noahmclaughlin commited on
Commit
4cd62fa
·
verified ·
1 Parent(s): 2f6fa74

Publish verified eleven-model Predict and Learn results

Browse files
README.md CHANGED
@@ -32,30 +32,15 @@ configs:
32
 
33
  This release contains **165 prompt windows representing 39,828 submissions**. One row is one aggregate prompt window, not one person or submission. The outcomes describe the Mixed Messages audience, not objectively correct meanings or a representative sample of the general population.
34
 
35
- **Learn the Room** is the interactive track: a model chooses historical examples to inspect, then predicts the audience's response to a new target. **Distribution Prediction** measures the same prediction ability without examples and supplies its essential control. The original RTR winning-quadrant benchmark is also supported by the same data, with a separate scoring protocol.
36
 
37
- [Code and runners](https://github.com/1999labs/RTR-bench) · [Learn the Room protocol](https://github.com/1999labs/RTR-bench/blob/644d481d335af9a632c7ed0599108b13e4a5653e/docs/learn-room.md) · [Distribution protocol](https://github.com/1999labs/RTR-bench/blob/644d481d335af9a632c7ed0599108b13e4a5653e/docs/distribution.md) · [First final results](https://github.com/1999labs/RTR-bench/blob/644d481d335af9a632c7ed0599108b13e4a5653e/results/learn-room/z-ai-glm-5-3-flash-e9d0327447ef1856/README.md)
38
 
39
- ## Tasks
40
 
41
- ### Learn the Room
42
 
43
- Each episode starts with a target prompt and a fresh conversation. The target's recorded crowd response remains hidden from the model.
44
-
45
- 1. The model can request a catalog of the **90 reference prompt names and IDs**, without their labels.
46
- 2. It selects a reference to inspect and receives that reference's quadrant distribution and submission count.
47
- 3. It can use that observation to choose another reference, up to **three distinct inspections**, or stop early.
48
- 4. It submits four probabilities for the target, summing to one.
49
-
50
- The default environment allows eight actions and eight model turns. The model receives no target score or correctness feedback during an episode. Learning is within the conversation; there is no weight update or memory carried between target episodes.
51
-
52
- Report the agent's score **and its paired gain against no examples**. Random-example and text-similarity conditions are comparison controls, not separate benchmark tracks. The similarity control uses a frozen lexical TF–IDF method, not an embedding model. Agent selection and evidence use are evaluated together; differences also include the agent's extra inference compute and multi-turn instructions.
53
-
54
- The first final run used the explicit `transfer-check-v1` policy, which asks the model to assess whether historical evidence transfers on each axis and to discount weak analogies. The original CLI policy remains the default; always identify the policy and settings when comparing results.
55
-
56
- ### Distribution Prediction
57
-
58
- Given only the target, audience description, and quadrant definitions, estimate the fraction of audience responses in each quadrant. Example output format, with illustrative values:
59
 
60
  ```json
61
  {
@@ -66,11 +51,26 @@ Given only the target, audience description, and quadrant definitions, estimate
66
  }
67
  ```
68
 
69
- The model outputs proportions, not a single winning label or its own preference. Learn the Room's no-example condition uses the same messages and scorer, so it can supply this result without duplicate model calls.
 
 
70
 
71
- ### Original RTR: winning-quadrant prediction
72
 
73
- Predict a quadrant with the largest crowd count; a second question asks for the model's own interpretation. Each question uses four cyclic option orders. The headline score is the percentage of crowd-prediction windows answered correctly after a unique modal vote; unresolved model-vote ties score zero. Accept **any crowd co-winner**, not only the legacy single `dominant` field. The original benchmark uses all 165 windows; the 90/30/45 partition belongs to the two new tracks. See the [original protocol](https://github.com/1999labs/RTR-bench/blob/644d481d335af9a632c7ed0599108b13e4a5653e/README.md) for complete failure, tie, and consistency rules.
 
 
 
 
 
 
 
 
 
 
 
 
 
74
 
75
  ## Data and collection
76
 
@@ -97,11 +97,11 @@ There is **no training split** in this protocol. All reference windows precede a
97
 
98
  **Target labels are public in these files for scoring, but must not be passed into model prompts or unrestricted agent retrieval.** Only the reference catalog and permitted inspected reference outcomes may enter the agent's context. Development outcomes never become reference evidence. Loading a dataset row is not itself a compliant evaluation environment; use the linked runner to enforce that boundary.
99
 
100
- The data were public and used in the original RTR benchmark before these new partitions were defined. The final split has now been evaluated with GLM 5.3 Flash. It is a holdout from that development process, not a claim of unseen model training data. Any later tuning informed by these outcomes must disclose that exposure.
101
 
102
  ## Loading
103
 
104
- After publication, load all three splits with Hugging Face Datasets:
105
 
106
  ```python
107
  from datasets import load_dataset
@@ -143,47 +143,89 @@ Pin a Hugging Face commit with `revision="<dataset-commit>"` when publishing res
143
 
144
  There are three crowd-winner ties: Republicans (day 60), Police Departments (68), and Freedom (83). The original source contains a legacy single winner and rounded winning share; this export derives **all co-winners and full-precision shares from integer counts**. Conviction, polarization, and centroid fields are descriptive and do not enter the current scores.
145
 
146
- ## Scoring the two distribution tasks
147
 
148
  For target probabilities `p` and observed shares `y`, the error is the sum of `(p[q] - y[q])²` over the four quadrants. Average errors equally across declared targets, then compute:
149
 
150
  ```text
151
  score = 100 × (1 − mean_error / 2)
152
- agent_gain = agent_score − no_example_score
153
  ```
154
 
155
  A score of 100 is an exact distribution match. These scores are **not percentage accuracy**, and 50 is not a chance baseline. Report uniform and constant-reference predictions on the same targets. The constant baseline averages the 90 reference-window distributions equally, rather than pooling their submissions.
156
 
157
  Predictions must contain exactly the four canonical keys and finite numbers between zero and one. Sums within 0.001 of one are normalized; other invalid or missing terminal predictions receive zero and remain in the denominator. Pending targets keep a run incomplete. Saved requests, responses, costs, and traces support offline replay. The protocol documents the bounded transport retries and agent action recovery rules.
158
 
159
- ## Initial final results
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
160
 
161
- One GLM 5.3 Flash run through Relace, using all 45 evaluation targets, temperature 0, a 4,000-token allowance, seed 7, and agent policy `transfer-check-v1`:
162
 
163
- | Result | Score / 100 | Gain vs no examples |
164
- |---|---:|---:|
165
- | **Learn the Room: agent** | **85.57** | **-2.37** |
166
- | **Distribution Prediction / no examples** | **87.93** | — |
167
- | Control: random examples | 86.47 | -1.46 |
168
- | Control: text-similar examples | 85.43 | -2.51 |
169
- | Constant reference distribution | 85.56 | — |
170
- | Uniform distribution | 85.39 | — |
171
 
172
- All 180 predictions were valid. The agent's exploratory paired 95% interval against no examples was **-5.14 to +0.12 points**. This run demonstrated no average benefit from gathering examples; a score in the mid-80s is close to simple baselines on this split. The interval resamples target entity groups, is not corrected for multiple comparisons, and does not measure variability across repeated model runs. These are one model's results on one audience, not a general model ranking. See the linked final report for complete records and settings.
 
 
 
 
 
 
 
 
 
 
173
 
174
  ## Running the benchmark
175
 
176
- The data are usable independently of any model provider. The current reference CLI uses OpenRouter for paid model calls; support for other endpoints is not yet implemented. To verify the environment locally with **zero API calls**, use Node.js 22.18 or newer:
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
177
 
178
  ```bash
179
- git clone https://github.com/1999labs/RTR-bench.git
180
  cd RTR-bench
181
- git checkout 644d481d335af9a632c7ed0599108b13e4a5653e
182
  npm ci
183
- node src/cli.ts learn-room --fixture tests/fixtures/learn-room-four-strategies.json --strategy all --limit 5 --max-tokens 4000 --agent-policy transfer-check-v1
 
184
  ```
185
 
186
- The fixture results are synthetic and are not model performance. The code repository contains the scoring, interactive environment, CLI, tests, and report generation. This Hub repository distributes data and documentation; it does not automatically execute or host an evaluation service.
 
 
187
 
188
  ## Limitations and intended use
189
 
@@ -202,6 +244,7 @@ Use this release to study crowd prediction, evidence selection, and audience ada
202
  - `data/*.jsonl`: the 165 derived rows in the three frozen partitions.
203
  - `source/windows.json`: byte-identical original export, including collection metadata.
204
  - `source/split-v1.json`: byte-identical chronological split manifest.
 
205
  - `release.json`: source repository commit, source hashes, row counts, and SHA-256 checksums for package files.
206
  - `LICENSE`, `LICENSE-CC0`, `LICENSE-MIT`: license scope and full texts.
207
  - `CITATION.cff`: machine-readable citation.
 
32
 
33
  This release contains **165 prompt windows representing 39,828 submissions**. One row is one aggregate prompt window, not one person or submission. The outcomes describe the Mixed Messages audience, not objectively correct meanings or a representative sample of the general population.
34
 
35
+ **Predict** tests whether a model can predict the audience's response without examples. **Learn** lets it choose and inspect up to three historical examples before making the same prediction. Both return four probabilities and receive a distribution score out of 100. **Learn Gain = Learn score − Predict score** shows whether the complete evidence-gathering workflow helped.
36
 
37
+ [Code and CLI](https://github.com/1999labs/RTR-bench) · [Predict protocol](https://github.com/1999labs/RTR-bench/blob/main/docs/distribution.md) · [Learn protocol](https://github.com/1999labs/RTR-bench/blob/main/docs/learn-room.md) · [Verified final results](https://github.com/1999labs/RTR-bench/tree/main/results/final-8000)
38
 
39
+ ## The two tests
40
 
41
+ ### Predict
42
 
43
+ The model receives a target word or phrase, a description of the audience, and the four quadrant definitions. It predicts the fraction of audience responses in each quadrant, without historical examples or target outcomes. For example:
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
44
 
45
  ```json
46
  {
 
51
  }
52
  ```
53
 
54
+ These are proportions, not a single winning label or the model's own preference.
55
+
56
+ ### Learn
57
 
58
+ The model receives the same prediction task, but can gather evidence first:
59
 
60
+ 1. Request a catalog of the **90 reference prompt names and IDs**, without their outcomes.
61
+ 2. Choose a reference to inspect, revealing its quadrant distribution and submission count.
62
+ 3. Use that observation to choose another reference, up to **three distinct inspections**, or stop early.
63
+ 4. Submit four probabilities for the target, summing to one.
64
+
65
+ Each target starts a fresh conversation. The model receives no target score or correctness feedback, and nothing carries over to the next target. Learning happens through the conversation; model weights are not updated. The environment permits eight actions and eight model turns.
66
+
67
+ The public CLI uses the `transfer-check-v1` policy: assess whether each example transfers on each axis and discount weak analogies. Predict and Learn default to **8,000 output tokens per response** in the current release. Record the model, provider, policy, token allowance, and other settings when reporting results.
68
+
69
+ Random-example and text-similarity conditions are optional comparison controls, not additional tracks. Similarity uses a frozen lexical TF–IDF method. Learn Gain measures the whole workflow, including example selection, additional inference, and answer completion; it does not isolate the causal effect of examples alone.
70
+
71
+ ### Original winning-quadrant benchmark
72
+
73
+ The earlier RTR benchmark predicts a winning quadrant using cyclic answer orders. Its score is percentage accuracy over all 165 windows and is **not comparable** to the new distribution scores. Its [archived results and protocol](https://github.com/1999labs/RTR-bench/tree/main/results/history/original-benchmark) remain available. Predict and Learn use the frozen 90/30/45 partitions below.
74
 
75
  ## Data and collection
76
 
 
97
 
98
  **Target labels are public in these files for scoring, but must not be passed into model prompts or unrestricted agent retrieval.** Only the reference catalog and permitted inspected reference outcomes may enter the agent's context. Development outcomes never become reference evidence. Loading a dataset row is not itself a compliant evaluation environment; use the linked runner to enforce that boundary.
99
 
100
+ The data were public and used in the original RTR benchmark before these new partitions were defined. The final split has now been evaluated with multiple models. It is a holdout from the declared development process, not a claim of unseen model training data. Any later tuning informed by these outcomes must disclose that exposure.
101
 
102
  ## Loading
103
 
104
+ Load all three splits with Hugging Face Datasets:
105
 
106
  ```python
107
  from datasets import load_dataset
 
