +1,38 @@
 {
+  "append_history": [
+    {
+      "accepted_by_source": {
+        "agentsuite_bfcl_v4": 6160,
+        "delm_deepswe": 360,
+        "delm_terminalbench4": 358,
+        "swebench_multilingual_glm5": 300,
+        "swebench_multilingual_minimax_m25": 299,
+        "toolbench_full": 73121
+      },
+      "accepted_new_trajectories": 80598,
+      "affected_tasks": 74106,
+      "base_manifest_sha256": "6b2e45c02e8a0c09d87d7ced028b55072b5ba93ba89c443bcccc523206bfbb4a",
+      "base_scores_recomputed": false,
+      "base_shards_opened": false,
+      "base_trajectories": 155278,
+      "exact_duplicates_by_source": {
+        "agentsuite_bfcl_v4": 1
+      },
+      "existing_trajectory_ids_by_source": {},
+      "new_source_records": 80599,
+      "release": "v0.4.0",
+      "schema_version": 1,
+      "tasks_promoted_to_protected_evaluation": 0
+    }
+  ],
+  "base_release": {
+    "manifest_sha256": "6b2e45c02e8a0c09d87d7ced028b55072b5ba93ba89c443bcccc523206bfbb4a",
+    "trajectories": 155278
+  },
   "content_complete": false,
-  "duplicate_provenance_records": 1208,
+  "duplicate_provenance_records": 1209,
   "duplicate_trajectory_ids_skipped": 0,
-  "evaluated": 155278,
+  "evaluated": 235876,
   "files": {
     "all": null,
     "duplicate_provenance": "duplicate_provenance.jsonl.gz",
@@ -15,6 +45,8 @@
     }
   },
   "hosted_source_counts": {
+    "delm_deepswe": 360,
+    "delm_terminalbench4": 358,
     "mind2web": 1009,
     "nvidia_open_swe_traces": 51626,
     "official_tau3": 200,
@@ -24,10 +56,11 @@
     "swesmith_trajectories": 24100,
     "terminalbench_trajectories": 34263,
     "terminaltraj": 20000,
-    "toolbench": 11
+    "toolbench": 11,
+    "toolbench_full": 73121
   },
-  "hosted_trajectories": 136373,
-  "identical_records_collapsed": 1208,
+  "hosted_trajectories": 210212,
+  "identical_records_collapsed": 1209,
   "inputs": [
     {
       "path": "canonical/mind2web/default--train--0000.jsonl.gz",
@@ -479,6 +512,166 @@
       "source_counts": {
         "terminaltraj": 10000
       }
+    },
+    {
+      "path": "canonical/delm_deepswe/astra-ds-n2.jsonl.gz",
+      "records": 60,
+      "release_added": "v0.4.0",
+      "sha256": "81abee17b8c2be577d71333d705935448b4b9ac7cd2bba404493ce1094a9e612",
+      "size_bytes": 5258861,
+      "source_counts": {
+        "delm_deepswe": 60
+      }
+    },
+    {
+      "path": "canonical/delm_deepswe/astra-ds-n4.jsonl.gz",
+      "records": 120,
+      "release_added": "v0.4.0",
+      "sha256": "a000abca021184d197387ee3b8f35345143a1094e8577be1bba80ceeda75be8d",
+      "size_bytes": 10854442,
+      "source_counts": {
+        "delm_deepswe": 120
+      }
+    },
+    {
+      "path": "canonical/delm_terminalbench4/astra-tb-n2.jsonl.gz",
+      "records": 60,
+      "release_added": "v0.4.0",
+      "sha256": "d0605eeb9a0774129dcccfb5707619fcef7a08e807af2ec603969f901aa080d1",
+      "size_bytes": 4039168,
+      "source_counts": {
+        "delm_terminalbench4": 60
+      }
+    },
+    {
+      "path": "canonical/delm_terminalbench4/astra-tb-n4.jsonl.gz",
+      "records": 120,
+      "release_added": "v0.4.0",
+      "sha256": "79119caa42f08dc1ffe4f93e309b7cb21108abea4a6896d63288bc1bf4f04be9",
+      "size_bytes": 8210876,
+      "source_counts": {
+        "delm_terminalbench4": 120
+      }
+    },
+    {
+      "path": "canonical/delm_deepswe/opus55-ds-n2.jsonl.gz",
+      "records": 60,
+      "release_added": "v0.4.0",
+      "sha256": "aaa1f62cae7f04378c60b156a001f46c776e9dbe42d5879211104b4aa28232d6",
+      "size_bytes": 5280025,
+      "source_counts": {
+        "delm_deepswe": 60
+      }
+    },
+    {
+      "path": "canonical/delm_deepswe/opus55-ds-n4.jsonl.gz",
+      "records": 120,
+      "release_added": "v0.4.0",
+      "sha256": "3b6ff27dbe4299271da0d35965cb995d764b2a0cd56b1672b4a1aef47b4eaa82",
+      "size_bytes": 11092230,
+      "source_counts": {
+        "delm_deepswe": 120
+      }
+    },
+    {
+      "path": "canonical/delm_terminalbench4/opus55-tb-n2.jsonl.gz",
+      "records": 58,
+      "release_added": "v0.4.0",
+      "sha256": "f1f3e3b3a7f80e0f1afec5f307076a7252239238457561f07ccf24eff9518b65",
+      "size_bytes": 5909974,
+      "source_counts": {
+        "delm_terminalbench4": 58
+      }
+    },
+    {
+      "path": "canonical/delm_terminalbench4/opus55-tb-n4.jsonl.gz",
+      "records": 120,
+      "release_added": "v0.4.0",
+      "sha256": "aa994fc8ab96758ff549dd2ea123078499981ec2d805003a132e09de18ceec78",
+      "size_bytes": 18021731,
+      "source_counts": {
+        "delm_terminalbench4": 120
+      }
+    },
+    {
+      "path": "canonical/toolbench_full/part-0000.jsonl.gz",
+      "records": 10000,
+      "release_added": "v0.4.0",
+      "sha256": "c675b9916db47c5ea5cf21695214db6c59f16a40f6e7f2d8749ce4b9ff573bb3",
+      "size_bytes": 25355904,
+      "source_counts": {
+        "toolbench_full": 10000
+      }
+    },
+    {
+      "path": "canonical/toolbench_full/part-0001.jsonl.gz",
+      "records": 10000,
+      "release_added": "v0.4.0",
+      "sha256": "867fa4371da83f9dbc01286fc2eb787ea4713f738ecf57e7aa0ba67d5994b89b",
+      "size_bytes": 25459527,
+      "source_counts": {
+        "toolbench_full": 10000
+      }
+    },
+    {
+      "path": "canonical/toolbench_full/part-0002.jsonl.gz",
+      "records": 10000,
+      "release_added": "v0.4.0",
+      "sha256": "7d14e67d3d5eba9d9c5eea73cb5f99610a929f32b8ee51ca02ca5154352c5a75",
+      "size_bytes": 25109866,
+      "source_counts": {
+        "toolbench_full": 10000
+      }
+    },
+    {
+      "path": "canonical/toolbench_full/part-0003.jsonl.gz",
+      "records": 10000,
+      "release_added": "v0.4.0",
+      "sha256": "f73c83e5d5ed2241c060ac1cd12be331c2717b745cd83900049e82ea1feebc68",
+      "size_bytes": 25119809,
+      "source_counts": {
+        "toolbench_full": 10000
+      }
+    },
+    {
+      "path": "canonical/toolbench_full/part-0004.jsonl.gz",
+      "records": 10000,
+      "release_added": "v0.4.0",
+      "sha256": "4342aa84d49c7fef7e890cdb42d66b49d26ccb9c1e047b54626c3ffce11ab864",
+      "size_bytes": 25115680,
+      "source_counts": {
+        "toolbench_full": 10000
+      }
+    },
+    {
+      "path": "canonical/toolbench_full/part-0005.jsonl.gz",
+      "records": 10000,
+      "release_added": "v0.4.0",
+      "sha256": "98362fccba3b4054b3c419b351718bee0a4f48d7a3e6f574cce0b1c059cfbeb4",
+      "size_bytes": 25189501,
+      "source_counts": {
+        "toolbench_full": 10000
+      }
+    },
+    {
+      "path": "canonical/toolbench_full/part-0006.jsonl.gz",
+      "records": 10000,
