DeBERTa ablations and learning curve

README.md

@@ -223,6 +223,14 @@ context features (−0.006 typed F1 [−0.027, +0.020]) or the part-of-speech fe
 documents to 0.143 with 120, then stays flat (0.145 with all 239); span agreement keeps rising
 slowly. See [`reports/crf-ablations/`](reports/crf-ablations/README.md).
 
+**Ablations and learning curve (DeBERTa, held-out half, three seeds each).** DeBERTa needs far
+more data than the CRF: typed agreement is 0.075 with 60 training documents, 0.097 with 120 and
+0.197 with all 239, and every reduced-data seed is clearly worse than the full model. Removing
+heading/list context or the group objective makes no clear difference to held-out typed
+agreement. Without context, the false-prerequisite rate is higher for all three seeds, and
+hand-labelled typed F1 is lower for all three (clear for one seed each). See
+[`reports/neural/ablation-1/`](reports/neural/ablation-1/README.md).
+
 95% project-group bootstrap intervals (1,000 resamples). Full reports, including scores against
 the 28 model-made dev records (which measure agreement with the drafting model, not accuracy),
 are in [`reports/pilot/`](reports/pilot/README.md). The paired difference from adding `batch-01` (+0.214 typed F1,

docs/model_card.md

@@ -143,7 +143,16 @@ Measured on the CRF, as agreement with the 172 model-drafted dev records
 - **Learning curve:** typed agreement rises from 0.054 with 30 training documents to 0.143 with
   120, then stays flat at about 0.14. Span agreement keeps rising slowly.
 
-The group-objective ablation and transformer learning curves need a GPU and have not been run.
+Measured on DeBERTa-v3-small, as agreement with the held-out half (61 documents), three seeds
+each ([`reports/neural/ablation-1/`](../reports/neural/ablation-1/README.md)):
+- **Learning curve:** typed agreement is 0.075 with 60 training documents, 0.097 with 120 and
+  0.197 with 239. Every reduced-data seed is clearly worse, so unlike the CRF, DeBERTa is
+  still data-limited.
+- **Context:** removing it makes no clear difference to held-out typed agreement. The
+  false-prerequisite rate is higher, and hand-labelled typed F1 lower, for all three seeds
+  (each clear for one seed).
+- **Group objective:** not measurable, because no run reproduces a drafted group exactly.
+  Without it, results vary four times as much across seeds.
 
 ## Release process
 

docs/training_on_gpu.md

@@ -68,8 +68,10 @@ On a T4 it takes about 1–1.5 hours:
 ```
 
 Only `triggermap-neural-results.tgz` is needed back. `score_neural.sh` picks one best run per
-grid, which suits model comparisons but not ablations, so these results are scored per variant
-instead.
+grid, which suits model comparisons but not ablations. Score this run with
+`bash scripts/gpu/score_ablations.sh <archive>` instead: it reports every seed and compares
+each variant with the full model trained with the same seed. Results of the first run are in
+[`reports/neural/ablation-1/`](../reports/neural/ablation-1/README.md).
 
 ## Scoring on the development machine
 

reports/crf-ablations/README.md

@@ -75,8 +75,8 @@ Every run, with intervals, is in [`experiment.md`](experiment.md).
   agreement much. The limit is the attribute classifiers (category and role), not the number of
   spans.
 - **The group-objective ablation does not apply to the CRF.** Its group classifier is separate
-  and predicts no matching group. That ablation is planned for the transformer
-  (`configs/experiments/ablations.toml`), which needs a GPU.
+  and predicts no matching group. The transformer's ablations, including the group objective,
+  are in [`../neural/ablation-1/`](../neural/ablation-1/README.md).
 
 Every number here is agreement with model drafts or a 10-document slice; neither is a release
 result.

reports/neural/ablation-1/README.md

@@ -0,0 +1,107 @@
+# Transformer ablations and learning curve (ablation-1, 2026-11-08)
+
+A Kaggle run (one Tesla T4) of
+[`configs/experiments/ablations.toml`](../../../configs/experiments/ablations.toml).
+
+**Recipe:** the best grid cell of neural run 3, with one factor changed at a time:
+- DeBERTa-v3-small, `sqrt` span-tag weights, learning rate 5e-5;
+- up to 20 epochs, patience 3, no early stop before epoch 6;
+- three seeds for every variant.