143
 
144
  There are three crowd-winner ties: Republicans (day 60), Police Departments (68), and Freedom (83). The original source contains a legacy single winner and rounded winning share; this export derives **all co-winners and full-precision shares from integer counts**. Conviction, polarization, and centroid fields are descriptive and do not enter the current scores.
145
 
146
+ ## Scoring Predict and Learn
147
 
148
  For target probabilities `p` and observed shares `y`, the error is the sum of `(p[q] - y[q])²` over the four quadrants. Average errors equally across declared targets, then compute:
149
 
150
  ```text
151
  score = 100 × (1 − mean_error / 2)
152
+ Learn Gain = Learn score − Predict score
153
  ```
154
 
155
  A score of 100 is an exact distribution match. These scores are **not percentage accuracy**, and 50 is not a chance baseline. Report uniform and constant-reference predictions on the same targets. The constant baseline averages the 90 reference-window distributions equally, rather than pooling their submissions.
156
 
157
  Predictions must contain exactly the four canonical keys and finite numbers between zero and one. Sums within 0.001 of one are normalized; other invalid or missing terminal predictions receive zero and remain in the denominator. Pending targets keep a run incomplete. Saved requests, responses, costs, and traces support offline replay. The protocol documents the bounded transport retries and agent action recovery rules.
158
 
159
+ ## Results
160
+
161
+ <!-- FINAL_RESULTS_START -->
162
+ All eleven models use the same 45 evaluation targets and 8,000-token response allowance. Scores measure distribution agreement, not percentage accuracy. Counts show valid answers; failed terminal answers score zero and were not rerolled. [Verification, intervals, and raw records](https://github.com/1999labs/RTR-bench/tree/main/results/final-8000).
163
+
164
+ | Model | Provider | Predict | Predict Cost | Learn | Learn Gain | Learn Cost | Total Cost |
165
+ |---|---|---:|---:|---:|---:|---:|---:|
166
+ | Historical-average baseline | No model | 85.56 | — | 85.56 | — | — | — |
167
+ | glm-5.3-flash | Relace (fp4) | 87.66<br>45/45 valid | $0.00223 | 86.00<br>45/45 valid | −1.66 | $0.02315 | $0.03363 |
168
+ | glm-5.3 | Wafer (fp4) | 86.04<br>44/45 valid | $0.00551 | 82.87<br>44/45 valid | −3.17 | $0.03842 | $0.05741 |
169
+ | minimax-m3.1-flash | Inferact vLLM (nvfp4) | 86.95<br>45/45 valid | $0.01748 | 84.24<br>45/45 valid | −2.71 | $0.06345 | $0.11698 |
170
+ | mimo-v2.6-flash | Xiaomi (fp8) | 86.98<br>45/45 valid | $0.00390 | 83.53<br>45/45 valid | −3.45 | ≥$0.02217 | ≥$0.04005 |
171
+ | deepseek-v4.1-flash-1017 | Sail Research (fp4) | 74.82<br>38/45 valid | ≥$0.02286 | 86.75<br>45/45 valid | +11.93 | ≥$0.02614 | ≥$0.10495 |
172
+ | hy3 | Tencent (fp8) | 86.14<br>45/45 valid | $0.03331 | 83.58<br>45/45 valid | −2.56 | $0.12577 | ≥$0.27726 |
173
+ | gemini-3.8-flash | Google AI Studio | 86.48<br>45/45 valid | $0.10883 | 87.94<br>45/45 valid | +1.46 | $0.47906 | $0.89724 |
174
+ | gpt-6-luna | OpenAI (Flex) | 85.99<br>44/45 valid | $0.00286 | 82.75<br>44/45 valid | −3.24 | $0.01475 | $0.02883 |
175
+ | claude-haiku-5.5 | Anthropic | 86.35<br>45/45 valid | $0.00504 | 86.79<br>45/45 valid | +0.44 | $0.05925 | $0.07958 |
176
+ | inkling-small | DeepInfra (fp4) | 86.57<br>45/45 valid | $0.02011 | 82.06<br>44/45 valid | −4.51 | $0.12605 | $0.22916 |
177
+ | gpt-oss-120b | DekaLLM (bf16) | 86.53<br>45/45 valid | $0.00291 | 82.84<br>45/45 valid | −3.68 | ≥$0.03264 | ≥$0.04415 |
178
 
179
+ Predict Cost and Learn Cost cover their respective 45-target tracks. Total Cost also includes the random and text-similarity controls. A **≥** marks a known subtotal with unresolved charges; conservative reservations remain in the spending ledger. The plots use those same subtotals, so affected cost positions and the purple Pareto frontier are provisional. The dashed guide shows the best observed score available at each budget; GLM Flash is the sole Predict frontier model. Higher scores and lower costs are better. In the combined plot, **P = Predict** and **L = Learn**.
180
 
181
+ ![RTR-Bench (Predict)](results/predict-scatter.png)
 
 
 
 
 
 
 
182
 
183
+ ![RTR-Bench (Learn)](results/learn-scatter.png)
184
+
185
+ ![RTR-Bench (Learn + Predict)](results/predict-learn-scatter.png)
186
+
187
+ ### Findings
188
+
189
+ **glm-5.3-flash led Predict at 87.66; gemini-3.8-flash led Learn at 87.94.** 10 of 11 models beat the 85.56 historical-average baseline on Predict, and 4 did on Learn. That baseline uses the same reference-derived distribution for every target; uniform predictions score 85.39. Scores in the mid-80s therefore indicate closeness to a simple baseline, not 80%-plus classification accuracy.
190
+
191
+ Learn scored lower than Predict for 8 of 11 models, so gathering examples did not consistently improve this workflow. DeepSeek's +11.93-point gain includes an answer-completion difference: 38/45 valid Predict answers versus 45/45 Learn answers. It should not be read as a pure benefit from historical evidence. These are single runs on 45 targets from one audience; small rank differences and evidence-only explanations remain uncertain.
192
+
193
+ <!-- FINAL_RESULTS_END -->
194
 
195
  ## Running the benchmark
196
 
197
+ Install **Node.js 22.18 or newer**. It includes `npx`, which downloads and starts the npm package; you do not need to clone the code. Open Terminal in a folder for your results and create a private `.env` file there containing `OPENROUTER_API_KEY=your-key`. Add credits to that OpenRouter account.
198
+
199
+ This small, paid development check runs both tests on five targets:
200
+
201
+ ```bash
202
+ npx rtr-bench@0.2.2 read-the-room z-ai/glm-5.3-flash \
203
+ --provider Relace --split development --limit 5 --max-usd 0.25
204
+ ```
205
+
206
+ | Command | What it runs |
207
+ |---|---|
208
+ | `npx rtr-bench@0.2.2 read-the-room <model>` | Predict and Learn, with Learn Gain |
209
+ | `npx rtr-bench@0.2.2 predict <model>` | Predict only |
210
+ | `npx rtr-bench@0.2.2 learn <model>` | Learn only; no Learn Gain without matching Predict |
211
+
212
+ Keep the provider and run-setting flags when replacing the command. To reproduce the full final protocol, use `--split evaluation --limit 45 --controls --max-tokens 8000`; select the reported model/provider and an appropriate spending cap. The cap limits spending rather than setting a fixed price. It can stop a run incomplete. `--controls` adds random and text-similar examples to the two headline tracks.
213
+
214
+ The runner automatically saves reports, predictions, evidence traces, and spending records under `results/learn-room/`. Repeat an identical command to resume its saved run. Terminal prints the path to `report.html`, which you can open in a browser. Changing settings creates a separate run.
215
+
216
+ For a **synthetic check with zero model calls**, use the GitHub source:
217
 
218
  ```bash
219
+ git clone https://github.com/1999labs/RTR-bench
220
  cd RTR-bench
 
221
  npm ci
222
+ node src/cli.ts read-the-room \
223
+ --fixture tests/fixtures/learn-room-four-strategies.json --controls --limit 5
224
  ```
225
 
226
+ Fixture scores demonstrate the pipeline, not model performance. For source-based real evaluations, replace `npx rtr-bench@0.2.2` with `node src/cli.ts` in the commands above. See the [full Quickstart](https://github.com/1999labs/RTR-bench#quickstart).
227
+
228
+ This Hub repository distributes the frozen data, documentation, and published result summaries. It does not automatically run evaluations. The reference CLI uses OpenRouter; direct connections to other endpoints are not yet implemented.
229
 
230
  ## Limitations and intended use
231
 
 
244
  - `data/*.jsonl`: the 165 derived rows in the three frozen partitions.
245
  - `source/windows.json`: byte-identical original export, including collection metadata.
246
  - `source/split-v1.json`: byte-identical chronological split manifest.
247
+ - `results/`: verified result summary, three score-versus-cost charts, and chart source metadata. Full request and response traces remain in GitHub.
248
  - `release.json`: source repository commit, source hashes, row counts, and SHA-256 checksums for package files.
249
  - `LICENSE`, `LICENSE-CC0`, `LICENSE-MIT`: license scope and full texts.
250
  - `CITATION.cff`: machine-readable citation.
release.json CHANGED
@@ -1,8 +1,9 @@
1
  {
2
  "packageVersion": "rtr-bench-hf-v1",
 
3
  "repository": "noahmclaughlin/RTR-bench",
4
  "sourceRepository": "https://github.com/1999labs/RTR-bench",
5
- "sourceCommit": "644d481d335af9a632c7ed0599108b13e4a5653e",
6
  "sourceGeneratedAt": "2026-11-12T06:20:26.692Z",
7
  "datasetSha256": "ec9d7582b5610a24f4e14dbbb05d13c636b692ec9e133b9381db0fdf483ffb2b",
8
  "splitSha256": "a165c72f7557f8128253148eb28e6906b39e80aec12e45c679fddde62fa31421",
@@ -198,9 +199,41 @@
198
  }
199
  },
200
  "files": {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
201
  "README.md": {
202
- "bytes": 15382,
203
- "sha256": "d336757780cc543ebb18a88d182680d7b27e31a734e564ee99b196e9ef8bf067"
204
  },
205
  "LICENSE": {
206
  "bytes": 760,
 
1
  {
2
  "packageVersion": "rtr-bench-hf-v1",
3
+ "cliVersion": "0.2.2",
4
  "repository": "noahmclaughlin/RTR-bench",
5
  "sourceRepository": "https://github.com/1999labs/RTR-bench",
6
+ "sourceCommit": "7a8cab7e0540deb44711cd35cc2ee3ef13f8b67d",
7
  "sourceGeneratedAt": "2026-11-12T06:20:26.692Z",
8
  "datasetSha256": "ec9d7582b5610a24f4e14dbbb05d13c636b692ec9e133b9381db0fdf483ffb2b",
9
  "splitSha256": "a165c72f7557f8128253148eb28e6906b39e80aec12e45c679fddde62fa31421",
 