+      "release_added": "v0.4.0",
+      "sha256": "d84c6fdd79b7c14e262fe87a214e8a87af4d13155856cf76c03135bbdc85467a",
+      "size_bytes": 25176940,
+      "source_counts": {
+        "toolbench_full": 10000
+      }
+    },
+    {
+      "path": "canonical/toolbench_full/part-0007.jsonl.gz",
+      "records": 3121,
+      "release_added": "v0.4.0",
+      "sha256": "9e13bbad290b10dfac0bf359d8e2497308ad5c9ea8cd8a35ce4f3f3129bf4b69",
+      "size_bytes": 7863644,
+      "source_counts": {
+        "toolbench_full": 3121
+      }
     }
   ],
   "inventory_inputs": [
@@ -1211,35 +1404,311 @@
       "upstream": [
         "https://huggingface.co/datasets/m-a-p/TerminalTraj/tree/5c1823f4a8b9ca0cf02c27d2db52c5b35b53a308"
       ]
+    },
+    {
+      "canonical_sha256": "3e7ae53433c2018b7975d34df8fc4e4e1fb4924b6d9d2865fe18220279d5b37b",
+      "content_availability": "metadata_only",
+      "path": null,
+      "records": 6161,
+      "release_added": "v0.4.0",
+      "size_bytes": 62243696,
+      "source_counts": {
+        "agentsuite_bfcl_v4": 6161
+      },
+      "upstream": [
+        "https://huggingface.co/datasets/AgentSuite/BFCL_V4-trajectories/tree/53730302150e4a9c24026e5f591964c500ba7f8a"
+      ]
+    },
+    {
+      "canonical_sha256": "81abee17b8c2be577d71333d705935448b4b9ac7cd2bba404493ce1094a9e612",
+      "content_availability": "hosted",
+      "path": "canonical/delm_deepswe/astra-ds-n2.jsonl.gz",
+      "records": 60,
+      "release_added": "v0.4.0",
+      "size_bytes": 5258861,
+      "source_counts": {
+        "delm_deepswe": 60
+      },
+      "upstream": [
+        "https://huggingface.co/datasets/yuzhenm/delm-720-trajectories/tree/15d5133454d5ecd702fa957be414c479f8afcdd0"
+      ]
+    },
+    {
+      "canonical_sha256": "a000abca021184d197387ee3b8f35345143a1094e8577be1bba80ceeda75be8d",
+      "content_availability": "hosted",
+      "path": "canonical/delm_deepswe/astra-ds-n4.jsonl.gz",
+      "records": 120,
+      "release_added": "v0.4.0",
+      "size_bytes": 10854442,
+      "source_counts": {
+        "delm_deepswe": 120
+      },
+      "upstream": [
+        "https://huggingface.co/datasets/yuzhenm/delm-720-trajectories/tree/15d5133454d5ecd702fa957be414c479f8afcdd0"
+      ]
+    },
+    {
+      "canonical_sha256": "d0605eeb9a0774129dcccfb5707619fcef7a08e807af2ec603969f901aa080d1",
+      "content_availability": "hosted",
+      "path": "canonical/delm_terminalbench4/astra-tb-n2.jsonl.gz",
+      "records": 60,
+      "release_added": "v0.4.0",
+      "size_bytes": 4039168,
+      "source_counts": {
+        "delm_terminalbench4": 60
+      },
+      "upstream": [
+        "https://huggingface.co/datasets/yuzhenm/delm-720-trajectories/tree/15d5133454d5ecd702fa957be414c479f8afcdd0"
+      ]
+    },
+    {
+      "canonical_sha256": "79119caa42f08dc1ffe4f93e309b7cb21108abea4a6896d63288bc1bf4f04be9",
+      "content_availability": "hosted",
+      "path": "canonical/delm_terminalbench4/astra-tb-n4.jsonl.gz",
+      "records": 120,
+      "release_added": "v0.4.0",
+      "size_bytes": 8210876,
+      "source_counts": {
+        "delm_terminalbench4": 120
+      },
+      "upstream": [
+        "https://huggingface.co/datasets/yuzhenm/delm-720-trajectories/tree/15d5133454d5ecd702fa957be414c479f8afcdd0"
+      ]
+    },
+    {
+      "canonical_sha256": "aaa1f62cae7f04378c60b156a001f46c776e9dbe42d5879211104b4aa28232d6",
+      "content_availability": "hosted",
+      "path": "canonical/delm_deepswe/opus55-ds-n2.jsonl.gz",
+      "records": 60,
+      "release_added": "v0.4.0",
+      "size_bytes": 5280025,
+      "source_counts": {
+        "delm_deepswe": 60
+      },
+      "upstream": [
+        "https://huggingface.co/datasets/yuzhenm/delm-720-trajectories/tree/15d5133454d5ecd702fa957be414c479f8afcdd0"
+      ]
+    },
+    {
+      "canonical_sha256": "3b6ff27dbe4299271da0d35965cb995d764b2a0cd56b1672b4a1aef47b4eaa82",
+      "content_availability": "hosted",
+      "path": "canonical/delm_deepswe/opus55-ds-n4.jsonl.gz",
+      "records": 120,
+      "release_added": "v0.4.0",
+      "size_bytes": 11092230,
+      "source_counts": {
+        "delm_deepswe": 120
+      },
+      "upstream": [
+        "https://huggingface.co/datasets/yuzhenm/delm-720-trajectories/tree/15d5133454d5ecd702fa957be414c479f8afcdd0"
+      ]
+    },
+    {
+      "canonical_sha256": "f1f3e3b3a7f80e0f1afec5f307076a7252239238457561f07ccf24eff9518b65",
+      "content_availability": "hosted",
+      "path": "canonical/delm_terminalbench4/opus55-tb-n2.jsonl.gz",
+      "records": 58,
+      "release_added": "v0.4.0",
+      "size_bytes": 5909974,
+      "source_counts": {
+        "delm_terminalbench4": 58
+      },
+      "upstream": [
+        "https://huggingface.co/datasets/yuzhenm/delm-720-trajectories/tree/15d5133454d5ecd702fa957be414c479f8afcdd0"
+      ]
+    },
+    {
+      "canonical_sha256": "aa994fc8ab96758ff549dd2ea123078499981ec2d805003a132e09de18ceec78",
+      "content_availability": "hosted",
+      "path": "canonical/delm_terminalbench4/opus55-tb-n4.jsonl.gz",
+      "records": 120,
+      "release_added": "v0.4.0",
+      "size_bytes": 18021731,
+      "source_counts": {
+        "delm_terminalbench4": 120
+      },
+      "upstream": [
+        "https://huggingface.co/datasets/yuzhenm/delm-720-trajectories/tree/15d5133454d5ecd702fa957be414c479f8afcdd0"
+      ]
+    },
+    {
+      "canonical_sha256": "c98dc541275a65e7dcba70730c75474673118a76937283c8e5c12a36e66c3c3b",
+      "content_availability": "metadata_only",
+      "path": null,
+      "records": 300,
+      "release_added": "v0.4.0",
+      "size_bytes": 14782256,
+      "source_counts": {
+        "swebench_multilingual_glm5": 300
+      },
+      "upstream": [
+        "https://huggingface.co/datasets/AlienKevin/SWE-bench-multilingual-glm-5-trajectories/tree/337a922eac45fc0d63271c2a50c3a7c84dae34de"
+      ]
+    },
+    {
+      "canonical_sha256": "f9a519a66c449639cbb3a83d8862a020484045f597179802e8d950a46d4bbf3d",
+      "content_availability": "metadata_only",
+      "path": null,
+      "records": 299,
+      "release_added": "v0.4.0",
+      "size_bytes": 13584021,
+      "source_counts": {
+        "swebench_multilingual_minimax_m25": 299
+      },
+      "upstream": [
+        "https://huggingface.co/datasets/AlienKevin/SWE-bench-multilingual-minimax-m2.5-trajectories/tree/479f2e1f49c71c6d8fa778e5de945b380c073b8c"
+      ]
+    },
+    {
+      "canonical_sha256": "c675b9916db47c5ea5cf21695214db6c59f16a40f6e7f2d8749ce4b9ff573bb3",
+      "content_availability": "hosted",
+      "path": "canonical/toolbench_full/part-0000.jsonl.gz",