+
+**Data:**
+- **Training:** the train split of `pilot-2`, `batch-01` and `batch-02` (239 documents), as in
+  runs 1–3.
+- **Early stopping:** the selection half of the model-drafted dev records
+  (`splits/selection-v1/selection.json`, 111 documents). Nothing else was used for it.
+- **Scored on:** the held-out half (61 documents in 46 other project groups) and the 10
+  hand-labelled dev documents. Against model drafts, the scores measure **agreement with the
+  drafting model, not accuracy**.
+- **Untouched:** the test split.
+
+**No run is selected:** every seed is reported, and each variant is compared with the full
+model trained with the same seed.
+
+**Variants:**
+
+| Variant | Change |
+|---|---|
+| `full` | none |
+| `no-context` | no heading or list context in the windows |
+| `no-group` | group loss weight 0 |
+| `docs-60`, `docs-120` | a seeded random subset of 60 or 120 training documents |
+
+## Results (held-out half, 61 documents)
+
+Mean ± SD across the three seeds. The last column is typed F1 on the 10 hand-labelled dev
+documents.
+
+| Variant | Typed F1 | Exact span F1 | Exact span precision | False-prerequisite rate | Hand-labelled typed F1 |
+|---|---:|---:|---:|---:|---:|
+| full | 0.197 ± 0.012 | 0.230 ± 0.012 | 0.166 ± 0.012 | 0.493 ± 0.070 | 0.328 ± 0.043 |
+| no-context | 0.199 ± 0.027 | 0.235 ± 0.028 | 0.168 ± 0.024 | 0.527 ± 0.074 | 0.256 ± 0.035 |
+| no-group | 0.185 ± 0.048 | 0.227 ± 0.059 | 0.163 ± 0.056 | 0.541 ± 0.111 | 0.281 ± 0.093 |
+| docs-120 | 0.097 ± 0.041 | 0.119 ± 0.040 | 0.076 ± 0.031 | 0.697 ± 0.086 | 0.186 ± 0.105 |
+| docs-60 | 0.075 ± 0.031 | 0.090 ± 0.040 | 0.060 ± 0.030 | 0.633 ± 0.047 | 0.104 ± 0.066 |
+
+For reference, the CRF (`crf-batches-01-02`) scores 0.127 typed F1 and a 0.412
+false-prerequisite rate on the same 61 documents ([`../run-3/`](../run-3/README.md)).
+
+### Paired differences against the full model, same seed (B − A)
+
+95% project-group bootstrap intervals, 1,000 resamples. Bold marks a clear difference.
+
+| Variant | Seed | Typed F1, held-out | False-prerequisite rate, held-out | Typed F1, hand-labelled (10) |
+|---|---:|---:|---:|---:|
+| no-context | 0 | +0.019 [−0.021, +0.062] | +0.038 [−0.021, +0.091] | −0.073 [−0.190, +0.021] |
+| no-context | 1 | +0.000 [−0.028, +0.027] | **+0.042 [+0.005, +0.078]** | −0.027 [−0.079, +0.009] |
+| no-context | 2 | −0.012 [−0.044, +0.026] | +0.023 [−0.022, +0.064] | **−0.117 [−0.232, −0.034]** |
+| no-group | 0 | −0.015 [−0.060, +0.028] | +0.031 [−0.024, +0.089] | −0.009 [−0.113, +0.101] |
+| no-group | 1 | +0.030 [−0.003, +0.064] | **−0.070 [−0.117, −0.024]** | +0.030 [−0.054, +0.134] |
+| no-group | 2 | **−0.051 [−0.081, −0.017]** | **+0.183 [+0.132, +0.238]** | **−0.162 [−0.325, −0.018]** |
+| docs-120 | 0 | **−0.120 [−0.180, −0.063]** | **+0.240 [+0.160, +0.312]** | **−0.215 [−0.308, −0.110]** |
+| docs-120 | 1 | **−0.057 [−0.092, −0.024]** | **+0.062 [+0.012, +0.108]** | +0.024 [−0.067, +0.104] |
+| docs-120 | 2 | **−0.124 [−0.169, −0.083]** | **+0.310 [+0.256, +0.369]** | **−0.233 [−0.428, −0.064]** |
+| docs-60 | 0 | **−0.162 [−0.219, −0.110]** | **+0.209 [+0.122, +0.293]** | **−0.335 [−0.565, −0.122]** |
+| docs-60 | 1 | **−0.125 [−0.174, −0.083]** | **+0.110 [+0.060, +0.166]** | **−0.133 [−0.307, −0.036]** |
+| docs-60 | 2 | **−0.078 [−0.123, −0.031]** | **+0.101 [+0.038, +0.170]** | **−0.204 [−0.431, −0.003]** |
+
+Full reports: `compare_holdout_*.md`, `compare_manual_*.md`, and `<run>_dev_holdout.md` /
+`<run>_dev_manual.md` for every run. Selection-half scores are in
+[`experiment.md`](experiment.md).