199
  }
200
  },
201
  "files": {
202
+ "results/predict-scatter.png": {
203
+ "bytes": 70421,
204
+ "sha256": "ba796076f40f72fda72cdc89d0e66601973b67c75facde019979b9c3f39ec1e0"
205
+ },
206
+ "results/learn-scatter.png": {
207
+ "bytes": 69311,
208
+ "sha256": "f2ab151e13f4d8ab208bcff38071135d1139a213be71103a45bd1dfd6a2f6062"
209
+ },
210
+ "results/predict-learn-scatter.png": {
211
+ "bytes": 199909,
212
+ "sha256": "8251a848e0404824deae0ea671399d254c6ac60a4cd943480690f9274b7218d7"
213
+ },
214
+ "results/predict-scatter.svg": {
215
+ "bytes": 54158,
216
+ "sha256": "096df77ca78c31f41f1987c08fd51a44e302b3bcdc692948c256649bd1db9f9a"
217
+ },
218
+ "results/learn-scatter.svg": {
219
+ "bytes": 53154,
220
+ "sha256": "dd2d15bc062e66b8e465ac57aa1961495cefa3c66f83c770b35dab402173820b"
221
+ },
222
+ "results/predict-learn-scatter.svg": {
223
+ "bytes": 83719,
224
+ "sha256": "684730a508b633422cd4bfe6f5d15abbc7dfe05126b8deee3240b1c7c8c15448"
225
+ },
226
+ "results/predict-learn-scatter.json": {
227
+ "bytes": 11955,
228
+ "sha256": "eb56b68a172ed909c9a5fd090ac0ef001111e57dd4b871578d075a8fa4781f1a"
229
+ },
230
+ "results/verified-results.json": {
231
+ "bytes": 37810,
232
+ "sha256": "2fd106239430d3679e794ba119b29f87a8bc47e9981987b782a487b35d4b9c52"
233
+ },
234
  "README.md": {
235
+ "bytes": 18752,
236
+ "sha256": "f61ddf0c9041596031f6f29ff9fe6b5582e98f0b6153403838aaec45cf8639b0"
237
  },
238
  "LICENSE": {
239
  "bytes": 760,
results/learn-scatter.png ADDED

Git LFS Details

  • SHA256: f2ab151e13f4d8ab208bcff38071135d1139a213be71103a45bd1dfd6a2f6062
  • Pointer size: 130 Bytes
  • Size of remote file: 69.3 kB
results/learn-scatter.svg ADDED
results/predict-learn-scatter.json ADDED
@@ -0,0 +1,486 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "matplotlibVersion": "3.11.2",
3
+ "verifiedResultsSha256": "2fd106239430d3679e794ba119b29f87a8bc47e9981987b782a487b35d4b9c52",
4
+ "protocol": {
5
+ "partition": "evaluation",
6
+ "targets": [
7
+ 130,
8
+ 131,
9
+ 132,
10
+ 133,
11
+ 134,
12
+ 135,
13
+ 136,
14
+ 137,
15
+ 138,
16
+ 139,
17
+ 140,
18
+ 141,
19
+ 142,
20
+ 143,
21
+ 144,
22
+ 145,
23
+ 146,
24
+ 147,
25
+ 148,
26
+ 149,
27
+ 150,
28
+ 151,
29
+ 152,
30
+ 153,
31
+ 154,
32
+ 155,
33
+ 156,
34
+ 157,
35
+ 158,
36
+ 159,
37
+ 160,
38
+ 161,
39
+ 162,
40
+ 163,
41
+ 164,
42
+ 165,
43
+ 166,
44
+ 167,
45
+ 168,
46
+ 169,
47
+ 170,
48
+ 171,
49
+ 172,
50
+ 173,
51
+ 174
52
+ ],
53
+ "reference": [
54
+ 1,
55
+ 2,
56
+ 3,
57
+ 4,
58
+ 5,
59
+ 6,
60
+ 7,
61
+ 8,
62
+ 9,
63
+ 10,
64
+ 11,
65
+ 12,
66
+ 13,
67
+ 14,
68
+ 15,
69
+ 16,
70
+ 17,
71
+ 18,
72
+ 19,
73
+ 20,
74
+ 21,
75
+ 22,
76
+ 23,
77
+ 24,
78
+ 25,
79
+ 28,
80
+ 29,
81
+ 30,
82
+ 31,
83
+ 35,
84
+ 36,
85
+ 37,
86
+ 38,
87
+ 39,
88
+ 42,
89
+ 43,
90
+ 44,
91
+ 45,
92
+ 46,
93
+ 47,
94
+ 48,
95
+ 49,
96
+ 50,
97
+ 51,
98
+ 52,
99
+ 53,
100
+ 54,
101
+ 55,
102
+ 56,
103
+ 57,
104
+ 58,
105
+ 59,
106
+ 60,
107
+ 61,
108
+ 62,
109
+ 63,
110
+ 64,
111
+ 65,
112
+ 66,
113
+ 67,
114
+ 68,
115
+ 69,
116
+ 70,
117
+ 71,
118
+ 72,
119
+ 73,
120
+ 74,
121
+ 77,
122
+ 78,
123
+ 79,
124
+ 80,
125
+ 81,
126
+ 82,
127
+ 83,
128
+ 84,
129
+ 85,
130
+ 86,
131
+ 87,
132
+ 88,
133
+ 89,
134
+ 90,
135
+ 91,
136
+ 92,
137
+ 93,
138
+ 94,
139
+ 95,
140
+ 96,
141
+ 97,
142
+ 98,
143
+ 99
144
+ ],
145
+ "datasetHash": "ec9d7582b5610a24f4e14dbbb05d13c636b692ec9e133b9381db0fdf483ffb2b",
146
+ "splitHash": "a165c72f7557f8128253148eb28e6906b39e80aec12e45c679fddde62fa31421",
147
+ "strategies": [
148
+ "none",
149
+ "random",
150
+ "similar",
151
+ "agent"
152
+ ],
153
+ "budget": 3,
154
+ "seed": 7,
155
+ "maxModelCalls": 8,
156
+ "maxActions": 8,
157
+ "maxTokens": 8000,
158
+ "maxAttempts": 3,
159
+ "timeoutMs": 60000,
160
+ "agentPolicy": "transfer-check-v1",
161
+ "temperature": "0 where supported; provider default otherwise",
162
+ "reasoning": "provider defaults, unchanged from development"
163
+ },
164
+ "costNote": "Unknown charges make some plotted costs lower bounds and the Pareto frontier provisional. Controls are excluded.",
165
+ "runs": [
166
+ {
167
+ "model": "glm-5.3-flash",
168
+ "provider": "relace/fp4",
169
+ "source": "results/final-8000/z-ai__glm-5.3-flash/z-ai-glm-5-3-flash-a3721d86cb3fffd7/report.json",
170
+ "sourceSha256": "64389da96134a8d91d3230f9861dc4438f69d72775cf8447345ac7f25f1e6547",
171
+ "points": [
172
+ {
173
+ "track": "Predict",
174
+ "score": 87.65682697223357,
175
+ "costUsd": 0.0022276652999999994,
176
+ "costIsLowerBound": false,
177
+ "valid": 45
178
+ },
179
+ {
180
+ "track": "Learn",
181
+ "score": 85.99932573408472,
182
+ "costUsd": 0.02314631309999999,
183
+ "costIsLowerBound": false,
184
+ "valid": 45
185
+ }
186
+ ]
187
+ },
188
+ {
189
+ "model": "glm-5.3",
190
+ "provider": "wafer/us",
191
+ "source": "results/final-8000/z-ai__glm-5.3/z-ai-glm-5-3-df181956130e550d/report.json",
192
+ "sourceSha256": "8f5320e0a6668086adc6877e627d4424997e3bda58cdd0a76da30a0308b9ebb9",
193
+ "points": [
194
+ {
195
+ "track": "Predict",
196
+ "score": 86.03929052235672,
197
+ "costUsd": 0.005509304100000001,
198
+ "costIsLowerBound": false,
199
+ "valid": 44
200
+ },
201
+ {
202
+ "track": "Learn",
203
+ "score": 82.86617668804384,
204
+ "costUsd": 0.03842064060000001,
205
+ "costIsLowerBound": false,
206
+ "valid": 44
207
+ }
208
+ ]
209
+ },
210
+ {
211
+ "model": "minimax-m3.1-flash",
212
+ "provider": "inferact-vllm/nvfp4",
213
+ "source": "results/final-8000/minimax__minimax-m3.1-flash/minimax-minimax-m3-1-flash-8985e19160d564f2/report.json",
214
+ "sourceSha256": "caf58ed54159069b5f8e1a5a550e19531d9e591b8a202da2a7c9f463c93d7bc1",
215
+ "points": [
216
+ {
217
+ "track": "Predict",
218
+ "score": 86.95313994076243,
219
+ "costUsd": 0.017481899999999995,
220
+ "costIsLowerBound": false,
221
+ "valid": 45
222
+ },
223
+ {
224
+ "track": "Learn",
225
+ "score": 84.24219981765467,
226
+ "costUsd": 0.06344557500000003,
227
+ "costIsLowerBound": false,
228
+ "valid": 45
229
+ }
230
+ ]
231
+ },
232
+ {
233
+ "model": "mimo-v2.6-flash",
234
+ "provider": "xiaomi/fp8",
235
+ "source": "results/final-8000/xiaomi__mimo-v2.6-flash/xiaomi-mimo-v2-6-flash-0e9297795e7f3f19/report.json",
236
+ "sourceSha256": "a46392cd45d615ef1816ff0ce8f4acb0c8de5531c8649561b32ba42676e15da9",
237
+ "points": [
238
+ {
239
+ "track": "Predict",
240
+ "score": 86.97924213353758,
241
+ "costUsd": 0.00389634,
242
+ "costIsLowerBound": false,
243
+ "valid": 45
244
+ },
245
+ {
246
+ "track": "Learn",
247
+ "score": 83.52752701775555,
248
+ "costUsd": 0.022174073600000014,
249