+      "records": 10000,
+      "release_added": "v0.4.0",
+      "size_bytes": 25355904,
+      "source_counts": {
+        "toolbench_full": 10000
+      },
+      "upstream": [
+        "https://github.com/OpenBMB/ToolBench/tree/d56fdd89faf8c91fa135090b212bb9057ee5cfc2"
+      ]
+    },
+    {
+      "canonical_sha256": "867fa4371da83f9dbc01286fc2eb787ea4713f738ecf57e7aa0ba67d5994b89b",
+      "content_availability": "hosted",
+      "path": "canonical/toolbench_full/part-0001.jsonl.gz",
+      "records": 10000,
+      "release_added": "v0.4.0",
+      "size_bytes": 25459527,
+      "source_counts": {
+        "toolbench_full": 10000
+      },
+      "upstream": [
+        "https://github.com/OpenBMB/ToolBench/tree/d56fdd89faf8c91fa135090b212bb9057ee5cfc2"
+      ]
+    },
+    {
+      "canonical_sha256": "7d14e67d3d5eba9d9c5eea73cb5f99610a929f32b8ee51ca02ca5154352c5a75",
+      "content_availability": "hosted",
+      "path": "canonical/toolbench_full/part-0002.jsonl.gz",
+      "records": 10000,
+      "release_added": "v0.4.0",
+      "size_bytes": 25109866,
+      "source_counts": {
+        "toolbench_full": 10000
+      },
+      "upstream": [
+        "https://github.com/OpenBMB/ToolBench/tree/d56fdd89faf8c91fa135090b212bb9057ee5cfc2"
+      ]
+    },
+    {
+      "canonical_sha256": "f73c83e5d5ed2241c060ac1cd12be331c2717b745cd83900049e82ea1feebc68",
+      "content_availability": "hosted",
+      "path": "canonical/toolbench_full/part-0003.jsonl.gz",
+      "records": 10000,
+      "release_added": "v0.4.0",
+      "size_bytes": 25119809,
+      "source_counts": {
+        "toolbench_full": 10000
+      },
+      "upstream": [
+        "https://github.com/OpenBMB/ToolBench/tree/d56fdd89faf8c91fa135090b212bb9057ee5cfc2"
+      ]
+    },
+    {
+      "canonical_sha256": "4342aa84d49c7fef7e890cdb42d66b49d26ccb9c1e047b54626c3ffce11ab864",
+      "content_availability": "hosted",
+      "path": "canonical/toolbench_full/part-0004.jsonl.gz",
+      "records": 10000,
+      "release_added": "v0.4.0",
+      "size_bytes": 25115680,
+      "source_counts": {
+        "toolbench_full": 10000
+      },
+      "upstream": [
+        "https://github.com/OpenBMB/ToolBench/tree/d56fdd89faf8c91fa135090b212bb9057ee5cfc2"
+      ]
+    },
+    {
+      "canonical_sha256": "98362fccba3b4054b3c419b351718bee0a4f48d7a3e6f574cce0b1c059cfbeb4",
+      "content_availability": "hosted",
+      "path": "canonical/toolbench_full/part-0005.jsonl.gz",
+      "records": 10000,
+      "release_added": "v0.4.0",
+      "size_bytes": 25189501,
+      "source_counts": {
+        "toolbench_full": 10000
+      },
+      "upstream": [
+        "https://github.com/OpenBMB/ToolBench/tree/d56fdd89faf8c91fa135090b212bb9057ee5cfc2"
+      ]
+    },
+    {
+      "canonical_sha256": "d84c6fdd79b7c14e262fe87a214e8a87af4d13155856cf76c03135bbdc85467a",
+      "content_availability": "hosted",
+      "path": "canonical/toolbench_full/part-0006.jsonl.gz",
+      "records": 10000,
+      "release_added": "v0.4.0",
+      "size_bytes": 25176940,
+      "source_counts": {
+        "toolbench_full": 10000
+      },
+      "upstream": [
+        "https://github.com/OpenBMB/ToolBench/tree/d56fdd89faf8c91fa135090b212bb9057ee5cfc2"
+      ]
+    },
+    {
+      "canonical_sha256": "9e13bbad290b10dfac0bf359d8e2497308ad5c9ea8cd8a35ce4f3f3129bf4b69",
+      "content_availability": "hosted",
+      "path": "canonical/toolbench_full/part-0007.jsonl.gz",
+      "records": 3121,
+      "release_added": "v0.4.0",
+      "size_bytes": 7863644,
+      "source_counts": {
+        "toolbench_full": 3121
+      },
+      "upstream": [
+        "https://github.com/OpenBMB/ToolBench/tree/d56fdd89faf8c91fa135090b212bb9057ee5cfc2"
+      ]
     }
   ],
   "metadata_only_source_counts": {
+    "agentsuite_bfcl_v4": 6160,
     "agentsuite_tau2": 8340,
     "browsecomp_rollouts": 5063,
+    "swebench_multilingual_glm5": 300,
+    "swebench_multilingual_minimax_m25": 299,
     "swegym_sampled": 5502
   },
-  "metadata_only_trajectories": 18905,
+  "metadata_only_trajectories": 25664,
   "outcome_counts": {
-    "failure": 79973,
-    "success": 54296,
-    "unknown": 21009
+    "failure": 81551,
+    "success": 60194,
+    "unknown": 94131
   },
   "partition_counts": {
-    "structurally_eligible": 143999,
-    "successful": 54296,
-    "training_eligible": 110208,
-    "unknown": 21009,
-    "unsuccessful": 79973
+    "structurally_eligible": 223930,
+    "successful": 60194,
+    "training_eligible": 183326,
+    "unknown": 94131,
+    "unsuccessful": 81551
   },
   "permissions": "registry/permissions.json",
-  "ranked": 155278,
+  "previous_hf_revision": "e9214095fdcb9ce6a7ddaefb09426ae531e0221d",
+  "ranked": 235876,
+  "release": "v0.4.0",
   "schema_version": 1,
   "source_counts": {
+    "agentsuite_bfcl_v4": 6160,
     "agentsuite_tau2": 8340,
     "browsecomp_rollouts": 5063,
+    "delm_deepswe": 360,
+    "delm_terminalbench4": 358,
     "mind2web": 1009,
     "nvidia_open_swe_traces": 51626,
     "official_tau3": 200,
+    "swebench_multilingual_glm5": 300,
+    "swebench_multilingual_minimax_m25": 299,
     "swebench_multilingual_trajectories": 129,
     "swebench_verified_trajectories": 4544,
     "swegym_openhands": 491,
@@ -1247,14 +1716,20 @@
     "swesmith_trajectories": 24100,
     "terminalbench_trajectories": 34263,
     "terminaltraj": 20000,
-    "toolbench": 11
+    "toolbench": 11,
+    "toolbench_full": 73121
   },
   "source_record_counts": {
+    "agentsuite_bfcl_v4": 6161,
     "agentsuite_tau2": 8340,
     "browsecomp_rollouts": 5063,
+    "delm_deepswe": 360,
+    "delm_terminalbench4": 358,
     "mind2web": 1009,
     "nvidia_open_swe_traces": 51626,
     "official_tau3": 200,
+    "swebench_multilingual_glm5": 300,
+    "swebench_multilingual_minimax_m25": 299,
     "swebench_multilingual_trajectories": 129,
     "swebench_verified_trajectories": 5000,
     "swegym_openhands": 491,
@@ -1262,16 +1737,17 @@
     "swesmith_trajectories": 24100,
     "terminalbench_trajectories": 34462,
     "terminaltraj": 20000,
-    "toolbench": 11
+    "toolbench": 11,
+    "toolbench_full": 73121
   },
-  "source_records": 156486,
-  "structurally_eligible": 143999,
+  "source_records": 237085,
+  "structurally_eligible": 223930,
   "task_split_counts": {
-    "protected_evaluation": 1948,
-    "test": 6442,
-    "train": 52592,
-    "validation": 6628
+    "protected_evaluation": 2890,
+    "test": 13636,
+    "train": 111191,
+    "validation": 13956
   },
-  "training_eligible": 110208,
-  "trajectories": 155278
+  "training_eligible": 183326,
+  "trajectories": 235876
 }