+
+## Reading
+
+- **DeBERTa needs much more data than the CRF.** With 60 or 120 training documents, every seed
+  is clearly worse than with all 239, on typed agreement and on invented prerequisites. Typed
+  agreement roughly doubles from 120 to 239 documents (0.097 → 0.197). The CRF, by contrast,
+  stops improving at about 120 documents ([`../../crf-ablations/`](../../crf-ablations/README.md);
+  measured on all 172 agreement records). So more training documents would likely raise
+  DeBERTa's agreement further. With model-drafted labels, though, that means agreement with the
+  drafting model, not shown accuracy.
+- **Heading and list context: no clear effect on held-out agreement,** but every sign points
+  the same way. Without context, the false-prerequisite rate is higher for all three seeds
+  (+0.023 to +0.042, clear for one), and hand-labelled typed F1 is lower for all three (0.328 →
+  0.256 on average, clear for one seed). That is weak evidence that context helps; 61 and 10
+  documents cannot settle it.
+- **Group objective: not measurable here.** No run reproduces a single drafted group exactly
+  (group F1 0.000 everywhere), so its own target cannot be scored. Its effect on spans is
+  inconsistent: one seed is clearly better without it, and one is clearly worse.
+- **Training is unstable at this data size.** The worse seeds stopped early: `no-group` seed 2
+  peaked at epoch 6, and `docs-120` seeds 0 and 2 at epochs 9 and 7. Without the group loss,
+  the spread across seeds is four times larger (SD 0.048 against 0.012). Single runs should
+  not be trusted; seed means are the unit of comparison.
+- **The recipe reproduces.** The three `full` seeds score 0.208, 0.198 and 0.185 held-out.
+  Run 3 scored 0.207, 0.198 and 0.185 with the same settings.
+
+## Provenance
+
+- Results archive SHA-256 `34451e81108226bb1067b8a458b7b1491c9979aa4309e8e69237c5044a5f3240`.
+- Per-run `metadata/` (configuration, base revision, training hash and origins, per-epoch
+  losses and dev scores, peak GPU memory) and every run's `predictions/` are committed. The
+  weights are not.