+ "costIsLowerBound": true,
250
+ "valid": 45
251
+ }
252
+ ]
253
+ },
254
+ {
255
+ "model": "deepseek-v4.1-flash-1017",
256
+ "provider": "sail-research/fp4/us",
257
+ "source": "results/final-8000/deepseek__deepseek-v4.1-flash-1017/deepseek-deepseek-v4-1-flash-1017-08a69325bd0d847d/report.json",
258
+ "sourceSha256": "c5110037eaac632a96906c892314b92fd20cf18e7be9de0ce16a71f36c3c43bf",
259
+ "points": [
260
+ {
261
+ "track": "Predict",
262
+ "score": 74.81725902258384,
263
+ "costUsd": 0.022857185520000003,
264
+ "costIsLowerBound": true,
265
+ "valid": 38
266
+ },
267
+ {
268
+ "track": "Learn",
269
+ "score": 86.74674526342939,
270
+ "costUsd": 0.02613514208000001,
271
+ "costIsLowerBound": true,
272
+ "valid": 45
273
+ }
274
+ ]
275
+ },
276
+ {
277
+ "model": "hy3",
278
+ "provider": "tencent/fp8",
279
+ "source": "results/final-8000/tencent__hy3/tencent-hy3-ad08a72daecee623/report.json",
280
+ "sourceSha256": "f3f07a08cb8c2b16f60eebb913667b3984b906b546a8aea9428ca52735c09695",
281
+ "points": [
282
+ {
283
+ "track": "Predict",
284
+ "score": 86.13985983617552,
285
+ "costUsd": 0.033309804000000005,
286
+ "costIsLowerBound": false,
287
+ "valid": 45
288
+ },
289
+ {
290
+ "track": "Learn",
291
+ "score": 83.5824847611388,
292
+ "costUsd": 0.125771448,
293
+ "costIsLowerBound": false,
294
+ "valid": 45
295
+ }
296
+ ]
297
+ },
298
+ {
299
+ "model": "gemini-3.8-flash",
300
+ "provider": "google-ai-studio",
301
+ "source": "results/final-8000/google__gemini-3.8-flash/google-gemini-3-8-flash-f27a116ea917dbe9/report.json",
302
+ "sourceSha256": "8e454e4d874deeecb4cf41d24ca8ce98865428a6ff6e7f43721bb856869add3d",
303
+ "points": [
304
+ {
305
+ "track": "Predict",
306
+ "score": 86.47655635922422,
307
+ "costUsd": 0.10882725,
308
+ "costIsLowerBound": false,
309
+ "valid": 45
310
+ },
311
+ {
312
+ "track": "Learn",
313
+ "score": 87.93711091197737,
314
+ "costUsd": 0.47905575,
315
+ "costIsLowerBound": false,
316
+ "valid": 45
317
+ }
318
+ ]
319
+ },
320
+ {
321
+ "model": "gpt-6-luna",
322
+ "provider": "openai",
323
+ "source": "results/final-8000/openai__gpt-6-luna/openai-gpt-6-luna-94d7b25701149ff6/report.json",
324
+ "sourceSha256": "df1865c6563b42acc3bec9c37f43120b63e9842e3eafb75ea54ab728cbfbd271",
325
+ "points": [
326
+ {
327
+ "track": "Predict",
328
+ "score": 85.99226007227816,
329
+ "costUsd": 0.0028580999999999997,
330
+ "costIsLowerBound": false,
331
+ "valid": 44
332
+ },
333
+ {
334
+ "track": "Learn",
335
+ "score": 82.74807125831057,
336
+ "costUsd": 0.014754977499999997,
337
+ "costIsLowerBound": false,
338
+ "valid": 44
339
+ }
340
+ ]
341
+ },
342
+ {
343
+ "model": "claude-haiku-5.5",
344
+ "provider": "anthropic",
345
+ "source": "results/final-8000/anthropic__claude-haiku-5.5/anthropic-claude-haiku-5-5-c0cc2752b65c1032/report.json",
346
+ "sourceSha256": "32ef3c87e73fe9d78420e8238a8c76c9a05dc7a1c865e0beb55e022810d7e34a",
347
+ "points": [
348
+ {
349
+ "track": "Predict",
350
+ "score": 86.34996950320857,
351
+ "costUsd": 0.0050422999999999996,
352
+ "costIsLowerBound": false,
353
+ "valid": 45
354
+ },
355
+ {
356
+ "track": "Learn",
357
+ "score": 86.79157210497499,
358
+ "costUsd": 0.05925300000000001,
359
+ "costIsLowerBound": false,
360
+ "valid": 45
361
+ }
362
+ ]
363
+ },
364
+ {
365
+ "model": "inkling-small",
366
+ "provider": "deepinfra/fp4",
367
+ "source": "results/final-8000/thinkingmachines__inkling-small/thinkingmachines-inkling-small-8e9fa2bbf79c5f60/report.json",
368
+ "sourceSha256": "10b89e94895a8497d08e11f64a637e8f3310b146e474044e6c8783c4d35eefef",
369
+ "points": [
370
+ {
371
+ "track": "Predict",
372
+ "score": 86.56687634653072,
373
+ "costUsd": 0.02011335,
374
+ "costIsLowerBound": false,
375
+ "valid": 45
376
+ },
377
+ {
378
+ "track": "Learn",
379
+ "score": 82.0589167782321,
380
+ "costUsd": 0.12605145000000004,
381
+ "costIsLowerBound": false,
382
+ "valid": 44
383
+ }
384
+ ]
385
+ },
386
+ {
387
+ "model": "gpt-oss-120b",
388
+ "provider": "dekallm/bf16",
389
+ "source": "results/final-8000/openai__gpt-oss-120b/openai-gpt-oss-120b-1fc08149dc845523/report.json",
390
+ "sourceSha256": "c843ac88dc67870904eb19691813e4eac23fe91b79b2bfe3dd8c2013c53b83cc",
391
+ "points": [
392
+ {
393
+ "track": "Predict",
394
+ "score": 86.52508033881153,
395
+ "costUsd": 0.002907540000000001,
396
+ "costIsLowerBound": false,
397
+ "valid": 45
398
+ },
399
+ {
400
+ "track": "Learn",
401
+ "score": 82.84109578384674,
402
+ "costUsd": 0.03264432000000001,
403
+ "costIsLowerBound": true,
404
+ "valid": 45
405
+ }
406
+ ]
407
+ }
408
+ ],
409
+ "charts": [
410
+ {
411
+ "title": "RTR-Bench (Predict)",
412
+ "file": "predict-scatter",
413
+ "xAxis": "per-track known USD subtotal (log scale)",
414
+ "yAxis": "distribution score out of 100",
415
+ "paretoUsesKnownSubtotals": true,
416
+ "frontier": [
417
+ {
418
+ "track": "Predict",
419
+ "score": 87.65682697223357,
420
+ "costUsd": 0.0022276652999999994,
421
+ "costIsLowerBound": false,
422
+ "valid": 45
423
+ }
424
+ ]
425
+ },
426
+ {
427
+ "title": "RTR-Bench (Learn)",
428
+ "file": "learn-scatter",
429
+ "xAxis": "per-track known USD subtotal (log scale)",
430
+ "yAxis": "distribution score out of 100",
431
+ "paretoUsesKnownSubtotals": true,
432
+ "frontier": [
433
+ {
434
+ "track": "Learn",
435
+ "score": 82.74807125831057,
436
+ "costUsd": 0.014754977499999997,
437
+ "costIsLowerBound": false,
438
+ "valid": 44
439
+ },
440
+ {
441
+ "track": "Learn",
442
+ "score": 83.52752701775555,
443
+ "costUsd": 0.022174073600000014,
444
+ "costIsLowerBound": true,
445
+ "valid": 45
446
+ },
447
+ {
448
+ "track": "Learn",
449
+ "score": 85.99932573408472,
450
+ "costUsd": 0.02314631309999999,
451
+ "costIsLowerBound": false,
452
+ "valid": 45
453
+ },
454
+ {
455
+ "track": "Learn",
456
+ "score": 86.74674526342939,
457
+ "costUsd": 0.02613514208000001,
458
+ "costIsLowerBound": true,
459
+ "valid": 45
460
+ },
461
+ {
462
+ "track": "Learn",
463
+ "score": 86.79157210497499,
464
+ "costUsd": 0.05925300000000001,
465
+ "costIsLowerBound": false,
466
+ "valid": 45
467
+ },
468
+ {
469
+ "track": "Learn",
470
+ "score": 87.93711091197737,
471
+ "costUsd": 0.47905575,
472
+ "costIsLowerBound": false,
473
+ "valid": 45
474
+ }
475
+ ]
476
+ },
477
+ {
478
+ "title": "RTR-Bench (Learn + Predict)",
479
+ "file": "predict-learn-scatter",
480
+ "xAxis": "per-track known USD subtotal (log scale)",
481
+ "yAxis": "distribution score out of 100",
482
+ "paretoUsesKnownSubtotals": false,
483
+ "frontier": []
484
+ }
485
+ ]
486
+ }
results/predict-learn-scatter.png ADDED