rankings.jsonl.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5a56a9ae57c860130af600fb534afaaa337124c15646573ec9f8dc86b950fa41
-size 10959026
+oid sha256:64edd36bd41f1dec7fdf89e6efb79f6c8fe9661fc7c6f4475f9528dc5ad8be85
+size 17271621

registry/benchmarks.json

@@ -44,6 +44,8 @@
       },
       "source_ids": [
         "nvidia_open_swe_traces",
+        "swebench_multilingual_glm5",
+        "swebench_multilingual_minimax_m25",
         "swebench_multilingual_trajectories",
         "swebench_verified_trajectories"
       ]
@@ -144,7 +146,49 @@
         "imported_evidence": "Published generation status.",
         "caveat": "This is not independent semantic proof of task success."
       },
-      "source_ids": ["toolbench"]
+      "source_ids": ["toolbench", "toolbench_full"]
+    },
+    {
+      "id": "bfcl-v4",
+      "name": "BFCL V4",
+      "description": "Agentic function-calling tasks with web and memory tool environments.",
+      "url": "https://gorilla.cs.berkeley.edu/leaderboard.html",
+      "metric": {
+        "name": "BFCL evaluator score",
+        "type": "native_binary",
+        "direction": "higher_is_better",
+        "imported_evidence": "AgentSuite-published V4 evaluator score.",
+        "caveat": "Only web and memory rows with returned tool observations qualify; scores were not rerun."
+      },
+      "source_ids": ["agentsuite_bfcl_v4"]
+    },
+    {
+      "id": "deepswe",
+      "name": "DeepSWE",
+      "description": "Repository engineering tasks executed by multi-agent DeLM teams.",
+      "url": "https://huggingface.co/datasets/yuzhenm/delm-720-trajectories",
+      "metric": {
+        "name": "CTRF test-suite pass",
+        "type": "native_binary",
+        "direction": "higher_is_better",
+        "imported_evidence": "Published per-agent CTRF grading summaries.",
+        "caveat": "Protected evaluation tasks contain anti-training canaries and were not replayed here."
+      },
+      "source_ids": ["delm_deepswe"]
+    },
+    {
+      "id": "terminal-bench-4",
+      "name": "Terminal-Bench 4",
+      "description": "Protected terminal tasks executed by multi-agent DeLM teams.",
+      "url": "https://huggingface.co/datasets/yuzhenm/delm-720-trajectories",
+      "metric": {
+        "name": "task reward",
+        "type": "native_reward",
+        "direction": "higher_is_better",
+        "imported_evidence": "Published per-agent grader reward.",
+        "caveat": "Protected evaluation tasks contain Harbor anti-training canaries and were not replayed here."
+      },
+      "source_ids": ["delm_terminalbench4"]
     }
   ]
 }