+
+To rescore from the archive:
+
+```bash
+bash scripts/gpu/score_ablations.sh ~/Downloads/triggermap-neural-results.tgz
+```

reports/neural/ablation-1/ablation-1-docs-120-seed0_dev_holdout.md

@@ -0,0 +1,57 @@
+# Agreement with model drafts (held-out half): ablation-1-docs-120-seed0
+
+- Gold: `annotations/batches/batch-03.jsonl`, `annotations/batches/batch-06.jsonl` (dev split of `splits/pool-v1.json`; origin: all; subset `splits/selection-v1/holdout.json`; 61 documents)
+- Predictions: `reports/neural/ablation-1/predictions/ablation-1-docs-120-seed0.jsonl` (ablation-1-docs-120-seed0)
+- Bootstrap: 1000 resamples, project clusters, seed 0
+- Documents: 61 (57 with gold conditions, 4 without)
+
+## Headline metrics (95% bootstrap intervals)
+
+| Metric | Estimate | 95% CI |
+|---|---:|:---:|
+| Typed condition micro-F1 (span + category + role) | 0.089 | [0.057, 0.122] |
+| Exact span F1 | 0.109 | [0.072, 0.148] |
+| Exact span precision | 0.068 | [0.044, 0.094] |
+| Exact span recall | 0.271 | [0.194, 0.339] |
+| Role macro-F1 (roles with gold support) | 0.049 | [0.021, 0.100] |
+| Category macro-F1 | 0.070 | [0.042, 0.100] |
+| Group exact-match F1 | 0.000 | [0.000, 0.000] |
+| Complete-document exact match, documents with conditions | 0.000 | [0.000, 0.000] |
+| Complete-document exact match, documents without conditions | 0.000 | [0.000, 0.000] |
+| False prerequisite rate (lower is better) | 0.668 | [0.602, 0.723] |
+| Reversed meaning rate (lower is better) | 0.033 | [0.006, 0.063] |
+| Category accuracy on exactly matched spans | 0.842 | [0.717, 0.951] |
+| Role accuracy on exactly matched spans | 0.974 | [0.905, 1.000] |
+
+## Per-role results (exact span + role)
+
+| Label | Precision | Recall | F1 | Gold support |
+|---|---:|---:|---:|---:|
+| REQUIRED | 0.070 | 0.292 | 0.113 | 120 |
+| EXCEPTION | 0.000 | 0.000 | 0.000 | 2 |
+| NOT_REQUIRED | 0.053 | 0.200 | 0.083 | 10 |
+| UNCERTAIN | 0.000 | 0.000 | 0.000 | 8 |
+
+## Per-category results (exact span + category)
+
+| Label | Precision | Recall | F1 | Gold support |
+|---|---:|---:|---:|---:|
+| CONFIGURATION | 0.014 | 0.056 | 0.023 | 18 |
+| USAGE | 0.064 | 0.315 | 0.106 | 54 |
+| ACCESS | 0.067 | 0.241 | 0.105 | 54 |
+| PLATFORM | 0.033 | 0.071 | 0.045 | 14 |
+
+## Precision versus coverage
+
+| Coverage | Confidence ≥ | Typed precision | Retained recall |
+|---:|---:|---:|---:|
+| 10% | 0.559 | 0.089 | 0.036 |
+| 20% | 0.519 | 0.089 | 0.071 |
+| 30% | 0.490 | 0.089 | 0.107 |
+| 40% | 0.463 | 0.089 | 0.143 |
+| 50% | 0.438 | 0.079 | 0.157 |
+| 60% | 0.413 | 0.083 | 0.200 |
+| 70% | 0.387 | 0.077 | 0.214 |
+| 80% | 0.346 | 0.067 | 0.214 |
+| 90% | 0.283 | 0.062 | 0.221 |
+| 100% | 0.150 | 0.055 | 0.221 |

reports/neural/ablation-1/ablation-1-docs-120-seed0_dev_manual.md

@@ -0,0 +1,57 @@
+# Hand-labelled dev: ablation-1-docs-120-seed0
+
+- Gold: `annotations/pilot/pilot-2.jsonl` (dev split of `splits/pool-v1.json`; origin: manual; 10 documents)
+- Predictions: `reports/neural/ablation-1/predictions/ablation-1-docs-120-seed0.jsonl` (ablation-1-docs-120-seed0)
+- Bootstrap: 1000 resamples, project clusters, seed 0
+- Documents: 10 (9 with gold conditions, 1 without)
+
+## Headline metrics (95% bootstrap intervals)
+
+| Metric | Estimate | 95% CI |