Git LFS Details

  • SHA256: 8251a848e0404824deae0ea671399d254c6ac60a4cd943480690f9274b7218d7
  • Pointer size: 131 Bytes
  • Size of remote file: 200 kB
results/predict-learn-scatter.svg ADDED
results/predict-scatter.png ADDED

Git LFS Details

  • SHA256: ba796076f40f72fda72cdc89d0e66601973b67c75facde019979b9c3f39ec1e0
  • Pointer size: 130 Bytes
  • Size of remote file: 70.4 kB
results/predict-scatter.svg ADDED
results/verified-results.json ADDED
@@ -0,0 +1,1473 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "protocol": {
3
+ "partition": "evaluation",
4
+ "targets": [
5
+ 130,
6
+ 131,
7
+ 132,
8
+ 133,
9
+ 134,
10
+ 135,
11
+ 136,
12
+ 137,
13
+ 138,
14
+ 139,
15
+ 140,
16
+ 141,
17
+ 142,
18
+ 143,
19
+ 144,
20
+ 145,
21
+ 146,
22
+ 147,
23
+ 148,
24
+ 149,
25
+ 150,
26
+ 151,
27
+ 152,
28
+ 153,
29
+ 154,
30
+ 155,
31
+ 156,
32
+ 157,
33
+ 158,
34
+ 159,
35
+ 160,
36
+ 161,
37
+ 162,
38
+ 163,
39
+ 164,
40
+ 165,
41
+ 166,
42
+ 167,
43
+ 168,
44
+ 169,
45
+ 170,
46
+ 171,
47
+ 172,
48
+ 173,
49
+ 174
50
+ ],
51
+ "reference": [
52
+ 1,
53
+ 2,
54
+ 3,
55
+ 4,
56
+ 5,
57
+ 6,
58
+ 7,
59
+ 8,
60
+ 9,
61
+ 10,
62
+ 11,
63
+ 12,
64
+ 13,
65
+ 14,
66
+ 15,
67
+ 16,
68
+ 17,
69
+ 18,
70
+ 19,
71
+ 20,
72
+ 21,
73
+ 22,
74
+ 23,
75
+ 24,
76
+ 25,
77
+ 28,
78
+ 29,
79
+ 30,
80
+ 31,
81
+ 35,
82
+ 36,
83
+ 37,
84
+ 38,
85
+ 39,
86
+ 42,
87
+ 43,
88
+ 44,
89
+ 45,
90
+ 46,
91
+ 47,
92
+ 48,
93
+ 49,
94
+ 50,
95
+ 51,
96
+ 52,
97
+ 53,
98
+ 54,
99
+ 55,
100
+ 56,
101
+ 57,
102
+ 58,
103
+ 59,
104
+ 60,
105
+ 61,
106
+ 62,
107
+ 63,
108
+ 64,
109
+ 65,
110
+ 66,
111
+ 67,
112
+ 68,
113
+ 69,
114
+ 70,
115
+ 71,
116
+ 72,
117
+ 73,
118
+ 74,
119
+ 77,
120
+ 78,
121
+ 79,
122
+ 80,
123
+ 81,
124
+ 82,
125
+ 83,
126
+ 84,
127
+ 85,
128
+ 86,
129
+ 87,
130
+ 88,
131
+ 89,
132
+ 90,
133
+ 91,
134
+ 92,
135
+ 93,
136
+ 94,
137
+ 95,
138
+ 96,
139
+ 97,
140
+ 98,
141
+ 99
142
+ ],
143
+ "datasetHash": "ec9d7582b5610a24f4e14dbbb05d13c636b692ec9e133b9381db0fdf483ffb2b",
144
+ "splitHash": "a165c72f7557f8128253148eb28e6906b39e80aec12e45c679fddde62fa31421",
145
+ "strategies": [
146
+ "none",
147
+ "random",
148
+ "similar",
149
+ "agent"
150
+ ],
151
+ "budget": 3,
152
+ "seed": 7,
153
+ "maxModelCalls": 8,
154
+ "maxActions": 8,
155
+ "maxTokens": 8000,
156
+ "maxAttempts": 3,
157
+ "timeoutMs": 60000,
158
+ "agentPolicy": "transfer-check-v1",
159
+ "temperature": "0 where supported; provider default otherwise",
160
+ "reasoning": "provider defaults, unchanged from development"
161
+ },
162
+ "results": [
163
+ {
164
+ "model": "z-ai/glm-5.3-flash",
165
+ "provider": "relace/fp4",
166
+ "quantization": "fp4",
167
+ "directory": "results/final-8000/z-ai__glm-5.3-flash/z-ai-glm-5-3-flash-a3721d86cb3fffd7",
168
+ "summaries": [
169
+ {
170
+ "strategy": "none",
171
+ "score": 87.65682697223357,
172
+ "scoreCI": [
173
+ 85.14961427102584,
174
+ 89.92094310021464
175
+ ],
176
+ "valid": 45,
177
+ "invalid": 0,
178
+ "missing": 0,
179
+ "pending": 0,
180
+ "costUsd": {
181
+ "known": 0.0022276652999999994,
182
+ "unknownAttempts": 0
183
+ },
184
+ "apiCalls": 45,
185
+ "modelCalls": 45,
186
+ "inspections": 0,
187
+ "unknownChargeReservations": 0
188
+ },
189
+ {
190
+ "strategy": "random",
191
+ "score": 85.72082689577341,
192
+ "scoreCI": [
193
+ 82.59402449443175,
194
+ 88.7419714232416
195
+ ],
196
+ "valid": 45,
197
+ "invalid": 0,
198
+ "missing": 0,
199
+ "pending": 0,
200
+ "costUsd": {
201
+ "known": 0.004247038300000001,
202
+ "unknownAttempts": 0
203
+ },
204
+ "apiCalls": 45,
205
+ "modelCalls": 45,
206
+ "inspections": 135,
207
+ "unknownChargeReservations": 0
208
+ },
209
+ {
210
+ "strategy": "similar",
211
+ "score": 85.87137296290045,
212
+ "scoreCI": [
213
+ 82.20795078736248,
214
+ 89.16360081221084
215
+ ],
216
+ "valid": 45,
217
+ "invalid": 0,
218
+ "missing": 0,
219
+ "pending": 0,
220
+ "costUsd": {
221
+ "known": 0.004007989300000001,
222
+ "unknownAttempts": 0
223
+ },
224
+ "apiCalls": 45,
225
+ "modelCalls": 45,
226
+ "inspections": 135,
227
+ "unknownChargeReservations": 0
228
+ },
229
+ {
230
+ "strategy": "agent",
231
+ "score": 85.99932573408472,
232
+ "scoreCI": [
233
+ 82.97390048451051,
234
+ 88.82239442274773
235
+ ],
236
+ "valid": 45,
237
+ "invalid": 0,
238
+ "missing": 0,
239
+ "pending": 0,
240
+ "costUsd": {
241
+ "known": 0.02314631309999999,
242
+ "unknownAttempts": 0
243
+ },
244
+ "apiCalls": 223,
245
+ "modelCalls": 223,
246
+ "inspections": 133,
247
+ "unknownChargeReservations": 0
248
+ }
249
+ ],
250
+ "learnGain": -1.6575012381488865,
251
+ "learnGainCI": [
252
+ -4.039703578553609,
253
+ 0.6888408447203508
254
+ ],
255
+ "baselines": {
256
+ "uniformScore": 85.38589005462313,
257
+ "constantScore": 85.55744059648882,
258
+ "constantPrediction": {
259
+ "performed+uniting": 0.2500381443429472,
260
+ "honest+uniting": 0.2078301757272207,
261
+ "performed+dividing": 0.2577803629097389,
262
+ "honest+dividing": 0.2843513170200932
263
+ }
264
+ },
265
+ "valid": 180,
266
+ "total": 180,
267
+ "apiCalls": 358,
268
+ "reportedCostUsd": 0.03362900599999999,
269
+ "accountedUsd": 0.033629175,
270
+ "capUsd": 7.866515078,
271
+ "unknownReservations": 0,
272
+ "unbilledRateLimits": 0,
273
+ "apiErrors": 0,
274
+ "lengthStops": 0,
275
+ "maxCompletionTokens": 3156,
276
+ "actualProviders": [
277
+ "Relace"
278
+ ],
279
+ "elapsedMinutes": 34.3668,
280
+ "offlineResumeVerified": true
281
+ },
282
+ {
283
+ "model": "z-ai/glm-5.3",
284
+ "provider": "wafer/us",
285
+ "quantization": "fp4",
286
+ "directory": "results/final-8000/z-ai__glm-5.3/z-ai-glm-5-3-df181956130e550d",
287
+ "summaries": [
288
+ {
289
+ "strategy": "none",
290
+ "score": 86.03929052235672,
291
+ "scoreCI": [
292
+ 81.23104161390178,
293
+ 89.68385726400506
294
+ ],
295
+ "valid": 44,
296
+ "invalid": 0,
297
+ "missing": 1,
298
+ "pending": 0,
299
+ "costUsd": {
300
+ "known": 0.005509304100000001,
301
+ "unknownAttempts": 0
302
+ },
303
+ "apiCalls": 45,
304
+ "modelCalls": 45,
305
+ "inspections": 0,
306
+ "unknownChargeReservations": 0
307
+ },
308
+ {
309
+ "strategy": "random",
310
+ "score": 85.62695726334336,
311
+ "scoreCI": [
312
+ 81.64060201355468,
313
+ 89.35786030832618
314
+ ],
315
+ "valid": 45,
316
+ "invalid": 0,
317
+ "missing": 0,
318
+ "pending": 0,
319
+ "costUsd": {
320
+ "known": 0.006686855499999999,
321
+ "unknownAttempts": 0
322
+ },
323
+ "apiCalls": 45,
324
+ "modelCalls": 45,
325
+ "inspections": 135,
326
+ "unknownChargeReservations": 0
327
+ },
328
+ {
329
+ "strategy": "similar",
330
+ "score": 87.11414687671163,
331
+ "scoreCI": [
332
+ 83.30583110462224,
333
+ 90.59804286037904
334
+ ],
335
+ "valid": 45,
336
+ "invalid": 0,
337
+ "missing": 0,
338
+ "pending": 0,
339
+ "costUsd": {
340
+ "known": 0.006796690499999999,
341
+ "unknownAttempts": 0
342
+ },
343
+ "apiCalls": 45,
344
+ "modelCalls": 45,
345
+ "inspections": 135,
346
+ "unknownChargeReservations": 0
347
+ },
348
+ {
349
+ "strategy": "agent",
350
+ "score": 82.86617668804384,
351
+ "scoreCI": [
352
+ 77.63967006911795,
353
+ 87.49948669626973
354
+ ],
355
+ "valid": 44,
356
+ "invalid": 1,
357
+ "missing": 0,
358
+ "pending": 0,
359
+ "costUsd": {
360
+ "known": 0.03842064060000001,
361
+ "unknownAttempts": 0
362
+ },
363
+ "apiCalls": 221,
364
+ "modelCalls": 221,
365
+ "inspections": 127,
366
+ "unknownChargeReservations": 0
367
+ }
368
+ ],
369
+ "learnGain": -3.1731138343128866,
370
+ "learnGainCI": [
371
+ -9.104741542233242,
372
+ 3.0943995588071163
373
+ ],
374
+ "baselines": {
375
+ "uniformScore": 85.38589005462313,
376
+ "constantScore": 85.55744059648882,
377
+ "constantPrediction": {
378
+ "performed+uniting": 0.2500381443429472,
379
+ "honest+uniting": 0.2078301757272207,
380
+ "performed+dividing": 0.2577803629097389,
381
+ "honest+dividing": 0.2843513170200932
382
+ }
383
+ },
384
+ "valid": 178,
385
+ "total": 180,
386
+ "apiCalls": 356,
387
+ "reportedCostUsd": 0.05741349070000001,
388
+ "accountedUsd": 0.057413658,
389
+ "capUsd": 0.25,
390
+ "unknownReservations": 0,
391
+ "unbilledRateLimits": 0,
392
+ "apiErrors": 0,
393
+ "lengthStops": 0,
394
+ "maxCompletionTokens": 317,
395
+ "actualProviders": [
396
+ "Wafer"
397
+ ],
398
+ "elapsedMinutes": 4.303716666666666,
399
+ "offlineResumeVerified": true
400
+ },
401
+ {
402
+ "model": "minimax/minimax-m3.1-flash",
403
+ "provider": "inferact-vllm/nvfp4",
404
+ "quantization": "nvfp4",
405
+ "directory": "results/final-8000/minimax__minimax-m3.1-flash/minimax-minimax-m3-1-flash-8985e19160d564f2",
406
+ "summaries": [
407
+ {
408
+ "strategy": "none",
409
+ "score": 86.95313994076243,
410
+ "scoreCI": [
411
+ 84.60410563195533,
412
+ 89.13138791179802
413
+ ],
414
+ "valid": 45,