registry/permissions.json

@@ -180,6 +180,22 @@
         {"finding": "Pinned dataset card declares Apache-2.0, describes complete trajectories, and identifies the public leaderboard source.", "url": "https://huggingface.co/datasets/yoonholee/terminalbench-trajectories/blob/04e8940f5b6736a7ce8d22224fe2f2af74163ed2/README.md"}
       ]
     },
+    {
+      "source_id": "agentsuite_bfcl_v4",
+      "upstream": "https://huggingface.co/datasets/AgentSuite/BFCL_V4-trajectories/tree/53730302150e4a9c24026e5f591964c500ba7f8a",
+      "declared_license": null,
+      "public_access": {"permission": "permitted", "verification": "verified"},
+      "local_research_use": {"permission": "permitted_with_conditions", "verification": "inferred", "conditions": ["public benchmark research context", "honor BFCL and model-provider terms"]},
+      "model_training": {"permission": "uncertain", "verification": "uncertain", "conditions": ["avoid protected evaluation leakage", "obtain or confirm trajectory-output rights"]},
+      "commercial_training": {"permission": "uncertain", "verification": "uncertain"},
+      "redistribution_original": {"permission": "uncertain", "verification": "uncertain"},
+      "redistribution_transformed": {"permission": "uncertain", "verification": "uncertain"},
+      "public_payload": "metadata_only",
+      "evidence": [
+        {"finding": "The pinned public release supplies 30 model output files and provenance but no license or use grant.", "url": "https://huggingface.co/datasets/AgentSuite/BFCL_V4-trajectories/blob/53730302150e4a9c24026e5f591964c500ba7f8a/README.md"},
+        {"finding": "BFCL benchmark materials do not independently grant rights over AgentSuite's separately published model interaction histories.", "url": "https://github.com/ShishirPatil/gorilla"}
+      ]
+    },
     {
       "source_id": "browsecomp_rollouts",
       "upstream": "https://huggingface.co/datasets/rl-rag/browsecomp-gpt-oss-120b-260222/tree/ac4854be0eb527cdf90f6c7d09446eeb7a3d1dbd",
@@ -196,6 +212,102 @@
         {"finding": "Underlying BrowseComp questions are MIT in OpenAI simple-evals; that does not automatically license the separate rollouts or retrieved web content.", "url": "https://github.com/openai/simple-evals/tree/652c89d0ca9df547706735883097e9537d40dc47"}
       ]
     },
+    {
+      "source_id": "alfworld_replayed_experts",
+      "upstream": "https://huggingface.co/datasets/Joshyxwa/cp2107-textworld-trajectories/tree/6402e311e4f0f73e79bc4ebf8fbb49e0b3b3b898",
+      "declared_license": null,
+      "public_access": {"permission": "restricted", "verification": "verified", "conditions": ["accept the publisher's Hugging Face gated-access terms"]},
+      "local_research_use": {"permission": "uncertain", "verification": "uncertain"},
+      "model_training": {"permission": "uncertain", "verification": "uncertain"},
+      "commercial_training": {"permission": "uncertain", "verification": "uncertain"},
+      "redistribution_original": {"permission": "uncertain", "verification": "uncertain"},
+      "redistribution_transformed": {"permission": "uncertain", "verification": "uncertain"},
+      "public_payload": "metadata_only",
+      "evidence": [
+        {"finding": "The pinned repository is publicly discoverable but auto-gated and declares no dataset license in its card.", "url": "https://huggingface.co/datasets/Joshyxwa/cp2107-textworld-trajectories/tree/6402e311e4f0f73e79bc4ebf8fbb49e0b3b3b898"},
+        {"finding": "The card reports successful deterministic simulator replays but does not grant redistribution rights for this separate trajectory publication.", "url": "https://huggingface.co/datasets/Joshyxwa/cp2107-textworld-trajectories/blob/6402e311e4f0f73e79bc4ebf8fbb49e0b3b3b898/README.md"}
+      ]
+    },
+    {
+      "source_id": "delm_deepswe",
+      "upstream": "https://huggingface.co/datasets/yuzhenm/delm-720-trajectories/tree/15d5133454d5ecd702fa957be414c479f8afcdd0",
+      "declared_license": "MIT",
+      "public_access": {"permission": "permitted", "verification": "verified"},
+      "local_research_use": {"permission": "permitted", "verification": "verified", "conditions": ["retain MIT notice", "treat benchmark material as protected evaluation data"]},
+      "model_training": {"permission": "restricted", "verification": "verified", "conditions": ["embedded benchmark canaries explicitly prohibit inclusion in training corpora"]},
+      "commercial_training": {"permission": "restricted", "verification": "verified", "conditions": ["embedded benchmark canaries explicitly prohibit inclusion in training corpora"]},
+      "redistribution_original": {"permission": "permitted_with_conditions", "verification": "verified", "conditions": ["retain MIT notice", "retain benchmark protection warnings"]},
+      "redistribution_transformed": {"permission": "permitted_with_conditions", "verification": "verified", "conditions": ["retain MIT notice and source attribution", "retain benchmark protection warnings"]},
+      "public_payload": "hosted",
+      "evidence": [
+        {"finding": "The pinned dataset card declares MIT for the published multi-agent trajectory archives.", "url": "https://huggingface.co/datasets/yuzhenm/delm-720-trajectories/blob/15d5133454d5ecd702fa957be414c479f8afcdd0/README.md"},
+        {"finding": "Imported task prompts and tool observations contain explicit benchmark anti-training canaries; the source is therefore evaluation-only regardless of archive licensing.", "url": "https://huggingface.co/datasets/yuzhenm/delm-720-trajectories/tree/15d5133454d5ecd702fa957be414c479f8afcdd0"}
+      ]
+    },
+    {
+      "source_id": "delm_terminalbench4",
+      "upstream": "https://huggingface.co/datasets/yuzhenm/delm-720-trajectories/tree/15d5133454d5ecd702fa957be414c479f8afcdd0",
+      "declared_license": "MIT",
+      "public_access": {"permission": "permitted", "verification": "verified"},
+      "local_research_use": {"permission": "permitted", "verification": "verified", "conditions": ["retain MIT notice", "treat benchmark material as protected evaluation data"]},
+      "model_training": {"permission": "restricted", "verification": "verified", "conditions": ["Harbor canaries explicitly prohibit inclusion in training corpora"]},