+|---|---:|:---:|
+| Typed condition micro-F1 (span + category + role) | 0.148 | [0.000, 0.346] |
+| Exact span F1 | 0.173 | [0.038, 0.367] |
+| Exact span precision | 0.109 | [0.023, 0.250] |
+| Exact span recall | 0.412 | [0.091, 0.714] |
+| Role macro-F1 (roles with gold support) | 0.096 | [0.021, 0.340] |
+| Category macro-F1 | 0.105 | [0.000, 0.333] |
+| Group exact-match F1 | 0.100 | [0.000, 0.353] |
+| Complete-document exact match, documents with conditions | 0.111 | [0.000, 0.333] |
+| Complete-document exact match, documents without conditions | 1.000 | [0.000, 1.000] |
+| False prerequisite rate (lower is better) | 0.632 | [0.425, 0.768] |
+| Reversed meaning rate (lower is better) | 0.000 | [0.000, 0.000] |
+| Category accuracy on exactly matched spans | 0.857 | [0.000, 1.000] |
+| Role accuracy on exactly matched spans | 1.000 | [1.000, 1.000] |
+
+## Per-role results (exact span + role)
+
+| Label | Precision | Recall | F1 | Gold support |
+|---|---:|---:|---:|---:|
+| REQUIRED | 0.123 | 0.438 | 0.192 | 16 |
+| EXCEPTION | 0.000 | 0.000 | 0.000 | 0 |
+| NOT_REQUIRED | 0.000 | 0.000 | 0.000 | 0 |
+| UNCERTAIN | 0.000 | 0.000 | 0.000 | 1 |
+
+## Per-category results (exact span + category)
+
+| Label | Precision | Recall | F1 | Gold support |
+|---|---:|---:|---:|---:|
+| CONFIGURATION | 0.000 | 0.000 | 0.000 | 1 |
+| USAGE | 0.105 | 0.444 | 0.170 | 9 |
+| ACCESS | 0.182 | 0.400 | 0.250 | 5 |
+| PLATFORM | 0.000 | 0.000 | 0.000 | 2 |
+
+## Precision versus coverage
+
+| Coverage | Confidence ≥ | Typed precision | Retained recall |
+|---:|---:|---:|---:|
+| 9% | 0.611 | 0.000 | 0.000 |
+| 20% | 0.565 | 0.077 | 0.059 |
+| 30% | 0.530 | 0.158 | 0.176 |
+| 41% | 0.497 | 0.192 | 0.294 |
+| 50% | 0.488 | 0.156 | 0.294 |
+| 59% | 0.452 | 0.132 | 0.294 |
+| 70% | 0.405 | 0.111 | 0.294 |
+| 80% | 0.381 | 0.098 | 0.294 |
+| 91% | 0.313 | 0.103 | 0.353 |
+| 100% | 0.163 | 0.094 | 0.353 |

reports/neural/ablation-1/ablation-1-docs-120-seed1_dev_holdout.md

@@ -0,0 +1,57 @@
+# Agreement with model drafts (held-out half): ablation-1-docs-120-seed1
+
+- Gold: `annotations/batches/batch-03.jsonl`, `annotations/batches/batch-06.jsonl` (dev split of `splits/pool-v1.json`; origin: all; subset `splits/selection-v1/holdout.json`; 61 documents)
+- Predictions: `reports/neural/ablation-1/predictions/ablation-1-docs-120-seed1.jsonl` (ablation-1-docs-120-seed1)
+- Bootstrap: 1000 resamples, project clusters, seed 0
+- Documents: 61 (57 with gold conditions, 4 without)
+
+## Headline metrics (95% bootstrap intervals)
+
+| Metric | Estimate | 95% CI |
+|---|---:|:---:|
+| Typed condition micro-F1 (span + category + role) | 0.141 | [0.103, 0.198] |
+| Exact span F1 | 0.163 | [0.119, 0.226] |
+| Exact span precision | 0.110 | [0.078, 0.159] |
+| Exact span recall | 0.314 | [0.237, 0.401] |
+| Role macro-F1 (roles with gold support) | 0.056 | [0.039, 0.099] |
+| Category macro-F1 | 0.102 | [0.073, 0.140] |
+| Group exact-match F1 | 0.000 | [0.000, 0.000] |
+| Complete-document exact match, documents with conditions | 0.018 | [0.000, 0.052] |
+| Complete-document exact match, documents without conditions | 0.750 | [0.000, 1.000] |
+| False prerequisite rate (lower is better) | 0.629 | [0.546, 0.700] |