415
+ "invalid": 0,
416
+ "missing": 0,
417
+ "pending": 0,
418
+ "costUsd": {
419
+ "known": 0.017481899999999995,
420
+ "unknownAttempts": 3
421
+ },
422
+ "apiCalls": 48,
423
+ "modelCalls": 45,
424
+ "inspections": 0,
425
+ "unknownChargeReservations": 0
426
+ },
427
+ {
428
+ "strategy": "random",
429
+ "score": 83.13784686248302,
430
+ "scoreCI": [
431
+ 77.44827401190956,
432
+ 87.57666794504942
433
+ ],
434
+ "valid": 44,
435
+ "invalid": 0,
436
+ "missing": 1,
437
+ "pending": 0,
438
+ "costUsd": {
439
+ "known": 0.017992124999999998,
440
+ "unknownAttempts": 6
441
+ },
442
+ "apiCalls": 51,
443
+ "modelCalls": 45,
444
+ "inspections": 135,
445
+ "unknownChargeReservations": 0
446
+ },
447
+ {
448
+ "strategy": "similar",
449
+ "score": 81.67465450070995,
450
+ "scoreCI": [
451
+ 74.91695420606257,
452
+ 87.4217820867126
453
+ ],
454
+ "valid": 43,
455
+ "invalid": 0,
456
+ "missing": 2,
457
+ "pending": 0,
458
+ "costUsd": {
459
+ "known": 0.01805805,
460
+ "unknownAttempts": 2
461
+ },
462
+ "apiCalls": 47,
463
+ "modelCalls": 45,
464
+ "inspections": 135,
465
+ "unknownChargeReservations": 0
466
+ },
467
+ {
468
+ "strategy": "agent",
469
+ "score": 84.24219981765467,
470
+ "scoreCI": [
471
+ 80.5616511727701,
472
+ 87.65993078381547
473
+ ],
474
+ "valid": 45,
475
+ "invalid": 0,
476
+ "missing": 0,
477
+ "pending": 0,
478
+ "costUsd": {
479
+ "known": 0.06344557500000003,
480
+ "unknownAttempts": 11
481
+ },
482
+ "apiCalls": 242,
483
+ "modelCalls": 231,
484
+ "inspections": 135,
485
+ "unknownChargeReservations": 0
486
+ }
487
+ ],
488
+ "learnGain": -2.71094012310773,
489
+ "learnGainCI": [
490
+ -6.13168766197505,
491
+ 0.5924653725976328
492
+ ],
493
+ "baselines": {
494
+ "uniformScore": 85.38589005462313,
495
+ "constantScore": 85.55744059648882,
496
+ "constantPrediction": {
497
+ "performed+uniting": 0.2500381443429472,
498
+ "honest+uniting": 0.2078301757272207,
499
+ "performed+dividing": 0.2577803629097389,
500
+ "honest+dividing": 0.2843513170200932
501
+ }
502
+ },
503
+ "valid": 177,
504
+ "total": 180,
505
+ "apiCalls": 388,
506
+ "reportedCostUsd": 0.11697765000000002,
507
+ "accountedUsd": 0.116977658,
508
+ "capUsd": 0.5,
509
+ "unknownReservations": 0,
510
+ "unbilledRateLimits": 22,
511
+ "apiErrors": 22,
512
+ "lengthStops": 9,
513
+ "maxCompletionTokens": 8000,
514
+ "actualProviders": [
515
+ "Inferact vLLM"
516
+ ],
517
+ "elapsedMinutes": 26.686766666666667,
518
+ "offlineResumeVerified": true
519
+ },
520
+ {
521
+ "model": "xiaomi/mimo-v2.6-flash",
522
+ "provider": "xiaomi/fp8",
523
+ "quantization": "fp8",
524
+ "directory": "results/final-8000/xiaomi__mimo-v2.6-flash/xiaomi-mimo-v2-6-flash-0e9297795e7f3f19",
525
+ "summaries": [
526
+ {
527
+ "strategy": "none",
528
+ "score": 86.97924213353758,
529
+ "scoreCI": [
530
+ 84.5975146410855,
531
+ 89.23277908048537
532
+ ],
533
+ "valid": 45,
534
+ "invalid": 0,
535
+ "missing": 0,
536
+ "pending": 0,
537
+ "costUsd": {
538
+ "known": 0.00389634,
539
+ "unknownAttempts": 0
540
+ },
541
+ "apiCalls": 45,
542
+ "modelCalls": 45,
543
+ "inspections": 0,
544
+ "unknownChargeReservations": 0
545
+ },
546
+ {
547
+ "strategy": "random",
548
+ "score": 85.35181412899313,
549
+ "scoreCI": [
550
+ 81.89272966410402,
551
+ 88.13561788623723
552
+ ],
553
+ "valid": 45,
554
+ "invalid": 0,
555
+ "missing": 0,
556
+ "pending": 0,
557
+ "costUsd": {
558
+ "known": 0.006863695999999999,
559
+ "unknownAttempts": 0
560
+ },
561
+ "apiCalls": 45,
562
+ "modelCalls": 45,
563
+ "inspections": 135,
564
+ "unknownChargeReservations": 0
565
+ },
566
+ {
567
+ "strategy": "similar",
568
+ "score": 86.25270267889361,
569
+ "scoreCI": [
570
+ 82.94989074570486,
571
+ 89.18470846120827
572
+ ],
573
+ "valid": 45,
574
+ "invalid": 0,
575
+ "missing": 0,
576
+ "pending": 0,
577
+ "costUsd": {
578
+ "known": 0.007120265599999998,
579
+ "unknownAttempts": 0
580
+ },
581
+ "apiCalls": 45,
582
+ "modelCalls": 45,
583
+ "inspections": 135,
584
+ "unknownChargeReservations": 0
585
+ },
586
+ {
587
+ "strategy": "agent",
588
+ "score": 83.52752701775555,
589
+ "scoreCI": [
590
+ 79.73187893623671,
591
+ 87.07132567471113
592
+ ],
593
+ "valid": 45,
594
+ "invalid": 0,
595
+ "missing": 0,
596
+ "pending": 0,
597
+ "costUsd": {
598
+ "known": 0.022174073600000014,
599
+ "unknownAttempts": 2
600
+ },
601
+ "apiCalls": 205,
602
+ "modelCalls": 203,
603
+ "inspections": 110,
604
+ "unknownChargeReservations": 2
605
+ }
606
+ ],
607
+ "learnGain": -3.4517151157820396,
608
+ "learnGainCI": [
609
+ -6.629202272622245,
610
+ -0.45061918291670855
611
+ ],
612
+ "baselines": {
613
+ "uniformScore": 85.38589005462313,
614
+ "constantScore": 85.55744059648882,
615
+ "constantPrediction": {
616
+ "performed+uniting": 0.2500381443429472,
617
+ "honest+uniting": 0.2078301757272207,
618
+ "performed+dividing": 0.2577803629097389,
619
+ "honest+dividing": 0.2843513170200932
620
+ }
621
+ },
622
+ "valid": 180,
623
+ "total": 180,
624
+ "apiCalls": 340,
625
+ "reportedCostUsd": 0.040054375200000006,
626
+ "accountedUsd": 0.338135717,
627
+ "capUsd": 0.5,
628
+ "unknownReservations": 2,
629
+ "unbilledRateLimits": 0,
630
+ "apiErrors": 2,
631
+ "lengthStops": 0,
632
+ "maxCompletionTokens": 795,
633
+ "actualProviders": [
634
+ "Xiaomi"
635
+ ],
636
+ "elapsedMinutes": 37.593,
637
+ "offlineResumeVerified": true
638
+ },
639
+ {
640
+ "model": "deepseek/deepseek-v4.1-flash-1017",
641
+ "provider": "sail-research/fp4/us",
642
+ "quantization": "fp4",
643
+ "directory": "results/final-8000/deepseek__deepseek-v4.1-flash-1017/deepseek-deepseek-v4-1-flash-1017-08a69325bd0d847d",
644
+ "summaries": [
645
+ {
646
+ "strategy": "none",
647
+ "score": 74.81725902258384,
648
+ "scoreCI": [
649
+ 64.92593224672528,
650
+ 83.78867819223935
651
+ ],
652
+ "valid": 38,
653
+ "invalid": 0,
654
+ "missing": 7,
655
+ "pending": 0,
656
+ "costUsd": {
657
+ "known": 0.022857185520000003,
658
+ "unknownAttempts": 1
659
+ },
660
+ "apiCalls": 46,
661
+ "modelCalls": 45,
662
+ "inspections": 0,
663
+ "unknownChargeReservations": 1
664
+ },
665
+ {
666
+ "strategy": "random",
667
+ "score": 62.20354213551938,
668
+ "scoreCI": [
669
+ 49.399666065294845,
670
+ 73.99177142337163
671
+ ],
672
+ "valid": 31,
673
+ "invalid": 0,
674
+ "missing": 14,
675
+ "pending": 0,
676
+ "costUsd": {
677
+ "known": 0.02864010311999999,
678
+ "unknownAttempts": 2
679
+ },
680
+ "apiCalls": 47,
681
+ "modelCalls": 45,
682
+ "inspections": 135,
683
+ "unknownChargeReservations": 2
684
+ },
685
+ {
686
+ "strategy": "similar",
687
+ "score": 64.98243948187805,
688
+ "scoreCI": [
689
+ 52.642307939981684,
690
+ 76.42703879052287
691
+ ],
692
+ "valid": 33,
693
+ "invalid": 0,
694
+ "missing": 12,
695
+ "pending": 0,
696
+ "costUsd": {
697
+ "known": 0.027315130320000006,
698
+ "unknownAttempts": 3
699
+ },
700
+ "apiCalls": 47,
701
+ "modelCalls": 45,
702
+ "inspections": 135,
703
+ "unknownChargeReservations": 3
704
+ },
705
+ {
706
+ "strategy": "agent",
707
+ "score": 86.74674526342939,
708
+ "scoreCI": [
709
+ 83.3237005555384,
710
+ 90.12876195551101
711
+ ],
712
+ "valid": 45,
713
+ "invalid": 0,
714
+ "missing": 0,
715
+ "pending": 0,
716
+ "costUsd": {
717
+ "known": 0.02613514208000001,
718
+ "unknownAttempts": 6
719
+ },
720
+ "apiCalls": 229,
721
+ "modelCalls": 223,
722
+ "inspections": 132,
723
+ "unknownChargeReservations": 6
724
+ }
725
+ ],
726
+ "learnGain": 11.929486240845526,
727
+ "learnGainCI": [
728
+ 2.428821818523353,
729
+ 22.766831994932044
730
+ ],
731
+ "baselines": {
732
+ "uniformScore": 85.38589005462313,
733
+ "constantScore": 85.55744059648882,
734
+ "constantPrediction": {
735
+ "performed+uniting": 0.2500381443429472,
736
+ "honest+uniting": 0.2078301757272207,
737
+ "performed+dividing": 0.2577803629097389,
738
+ "honest+dividing": 0.2843513170200932
739
+ }
740
+ },
741
+ "valid": 147,
742
+ "total": 180,
743
+ "apiCalls": 369,
744
+ "reportedCostUsd": 0.10494756104,
745
+ "accountedUsd": 1.030345096,
746
+ "capUsd": 1.2,
747
+ "unknownReservations": 12,
748
+ "unbilledRateLimits": 0,
749
+ "apiErrors": 11,
750
+ "lengthStops": 33,
751
+ "maxCompletionTokens": 8000,
752
+ "actualProviders": [
753
+ "Sail Research"
754
+ ],
755
+ "elapsedMinutes": 102.52823333333333,
756
+ "offlineResumeVerified": true
757
+ },
758
+ {
759
+ "model": "tencent/hy3",
760
+ "provider": "tencent/fp8",
761
+ "quantization": "fp8",
762
+ "directory": "results/final-8000/tencent__hy3/tencent-hy3-ad08a72daecee623",
763
+ "summaries": [
764
+ {
765
+ "strategy": "none",
766
+ "score": 86.13985983617552,
767
+ "scoreCI": [
768
+ 83.04456132619205,
769
+ 88.99019363652691
770
+ ],
771
+ "valid": 45,
772
+ "invalid": 0,
773