+      "commercial_training": {"permission": "restricted", "verification": "verified", "conditions": ["Harbor canaries explicitly prohibit inclusion in training corpora"]},
+      "redistribution_original": {"permission": "permitted_with_conditions", "verification": "verified", "conditions": ["retain MIT notice", "retain benchmark protection warnings"]},
+      "redistribution_transformed": {"permission": "permitted_with_conditions", "verification": "verified", "conditions": ["retain MIT notice and source attribution", "retain benchmark protection warnings"]},
+      "public_payload": "hosted",
+      "evidence": [
+        {"finding": "The pinned dataset card declares MIT for the published multi-agent trajectory archives.", "url": "https://huggingface.co/datasets/yuzhenm/delm-720-trajectories/blob/15d5133454d5ecd702fa957be414c479f8afcdd0/README.md"},
+        {"finding": "Terminal-Bench prompts and observations contain explicit Harbor anti-training canaries; the source is evaluation-only.", "url": "https://huggingface.co/datasets/yuzhenm/delm-720-trajectories/tree/15d5133454d5ecd702fa957be414c479f8afcdd0"}
+      ]
+    },
+    {
+      "source_id": "swebench_multilingual_glm5",
+      "upstream": "https://huggingface.co/datasets/AlienKevin/SWE-bench-multilingual-glm-5-trajectories/tree/337a922eac45fc0d63271c2a50c3a7c84dae34de",
+      "declared_license": null,
+      "public_access": {"permission": "permitted", "verification": "verified"},
+      "local_research_use": {"permission": "permitted_with_conditions", "verification": "inferred", "conditions": ["public research release", "honor underlying repository licenses"]},
+      "model_training": {"permission": "uncertain", "verification": "uncertain"},
+      "commercial_training": {"permission": "uncertain", "verification": "uncertain"},
+      "redistribution_original": {"permission": "uncertain", "verification": "uncertain"},
+      "redistribution_transformed": {"permission": "uncertain", "verification": "uncertain"},
+      "public_payload": "metadata_only",
+      "evidence": [
+        {"finding": "The pinned trajectory card provides schema and provenance but no license or usage grant.", "url": "https://huggingface.co/datasets/AlienKevin/SWE-bench-multilingual-glm-5-trajectories/blob/337a922eac45fc0d63271c2a50c3a7c84dae34de/README.md"},
+        {"finding": "The underlying benchmark is MIT; this does not independently license the separate model outputs.", "url": "https://huggingface.co/datasets/SWE-bench/SWE-bench_Multilingual/tree/846e647b9f33c0b51b739d005d13d85493c9af09"}
+      ]
+    },
+    {
+      "source_id": "swebench_multilingual_minimax_m25",
+      "upstream": "https://huggingface.co/datasets/AlienKevin/SWE-bench-multilingual-minimax-m2.5-trajectories/tree/479f2e1f49c71c6d8fa778e5de945b380c073b8c",
+      "declared_license": null,
+      "public_access": {"permission": "permitted", "verification": "verified"},
+      "local_research_use": {"permission": "permitted_with_conditions", "verification": "inferred", "conditions": ["public research release", "honor underlying repository licenses"]},
+      "model_training": {"permission": "uncertain", "verification": "uncertain"},
+      "commercial_training": {"permission": "uncertain", "verification": "uncertain"},
+      "redistribution_original": {"permission": "uncertain", "verification": "uncertain"},
+      "redistribution_transformed": {"permission": "uncertain", "verification": "uncertain"},
+      "public_payload": "metadata_only",
+      "evidence": [
+        {"finding": "The pinned trajectory card provides schema and provenance but no license or usage grant.", "url": "https://huggingface.co/datasets/AlienKevin/SWE-bench-multilingual-minimax-m2.5-trajectories/blob/479f2e1f49c71c6d8fa778e5de945b380c073b8c/README.md"},
+        {"finding": "The underlying benchmark is MIT; this does not independently license the separate model outputs.", "url": "https://huggingface.co/datasets/SWE-bench/SWE-bench_Multilingual/tree/846e647b9f33c0b51b739d005d13d85493c9af09"}
+      ]
+    },
+    {
+      "source_id": "toolbench_full",
+      "upstream": "https://github.com/OpenBMB/ToolBench/tree/d56fdd89faf8c91fa135090b212bb9057ee5cfc2",
+      "declared_license": "Apache-2.0 with research-and-education intended use",
+      "public_access": {"permission": "permitted", "verification": "verified"},
+      "local_research_use": {"permission": "permitted", "verification": "verified", "conditions": ["research or education", "consider third-party API rights"]},
+      "model_training": {"permission": "permitted_with_conditions", "verification": "verified", "conditions": ["research or education", "consider third-party API rights"]},
+      "commercial_training": {"permission": "restricted", "verification": "verified", "conditions": ["official README states ToolBench is intended solely for research and educational purposes"]},
+      "redistribution_original": {"permission": "permitted_with_conditions", "verification": "verified", "conditions": ["Apache-2.0 notice", "research/education intended use", "third-party API content remains subject to third-party rights"]},
+      "redistribution_transformed": {"permission": "permitted_with_conditions", "verification": "verified", "conditions": ["Apache-2.0 notice", "source attribution", "research/education intended use", "third-party API content remains subject to third-party rights"]},
+      "public_payload": "hosted",
+      "evidence": [
+        {"finding": "The official pinned README licenses ToolBench under Apache-2.0, documents the SFT solution paths, and limits intended use to research and education.", "url": "https://github.com/OpenBMB/ToolBench/blob/d56fdd89faf8c91fa135090b212bb9057ee5cfc2/README.md"},
+        {"finding": "The pinned Apache-2.0 conversion reproduces the official preprocessed G1/G2/G3 solution-path file in Parquet form.", "url": "https://huggingface.co/datasets/tuandunghcmut/toolbench-v1/tree/36de9b189753ad5de276181974f97df15e8c3202"}
+      ]
+    },
     {
       "source_id": "swesmith_trajectories",
       "upstream": "https://huggingface.co/datasets/SWE-bench/SWE-smith-trajectories/tree/08e109b4a59eaeebf80e4675cd125d42e7ac99a4",