+ "missing": 0,
774
+ "pending": 0,
775
+ "costUsd": {
776
+ "known": 0.033309804000000005,
777
+ "unknownAttempts": 0
778
+ },
779
+ "apiCalls": 45,
780
+ "modelCalls": 45,
781
+ "inspections": 0,
782
+ "unknownChargeReservations": 0
783
+ },
784
+ {
785
+ "strategy": "random",
786
+ "score": 82.39022498497121,
787
+ "scoreCI": [
788
+ 78.19446572506257,
789
+ 86.27460771618624
790
+ ],
791
+ "valid": 45,
792
+ "invalid": 0,
793
+ "missing": 0,
794
+ "pending": 0,
795
+ "costUsd": {
796
+ "known": 0.058402211999999995,
797
+ "unknownAttempts": 0
798
+ },
799
+ "apiCalls": 45,
800
+ "modelCalls": 45,
801
+ "inspections": 135,
802
+ "unknownChargeReservations": 0
803
+ },
804
+ {
805
+ "strategy": "similar",
806
+ "score": 86.36791618754778,
807
+ "scoreCI": [
808
+ 82.22368966340524,
809
+ 89.86949469349841
810
+ ],
811
+ "valid": 45,
812
+ "invalid": 0,
813
+ "missing": 0,
814
+ "pending": 0,
815
+ "costUsd": {
816
+ "known": 0.059776727999999994,
817
+ "unknownAttempts": 4
818
+ },
819
+ "apiCalls": 49,
820
+ "modelCalls": 45,
821
+ "inspections": 135,
822
+ "unknownChargeReservations": 4
823
+ },
824
+ {
825
+ "strategy": "agent",
826
+ "score": 83.5824847611388,
827
+ "scoreCI": [
828
+ 79.72401574181727,
829
+ 87.06289351424579
830
+ ],
831
+ "valid": 45,
832
+ "invalid": 0,
833
+ "missing": 0,
834
+ "pending": 0,
835
+ "costUsd": {
836
+ "known": 0.125771448,
837
+ "unknownAttempts": 0
838
+ },
839
+ "apiCalls": 198,
840
+ "modelCalls": 198,
841
+ "inspections": 108,
842
+ "unknownChargeReservations": 0
843
+ }
844
+ ],
845
+ "learnGain": -2.5573750750367363,
846
+ "learnGainCI": [
847
+ -6.088462930819695,
848
+ 0.9752903280162528
849
+ ],
850
+ "baselines": {
851
+ "uniformScore": 85.38589005462313,
852
+ "constantScore": 85.55744059648882,
853
+ "constantPrediction": {
854
+ "performed+uniting": 0.2500381443429472,
855
+ "honest+uniting": 0.2078301757272207,
856
+ "performed+dividing": 0.2577803629097389,
857
+ "honest+dividing": 0.2843513170200932
858
+ }
859
+ },
860
+ "valid": 180,
861
+ "total": 180,
862
+ "apiCalls": 337,
863
+ "reportedCostUsd": 0.277260192,
864
+ "accountedUsd": 0.432568232,
865
+ "capUsd": 8.29908331,
866
+ "unknownReservations": 4,
867
+ "unbilledRateLimits": 0,
868
+ "apiErrors": 4,
869
+ "lengthStops": 0,
870
+ "maxCompletionTokens": 4434,
871
+ "actualProviders": [
872
+ "Tencent"
873
+ ],
874
+ "elapsedMinutes": 116.0642,
875
+ "offlineResumeVerified": true
876
+ },
877
+ {
878
+ "model": "google/gemini-3.8-flash",
879
+ "provider": "google-ai-studio",
880
+ "quantization": "unknown",
881
+ "directory": "results/final-8000/google__gemini-3.8-flash/google-gemini-3-8-flash-f27a116ea917dbe9",
882
+ "summaries": [
883
+ {
884
+ "strategy": "none",
885
+ "score": 86.47655635922422,
886
+ "scoreCI": [
887
+ 83.5992524899307,
888
+ 89.23471286167829
889
+ ],
890
+ "valid": 45,
891
+ "invalid": 0,
892
+ "missing": 0,
893
+ "pending": 0,
894
+ "costUsd": {
895
+ "known": 0.10882725,
896
+ "unknownAttempts": 0
897
+ },
898
+ "apiCalls": 45,
899
+ "modelCalls": 45,
900
+ "inspections": 0,
901
+ "unknownChargeReservations": 0
902
+ },
903
+ {
904
+ "strategy": "random",
905
+ "score": 87.21769931025307,
906
+ "scoreCI": [
907
+ 84.26945550925288,
908
+ 90.12258788122982
909
+ ],
910
+ "valid": 45,
911
+ "invalid": 0,
912
+ "missing": 0,
913
+ "pending": 0,
914
+ "costUsd": {
915
+ "known": 0.15634575,
916
+ "unknownAttempts": 0
917
+ },
918
+ "apiCalls": 45,
919
+ "modelCalls": 45,
920
+ "inspections": 135,
921
+ "unknownChargeReservations": 0
922
+ },
923
+ {
924
+ "strategy": "similar",
925
+ "score": 86.57834442466589,
926
+ "scoreCI": [
927
+ 83.35487644929039,
928
+ 89.6369659649018
929
+ ],
930
+ "valid": 45,
931
+ "invalid": 0,
932
+ "missing": 0,
933
+ "pending": 0,
934
+ "costUsd": {
935
+ "known": 0.15300825000000004,
936
+ "unknownAttempts": 0
937
+ },
938
+ "apiCalls": 45,
939
+ "modelCalls": 45,
940
+ "inspections": 135,
941
+ "unknownChargeReservations": 0
942
+ },
943
+ {
944
+ "strategy": "agent",
945
+ "score": 87.93711091197737,
946
+ "scoreCI": [
947
+ 84.78783551728755,
948
+ 90.71822349457014
949
+ ],
950
+ "valid": 45,
951
+ "invalid": 0,
952
+ "missing": 0,
953
+ "pending": 0,
954
+ "costUsd": {
955
+ "known": 0.47905575,
956
+ "unknownAttempts": 0
957
+ },
958
+ "apiCalls": 229,
959
+ "modelCalls": 229,
960
+ "inspections": 135,
961
+ "unknownChargeReservations": 0
962
+ }
963
+ ],
964
+ "learnGain": 1.460554552753152,
965
+ "learnGainCI": [
966
+ -0.6881904861547372,
967
+ 3.5775829858397312
968
+ ],
969
+ "baselines": {
970
+ "uniformScore": 85.38589005462313,
971
+ "constantScore": 85.55744059648882,
972
+ "constantPrediction": {
973
+ "performed+uniting": 0.2500381443429472,
974
+ "honest+uniting": 0.2078301757272207,
975
+ "performed+dividing": 0.2577803629097389,
976
+ "honest+dividing": 0.2843513170200932
977
+ }
978
+ },
979
+ "valid": 180,
980
+ "total": 180,
981
+ "apiCalls": 364,
982
+ "reportedCostUsd": 0.8972370000000001,
983
+ "accountedUsd": 0.897237007,
984
+ "capUsd": 7.832885903,
985
+ "unknownReservations": 0,
986
+ "unbilledRateLimits": 0,
987
+ "apiErrors": 0,
988
+ "lengthStops": 0,
989
+ "maxCompletionTokens": 2314,
990
+ "actualProviders": [
991
+ "Google AI Studio"
992
+ ],
993
+ "elapsedMinutes": 31.785566666666668,
994
+ "offlineResumeVerified": true
995
+ },
996
+ {
997
+ "model": "openai/gpt-6-luna",
998
+ "provider": "openai",
999
+ "quantization": "unknown",
1000
+ "directory": "results/final-8000/openai__gpt-6-luna/openai-gpt-6-luna-94d7b25701149ff6",
1001
+ "summaries": [
1002
+ {
1003
+ "strategy": "none",
1004
+ "score": 85.99226007227816,
1005
+ "scoreCI": [
1006
+ 80.85957227090464,
1007
+ 89.76239857660065
1008
+ ],
1009
+ "valid": 44,
1010
+ "invalid": 1,
1011
+ "missing": 0,
1012
+ "pending": 0,
1013
+ "costUsd": {
1014
+ "known": 0.0028580999999999997,
1015
+ "unknownAttempts": 0
1016
+ },
1017
+ "apiCalls": 45,
1018
+ "modelCalls": 45,
1019
+ "inspections": 0,
1020
+ "unknownChargeReservations": 0
1021
+ },
1022
+ {
1023
+ "strategy": "random",
1024
+ "score": 83.7951999298725,
1025
+ "scoreCI": [
1026
+ 76.98127793601262,
1027
+ 89.09672902506287
1028
+ ],
1029
+ "valid": 43,
1030
+ "invalid": 2,
1031
+ "missing": 0,
1032
+ "pending": 0,
1033
+ "costUsd": {
1034
+ "known": 0.0055698499999999995,
1035
+ "unknownAttempts": 0
1036
+ },
1037
+ "apiCalls": 45,
1038
+ "modelCalls": 45,
1039
+ "inspections": 135,
1040
+ "unknownChargeReservations": 0
1041
+ },
1042
+ {
1043
+ "strategy": "similar",
1044
+ "score": 85.31099858101005,
1045
+ "scoreCI": [
1046
+ 79.71210567413544,
1047
+ 89.68302243699291
1048
+ ],
1049
+ "valid": 44,
1050
+ "invalid": 1,
1051
+ "missing": 0,
1052
+ "pending": 0,
1053
+ "costUsd": {
1054
+ "known": 0.005645449999999998,
1055
+ "unknownAttempts": 0
1056
+ },
1057
+ "apiCalls": 45,
1058
+ "modelCalls": 45,
1059
+ "inspections": 135,
1060
+ "unknownChargeReservations": 0
1061
+ },
1062
+ {
1063
+ "strategy": "agent",
1064
+ "score": 82.74807125831057,
1065
+ "scoreCI": [
1066
+ 77.31979689275076,
1067
+ 87.38326049880996
1068
+ ],
1069
+ "valid": 44,
1070
+ "invalid": 1,
1071
+ "missing": 0,
1072
+ "pending": 0,
1073
+ "costUsd": {
1074
+ "known": 0.014754977499999997,
1075
+ "unknownAttempts": 0
1076
+ },
1077
+ "apiCalls": 237,
1078
+ "modelCalls": 237,
1079
+ "inspections": 135,
1080
+ "unknownChargeReservations": 0
1081
+ }
1082
+ ],
1083
+ "learnGain": -3.2441888139675648,
1084
+ "learnGainCI": [
1085
+ -8.463496612221975,
1086
+ 1.4771904061060266
1087
+ ],
1088
+ "baselines": {
1089
+ "uniformScore": 85.38589005462313,
1090
+ "constantScore": 85.55744059648882,
1091
+ "constantPrediction": {
1092
+ "performed+uniting": 0.2500381443429472,
1093
+ "honest+uniting": 0.2078301757272207,
1094
+ "performed+dividing": 0.2577803629097389,
1095
+ "honest+dividing": 0.2843513170200932
1096
+ }
1097
+ },
1098
+ "valid": 175,
1099
+ "total": 180,
1100
+ "apiCalls": 372,
1101
+ "reportedCostUsd": 0.028828377499999995,
1102
+ "accountedUsd": 0.028828433,
1103
+ "capUsd": 6.935648896,
1104
+ "unknownReservations": 0,
1105
+ "unbilledRateLimits": 0,
1106
+ "apiErrors": 0,
1107
+ "lengthStops": 0,
1108
+ "maxCompletionTokens": 798,
1109
+ "actualProviders": [
1110
+ "OpenAI"
1111
+ ],
1112
+ "elapsedMinutes": 40.48925,
1113
+ "offlineResumeVerified": true
1114
+ },
1115
+ {
1116
+ "model": "anthropic/claude-haiku-5.5",
1117
+ "provider": "anthropic",
1118
+ "quantization": "unknown",
1119
+ "directory": "results/final-8000/anthropic__claude-haiku-5.5/anthropic-claude-haiku-5-5-c0cc2752b65c1032",
1120
+ "summaries": [
1121
+ {
1122
+ "strategy": "none",
1123
+ "score": 86.34996950320857,
1124
+ "scoreCI": [
1125
+ 83.86934770745339,