registry/sources.json

@@ -420,6 +420,27 @@
       "duplicate_risks": ["one attempt per model/task"],
       "limitations": ["No causal tool results; retained only as tier-5 diagnostics, never as trajectory SFT"]
     },
+    {
+      "id": "agentsuite_bfcl_v4",
+      "name": "AgentSuite BFCL V4 Agentic Trajectories",
+      "url": "https://huggingface.co/datasets/AgentSuite/BFCL_V4-trajectories",
+      "pinned_revision": "53730302150e4a9c24026e5f591964c500ba7f8a",
+      "available_splits": ["30 model JSONL files; web and memory categories only"],
+      "upstream_record_count": 175950,
+      "qualifying_trajectory_count": 6161,
+      "rejected_before_library": 169789,
+      "verified_record_count": 6161,
+      "reported_statistics": "Across 30 models, 6,161 web/memory rows contain valid paired calls and returned tool observations (4,832 score-1; 1,329 score-0). Rejected: 156,000 format-only answers, 13,759 rows without a complete action-observation history, and 30 malformed/unpaired histories.",
+      "schema_summary": "System tool schemas, user task, content-encoded function calls, returned tool observations, final answers, evaluator score, latency and token counts.",
+      "full_trajectory": "yes for 6,161/19,950 agentic-category rows",
+      "provenance": "30 model families evaluated through AgentSuite",
+      "verification": "Published BFCL V4 evaluator score",
+      "eligibility": "evaluation-only",
+      "license": "No license declared on the pinned trajectory dataset card",
+      "task_id": "meta.id shared across model attempts",
+      "duplicate_risks": ["one attempt per model/task", "protected BFCL evaluation overlap"],
+      "limitations": ["Raw payload redistribution and model-training permissions remain uncertain", "Evaluator results were not independently rerun"]
+    },
     {
       "id": "browsecomp_rollouts",
       "name": "BrowseComp GPT-OSS-120B Rollouts",
@@ -441,6 +462,131 @@
       "duplicate_risks": ["exactly four attempts per question"],
       "limitations": ["One malformed record", "Reference answers remain raw provenance and are excluded from model_input"]
     },
+    {
+      "id": "alfworld_replayed_experts",
+      "name": "ALFWorld Simulator-Replayed Expert Demonstrations",
+      "url": "https://huggingface.co/datasets/Joshyxwa/cp2107-textworld-trajectories",
+      "pinned_revision": "6402e311e4f0f73e79bc4ebf8fbb49e0b3b3b898",
+      "available_splits": ["alfworld/train", "alfworld/validation", "alfworld/test"],
+      "upstream_record_count": 3513,
+      "qualifying_trajectory_count": 0,
+      "blocked_record_count": 3513,
+      "verified_record_count": 3513,
+      "reported_statistics": "The card reports 3,513 successful simulator-replayed expert trajectories and 314 excluded failures.",
+      "schema_summary": "Goal, action, observation and optional reasoning blocks with deterministic simulator replay evidence.",
+      "full_trajectory": "reported yes; payload access is gated",
+      "provenance": "ALFWorld expert planner demonstrations",
+      "verification": "Publisher-reported deterministic successful replay",
+      "eligibility": "unavailable",
+      "license": "No dataset license was declared at the pinned revision; upstream ALFWorld terms do not establish rights for the separately published trajectories.",
+      "task_id": "Unavailable until gated access is accepted",
+      "duplicate_risks": ["multiple language renderings or replays per ALFWorld task"],
+      "limitations": ["The configured Hugging Face identity has not accepted this gated dataset's access terms", "No raw records were downloaded or imported"]
+    },
+    {
+      "id": "delm_deepswe",
+      "name": "DeLM DeepSWE Multi-Agent Trajectories",
+      "url": "https://huggingface.co/datasets/yuzhenm/delm-720-trajectories",
+      "pinned_revision": "15d5133454d5ecd702fa957be414c479f8afcdd0",
+      "available_splits": ["astra-ds-n2", "astra-ds-n4", "opus55-ds-n2", "opus55-ds-n4"],
+      "upstream_record_count": 360,
+      "qualifying_trajectory_count": 360,
+      "verified_record_count": 360,
+      "reported_statistics": "360 native per-agent histories imported: 330 pass the published CTRF test report and 30 fail.",
+      "schema_summary": "Codex and Claude native event logs with assistant messages, structured tool calls, returned tool observations, prompts, grading reports and completion metadata.",
+      "full_trajectory": "yes",
+      "provenance": "GPT-6 Astra and Claude Opus 5.5 agents in two- and four-agent DeLM teams",
+      "verification": "Published DeepSWE CTRF test report",
+      "eligibility": "evaluation-only",
+      "license": "MIT dataset card; benchmark tasks remain protected evaluation material",
+      "task_id": "native task directory",
+      "duplicate_risks": ["three runs per task and multiple agents per run", "shared context across agents"],
+      "limitations": ["Contains explicit benchmark anti-training canaries", "Scores were not independently replayed"]
+    },
+    {
+      "id": "delm_terminalbench4",
+      "name": "DeLM Terminal-Bench 4 Multi-Agent Trajectories",
+      "url": "https://huggingface.co/datasets/yuzhenm/delm-720-trajectories",
+      "pinned_revision": "15d5133454d5ecd702fa957be414c479f8afcdd0",
+      "available_splits": ["astra-tb-n2", "astra-tb-n4", "opus55-tb-n2", "opus55-tb-n4"],
+      "upstream_record_count": 360,
+      "qualifying_trajectory_count": 358,
+      "rejected_before_library": 2,
+      "verified_record_count": 357,
+      "reported_statistics": "358 native per-agent histories imported: 308 reward-1, 49 reward-0 and one unscored; two advertised agent records lack an importable native history/result.",
+      "schema_summary": "Codex and Claude native event logs with assistant messages, structured tool calls, returned tool observations, prompts, grader rewards and completion metadata.",
+      "full_trajectory": "yes for 358/360 advertised records",
+      "provenance": "GPT-6 Astra and Claude Opus 5.5 agents in two- and four-agent DeLM teams",
+      "verification": "Published Terminal-Bench 4 grader reward",
+      "eligibility": "evaluation-only",
+      "license": "MIT dataset card; Terminal-Bench anti-training restrictions remain applicable",
+      "task_id": "native task directory",
+      "duplicate_risks": ["three runs per task and multiple agents per run", "shared context across agents", "Terminal-Bench task overlap"],
+      "limitations": ["Contains explicit Harbor anti-training canaries", "Scores were not independently replayed"]
+    },
+    {
+      "id": "swebench_multilingual_glm5",
+      "name": "SWE-bench Multilingual GLM-5 Trajectories",
+      "url": "https://huggingface.co/datasets/AlienKevin/SWE-bench-multilingual-glm-5-trajectories",
+      "pinned_revision": "337a922eac45fc0d63271c2a50c3a7c84dae34de",
+      "available_splits": ["train"],
+      "upstream_record_count": 300,
+      "qualifying_trajectory_count": 300,
+      "verified_record_count": 300,
+      "reported_statistics": "All 300 rows map uniquely by exact problem statement to the 300-task pinned SWE-bench Multilingual release; 209 are resolved.",
+      "schema_summary": "Complete structured shell tool-call histories, observations, model, patch and resolved label.",
+      "full_trajectory": "yes",
+      "provenance": "GLM-5 software-engineering agent completions",
+      "verification": "Published resolved field",
+      "eligibility": "evaluation-only",
+      "license": "No trajectory-dataset license declared; underlying benchmark is MIT but does not license separate model outputs",
+      "task_id": "joined to official benchmark instance_id",
+      "duplicate_risks": ["same 300 protected benchmark tasks as other multilingual model sources"],
+      "limitations": ["Raw payload redistribution and model-training permissions remain uncertain", "Scores were not independently replayed"]
+    },
+    {
+      "id": "swebench_multilingual_minimax_m25",
+      "name": "SWE-bench Multilingual MiniMax-M2.5 Trajectories",
+      "url": "https://huggingface.co/datasets/AlienKevin/SWE-bench-multilingual-minimax-m2.5-trajectories",
+      "pinned_revision": "479f2e1f49c71c6d8fa778e5de945b380c073b8c",
+      "available_splits": ["train"],
+      "upstream_record_count": 299,
+      "qualifying_trajectory_count": 299,
+      "verified_record_count": 299,
+      "reported_statistics": "All 299 rows map uniquely by exact problem statement to pinned SWE-bench Multilingual task IDs; 220 are resolved.",
+      "schema_summary": "Complete structured shell tool-call histories, observations, model, patch and resolved label.",
+      "full_trajectory": "yes",
+      "provenance": "MiniMax-M2.5 software-engineering agent completions",
+      "verification": "Published resolved field",
+      "eligibility": "evaluation-only",
+      "license": "No trajectory-dataset license declared; underlying benchmark is MIT but does not license separate model outputs",
+      "task_id": "joined to official benchmark instance_id",
+      "duplicate_risks": ["299 protected benchmark tasks overlap other multilingual model sources"],
+      "limitations": ["Raw payload redistribution and model-training permissions remain uncertain", "Scores were not independently replayed"]
+    },
+    {
+      "id": "toolbench_full",
+      "name": "ToolBench Full Valid Solution Paths",
+      "url": "https://github.com/OpenBMB/ToolBench",
+      "transport_url": "https://huggingface.co/datasets/tuandunghcmut/toolbench-v1",
+      "pinned_revision": "36de9b189753ad5de276181974f97df15e8c3202",
+      "related_code_revision": "d56fdd89faf8c91fa135090b212bb9057ee5cfc2",
+      "available_splits": ["G1/G2/G3 DFS train"],
+      "upstream_record_count": 187542,
+      "qualifying_trajectory_count": 73121,
+      "rejected_before_library": 343,
+      "verified_record_count": 0,
+      "reported_statistics": "The official preprocessed release has 187,542 causal-prefix rows across 73,464 exact queries. Maximum-prefix grouping reconstructs 73,121 complete action-observation paths; 343 queries have no qualifying complete path.",
+      "schema_summary": "System tool definitions, user query, alternating Thought/Action/Action Input and function observations, ending in Finish.",
+      "full_trajectory": "yes after deterministic maximum-prefix grouping",
+      "provenance": "ChatGPT-generated ToolBench DFS paths",
+      "verification": "Official preprocessing includes only valid_data paths; search-win/task-success labels are absent",
+      "eligibility": "conditionally-usable",
+      "license": "Official Apache-2.0 release with research/education intent; third-party API content retains third-party rights",
+      "task_id": "stable exact-query hash",
+      "duplicate_risks": ["187,542 source rows are overlapping causal prefixes", "query overlap with the 11 checked-in examples"],
+      "limitations": ["Valid-data inclusion is not a semantic success score", "Exact model version is not provided"]
+    },
     {
       "id": "swesmith_trajectories",
       "name": "SWE-smith Trajectories",