1126
+ 88.75628747105861
1127
+ ],
1128
+ "valid": 45,
1129
+ "invalid": 0,
1130
+ "missing": 0,
1131
+ "pending": 0,
1132
+ "costUsd": {
1133
+ "known": 0.0050422999999999996,
1134
+ "unknownAttempts": 0
1135
+ },
1136
+ "apiCalls": 45,
1137
+ "modelCalls": 45,
1138
+ "inspections": 0,
1139
+ "unknownChargeReservations": 0
1140
+ },
1141
+ {
1142
+ "strategy": "random",
1143
+ "score": 87.44082673833633,
1144
+ "scoreCI": [
1145
+ 84.823647856549,
1146
+ 90.12569427086608
1147
+ ],
1148
+ "valid": 45,
1149
+ "invalid": 0,
1150
+ "missing": 0,
1151
+ "pending": 0,
1152
+ "costUsd": {
1153
+ "known": 0.0078316,
1154
+ "unknownAttempts": 0
1155
+ },
1156
+ "apiCalls": 45,
1157
+ "modelCalls": 45,
1158
+ "inspections": 135,
1159
+ "unknownChargeReservations": 0
1160
+ },
1161
+ {
1162
+ "strategy": "similar",
1163
+ "score": 87.21660125549595,
1164
+ "scoreCI": [
1165
+ 84.38001949748943,
1166
+ 89.5988170012143
1167
+ ],
1168
+ "valid": 45,
1169
+ "invalid": 0,
1170
+ "missing": 0,
1171
+ "pending": 0,
1172
+ "costUsd": {
1173
+ "known": 0.007457899999999999,
1174
+ "unknownAttempts": 0
1175
+ },
1176
+ "apiCalls": 45,
1177
+ "modelCalls": 45,
1178
+ "inspections": 135,
1179
+ "unknownChargeReservations": 0
1180
+ },
1181
+ {
1182
+ "strategy": "agent",
1183
+ "score": 86.79157210497499,
1184
+ "scoreCI": [
1185
+ 84.20365714649625,
1186
+ 89.16785380356887
1187
+ ],
1188
+ "valid": 45,
1189
+ "invalid": 0,
1190
+ "missing": 0,
1191
+ "pending": 0,
1192
+ "costUsd": {
1193
+ "known": 0.05925300000000001,
1194
+ "unknownAttempts": 0
1195
+ },
1196
+ "apiCalls": 170,
1197
+ "modelCalls": 170,
1198
+ "inspections": 77,
1199
+ "unknownChargeReservations": 0
1200
+ }
1201
+ ],
1202
+ "learnGain": 0.441602601766414,
1203
+ "learnGainCI": [
1204
+ -1.930202444779965,
1205
+ 2.867520266082274
1206
+ ],
1207
+ "baselines": {
1208
+ "uniformScore": 85.38589005462313,
1209
+ "constantScore": 85.55744059648882,
1210
+ "constantPrediction": {
1211
+ "performed+uniting": 0.2500381443429472,
1212
+ "honest+uniting": 0.2078301757272207,
1213
+ "performed+dividing": 0.2577803629097389,
1214
+ "honest+dividing": 0.2843513170200932
1215
+ }
1216
+ },
1217
+ "valid": 180,
1218
+ "total": 180,
1219
+ "apiCalls": 305,
1220
+ "reportedCostUsd": 0.07958480000000001,
1221
+ "accountedUsd": 0.079584802,
1222
+ "capUsd": 6.906820463,
1223
+ "unknownReservations": 0,
1224
+ "unbilledRateLimits": 0,
1225
+ "apiErrors": 0,
1226
+ "lengthStops": 0,
1227
+ "maxCompletionTokens": 579,
1228
+ "actualProviders": [
1229
+ "Anthropic"
1230
+ ],
1231
+ "elapsedMinutes": 10.733583333333334,
1232
+ "offlineResumeVerified": true
1233
+ },
1234
+ {
1235
+ "model": "thinkingmachines/inkling-small",
1236
+ "provider": "deepinfra/fp4",
1237
+ "quantization": "fp4",
1238
+ "directory": "results/final-8000/thinkingmachines__inkling-small/thinkingmachines-inkling-small-8e9fa2bbf79c5f60",
1239
+ "summaries": [
1240
+ {
1241
+ "strategy": "none",
1242
+ "score": 86.56687634653072,
1243
+ "scoreCI": [
1244
+ 83.45736185942403,
1245
+ 89.16030565580068
1246
+ ],
1247
+ "valid": 45,
1248
+ "invalid": 0,
1249
+ "missing": 0,
1250
+ "pending": 0,
1251
+ "costUsd": {
1252
+ "known": 0.02011335,
1253
+ "unknownAttempts": 0
1254
+ },
1255
+ "apiCalls": 45,
1256
+ "modelCalls": 45,
1257
+ "inspections": 0,
1258
+ "unknownChargeReservations": 0
1259
+ },
1260
+ {
1261
+ "strategy": "random",
1262
+ "score": 87.16847269785018,
1263
+ "scoreCI": [
1264
+ 83.66296453727804,
1265
+ 90.23384432258139
1266
+ ],
1267
+ "valid": 45,
1268
+ "invalid": 0,
1269
+ "missing": 0,
1270
+ "pending": 0,
1271
+ "costUsd": {
1272
+ "known": 0.040481399999999994,
1273
+ "unknownAttempts": 0
1274
+ },
1275
+ "apiCalls": 45,
1276
+ "modelCalls": 45,
1277
+ "inspections": 135,
1278
+ "unknownChargeReservations": 0
1279
+ },
1280
+ {
1281
+ "strategy": "similar",
1282
+ "score": 88.21638574125416,
1283
+ "scoreCI": [
1284
+ 84.65284421213155,
1285
+ 91.2279308312807
1286
+ ],
1287
+ "valid": 45,
1288
+ "invalid": 0,
1289
+ "missing": 0,
1290
+ "pending": 0,
1291
+ "costUsd": {
1292
+ "known": 0.04251495,
1293
+ "unknownAttempts": 0
1294
+ },
1295
+ "apiCalls": 45,
1296
+ "modelCalls": 45,
1297
+ "inspections": 135,
1298
+ "unknownChargeReservations": 0
1299
+ },
1300
+ {
1301
+ "strategy": "agent",
1302
+ "score": 82.0589167782321,
1303
+ "scoreCI": [
1304
+ 76.74347640483236,
1305
+ 86.50779879163885
1306
+ ],
1307
+ "valid": 44,
1308
+ "invalid": 0,
1309
+ "missing": 1,
1310
+ "pending": 0,
1311
+ "costUsd": {
1312
+ "known": 0.12605145000000004,
1313
+ "unknownAttempts": 0
1314
+ },
1315
+ "apiCalls": 260,
1316
+ "modelCalls": 260,
1317
+ "inspections": 125,
1318
+ "unknownChargeReservations": 0
1319
+ }
1320
+ ],
1321
+ "learnGain": -4.507959568298644,
1322
+ "learnGainCI": [
1323
+ -8.543548654819764,
1324
+ -0.9715515405916082
1325
+ ],
1326
+ "baselines": {
1327
+ "uniformScore": 85.38589005462313,
1328
+ "constantScore": 85.55744059648882,
1329
+ "constantPrediction": {
1330
+ "performed+uniting": 0.2500381443429472,
1331
+ "honest+uniting": 0.2078301757272207,
1332
+ "performed+dividing": 0.2577803629097389,
1333
+ "honest+dividing": 0.2843513170200932
1334
+ }
1335
+ },
1336
+ "valid": 179,
1337
+ "total": 180,
1338
+ "apiCalls": 395,
1339
+ "reportedCostUsd": 0.22916115000000004,
1340
+ "accountedUsd": 0.229161153,
1341
+ "capUsd": 6.827235661,
1342
+ "unknownReservations": 0,
1343
+ "unbilledRateLimits": 0,
1344
+ "apiErrors": 0,
1345
+ "lengthStops": 0,
1346
+ "maxCompletionTokens": 1697,
1347
+ "actualProviders": [
1348
+ "DeepInfra"
1349
+ ],
1350
+ "elapsedMinutes": 8.922566666666667,
1351
+ "offlineResumeVerified": true
1352
+ },
1353
+ {
1354
+ "model": "openai/gpt-oss-120b",
1355
+ "provider": "dekallm/bf16",
1356
+ "quantization": "bf16",
1357
+ "directory": "results/final-8000/openai__gpt-oss-120b/openai-gpt-oss-120b-1fc08149dc845523",
1358
+ "summaries": [
1359
+ {
1360
+ "strategy": "none",
1361
+ "score": 86.52508033881153,
1362
+ "scoreCI": [
1363
+ 83.62381871611136,
1364
+ 89.05738870623047
1365
+ ],
1366
+ "valid": 45,
1367
+ "invalid": 0,
1368
+ "missing": 0,
1369
+ "pending": 0,
1370
+ "costUsd": {
1371
+ "known": 0.002907540000000001,
1372
+ "unknownAttempts": 0
1373
+ },
1374
+ "apiCalls": 45,
1375
+ "modelCalls": 45,
1376
+ "inspections": 0,
1377
+ "unknownChargeReservations": 0
1378
+ },
1379
+ {
1380
+ "strategy": "random",
1381
+ "score": 86.76509071422502,
1382
+ "scoreCI": [
1383
+ 83.46138619015233,
1384
+ 89.72608210727962
1385
+ ],
1386
+ "valid": 45,
1387
+ "invalid": 0,
1388
+ "missing": 0,
1389
+ "pending": 0,
1390
+ "costUsd": {
1391
+ "known": 0.004263209999999998,
1392
+ "unknownAttempts": 0
1393
+ },
1394
+ "apiCalls": 45,
1395
+ "modelCalls": 45,
1396
+ "inspections": 135,
1397
+ "unknownChargeReservations": 0
1398
+ },
1399
+ {
1400
+ "strategy": "similar",
1401
+ "score": 85.63896522044583,
1402
+ "scoreCI": [
1403
+ 81.31909060254966,
1404
+ 89.31213017065038
1405
+ ],
1406
+ "valid": 45,
1407
+ "invalid": 0,
1408
+ "missing": 0,
1409
+ "pending": 0,
1410
+ "costUsd": {
1411
+ "known": 0.00433008,
1412
+ "unknownAttempts": 0
1413
+ },
1414
+ "apiCalls": 45,
1415
+ "modelCalls": 45,
1416
+ "inspections": 135,
1417
+ "unknownChargeReservations": 0
1418
+ },
1419
+ {
1420
+ "strategy": "agent",
1421
+ "score": 82.84109578384674,
1422
+ "scoreCI": [
1423
+ 78.6363115823974,
1424
+ 86.73627093878922
1425
+ ],
1426
+ "valid": 45,
1427
+ "invalid": 0,
1428
+ "missing": 0,
1429
+ "pending": 0,
1430
+ "costUsd": {
1431
+ "known": 0.03264432000000001,
1432
+ "unknownAttempts": 4
1433
+ },
1434
+ "apiCalls": 288,
1435
+ "modelCalls": 284,
1436
+ "inspections": 78,
1437
+ "unknownChargeReservations": 4
1438
+ }
1439
+ ],
1440
+ "learnGain": -3.6839845549647987,
1441
+ "learnGainCI": [
1442
+ -8.056022871667915,
1443
+ 0.2854349259434493
1444
+ ],
1445
+ "baselines": {
1446
+ "uniformScore": 85.38589005462313,
1447
+ "constantScore": 85.55744059648882,
1448
+ "constantPrediction": {
1449
+ "performed+uniting": 0.2500381443429472,
1450
+ "honest+uniting": 0.2078301757272207,
1451
+ "performed+dividing": 0.2577803629097389,
1452
+ "honest+dividing": 0.2843513170200932
1453
+ }
1454
+ },
1455
+ "valid": 180,
1456
+ "total": 180,
1457
+ "apiCalls": 423,
1458
+ "reportedCostUsd": 0.04414515000000001,
1459
+ "accountedUsd": 0.065633801,
1460
+ "capUsd": 6.586004353,
1461
+ "unknownReservations": 4,
1462
+ "unbilledRateLimits": 0,
1463
+ "apiErrors": 4,
1464
+ "lengthStops": 0,
1465
+ "maxCompletionTokens": 1360,
1466
+ "actualProviders": [
1467
+ "DekaLLM"
1468
+ ],
1469
+ "elapsedMinutes": 46.59251666666667,
1470
+ "offlineResumeVerified": true
1471
+ }
1472
+ ]
1473
+ }