staging_report.json

@@ -1,26 +1,12 @@
 {
-  "hosted_shards": 50,
-  "hosted_sources": [
-    "mind2web",
-    "nvidia_open_swe_traces",
-    "official_tau3",
-    "swebench_multilingual_trajectories",
-    "swebench_verified_trajectories",
-    "swegym_openhands",
-    "swesmith_trajectories",
-    "terminalbench_trajectories",
-    "terminaltraj",
-    "toolbench"
-  ],
-  "hosted_trajectories": 136373,
-  "inventory_shards": 56,
-  "inventory_trajectories": 155278,
-  "metadata_only_sources": [
-    "agentsuite_tau2",
-    "browsecomp_rollouts",
-    "swegym_sampled"
-  ],
-  "metadata_only_trajectories": 18905,
+  "base_hf_revision": "e9214095fdcb9ce6a7ddaefb09426ae531e0221d",
+  "hosted_trajectories": 210212,
+  "inventory_trajectories": 235876,
+  "metadata_only_trajectories": 25664,
+  "new_hosted_shards": 16,
+  "new_inventory_shards": 19,
+  "old_shards_reused": 50,
+  "release": "v0.4.0",
   "schema_version": 1,
   "scores_recomputed": false
 }

structurally_eligible.index.jsonl.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7abe69bdb9cd3d07b48a282c0bf371f7bcff26776ad3caba307e22b45faf40f8
-size 4326891
+oid sha256:81e23e4a128d27e308e0c8488ee4172947bb087cf37e7c6b066c014ce07e05af
+size 7062131

successful.index.jsonl.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:30b4ec2dd4faacb804e78a2826df649af7737807f542ddc6f667e314ae08c3e1
-size 1561257
+oid sha256:ef34de7b52e947c40a6a55c7be1aebbf28823e1001207eb3d6bc7493ea86cd74
+size 1726006

task_splits.jsonl.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f932f6d111636fa2306d1b2fef94a6a5b98e9fdab426c95f0da78fced367f776
-size 582841
+oid sha256:baff2285f79e5fe6109bfd3bc265ae6b77bfe3c7e1f9c842e971284d0a5cd942
+size 1676404

training_eligible.index.jsonl.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:fd0619062d95f44824710e32922c7fc330c44c367f41c2adcd3fb2002308a00f
-size 3426523
+oid sha256:760ad47b9291eb25c65239df89e99533ddb8aceb557ef55b187873afa7ce2fc5
+size 5953469

unknown.index.jsonl.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1a11f7e31b6bf29ee513709a4118ad99823a8634b6f0b3f5ecd8cf605b58cc45
-size 663458
+oid sha256:33cffe8ee963d5807e2aea69989f449e09df39c8de4b907a79ea20ce7838ce48
+size 2958743

unsuccessful.index.jsonl.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a0e83f5fa6f85137c64534acc9f400663ac458d44fb88758db407007ae259273
-size 2233251
+oid sha256:06c9cdcc75e0e7a93ec7dc26f67060e3f26c509733a856ff17fbd6c990243584
+size 2276641