+| Reversed meaning rate (lower is better) | 0.056 | [0.013, 0.109] |
+| Category accuracy on exactly matched spans | 0.886 | [0.762, 0.978] |
+| Role accuracy on exactly matched spans | 0.955 | [0.889, 1.000] |
+
+## Per-role results (exact span + role)
+
+| Label | Precision | Recall | F1 | Gold support |
+|---|---:|---:|---:|---:|
+| REQUIRED | 0.128 | 0.342 | 0.186 | 120 |
+| EXCEPTION | 0.000 | 0.000 | 0.000 | 2 |
+| NOT_REQUIRED | 0.000 | 0.000 | 0.000 | 10 |
+| UNCERTAIN | 0.022 | 0.125 | 0.037 | 8 |
+
+## Per-category results (exact span + category)
+
+| Label | Precision | Recall | F1 | Gold support |
+|---|---:|---:|---:|---:|
+| CONFIGURATION | 0.034 | 0.111 | 0.052 | 18 |
+| USAGE | 0.107 | 0.389 | 0.167 | 54 |
+| ACCESS | 0.139 | 0.296 | 0.189 | 54 |
+| PLATFORM | 0.000 | 0.000 | 0.000 | 14 |
+
+## Precision versus coverage
+
+| Coverage | Confidence ≥ | Typed precision | Retained recall |
+|---:|---:|---:|---:|
+| 10% | 0.844 | 0.200 | 0.057 |
+| 20% | 0.735 | 0.225 | 0.129 |
+| 30% | 0.667 | 0.200 | 0.171 |
+| 40% | 0.611 | 0.175 | 0.200 |
+| 50% | 0.555 | 0.145 | 0.207 |
+| 60% | 0.514 | 0.121 | 0.207 |
+| 70% | 0.466 | 0.114 | 0.229 |
+| 80% | 0.429 | 0.116 | 0.264 |
+| 90% | 0.360 | 0.103 | 0.264 |
+| 100% | 0.188 | 0.095 | 0.271 |

reports/neural/ablation-1/ablation-1-docs-120-seed1_dev_manual.md

@@ -0,0 +1,57 @@
+# Hand-labelled dev: ablation-1-docs-120-seed1
+
+- Gold: `annotations/pilot/pilot-2.jsonl` (dev split of `splits/pool-v1.json`; origin: manual; 10 documents)
+- Predictions: `reports/neural/ablation-1/predictions/ablation-1-docs-120-seed1.jsonl` (ablation-1-docs-120-seed1)
+- Bootstrap: 1000 resamples, project clusters, seed 0
+- Documents: 10 (9 with gold conditions, 1 without)
+
+## Headline metrics (95% bootstrap intervals)
+
+| Metric | Estimate | 95% CI |
+|---|---:|:---:|
+| Typed condition micro-F1 (span + category + role) | 0.305 | [0.197, 0.455] |
+| Exact span F1 | 0.339 | [0.233, 0.480] |
+| Exact span precision | 0.238 | [0.151, 0.387] |
+| Exact span recall | 0.588 | [0.444, 0.727] |
+| Role macro-F1 (roles with gold support) | 0.204 | [0.164, 0.500] |
+| Category macro-F1 | 0.199 | [0.130, 0.404] |
+| Group exact-match F1 | 0.167 | [0.000, 0.421] |
+| Complete-document exact match, documents with conditions | 0.111 | [0.000, 0.300] |
+| Complete-document exact match, documents without conditions | 1.000 | [0.000, 1.000] |
+| False prerequisite rate (lower is better) | 0.545 | [0.360, 0.698] |
+| Reversed meaning rate (lower is better) | 0.000 | [0.000, 0.000] |
+| Category accuracy on exactly matched spans | 0.900 | [0.700, 1.000] |
+| Role accuracy on exactly matched spans | 1.000 | [1.000, 1.000] |
+
+## Per-role results (exact span + role)
+
+| Label | Precision | Recall | F1 | Gold support |
+|---|---:|---:|---:|---:|
+| REQUIRED | 0.303 | 0.625 | 0.408 | 16 |
+| EXCEPTION | 0.000 | 0.000 | 0.000 | 0 |
+| NOT_REQUIRED | 0.000 | 0.000 | 0.000 | 0 |
+| UNCERTAIN | 0.000 | 0.000 | 0.000 | 1 |
+
+## Per-category results (exact span + category)
+
+| Label | Precision | Recall | F1 | Gold support |
+|---|---:|---:|---:|---:|
+| CONFIGURATION | 0.000 | 0.000 | 0.000 | 1 |
+| USAGE | 0.222 | 0.667 | 0.333 | 9 |
+| ACCESS | 0.375 | 0.600 | 0.462 | 5 |
+| PLATFORM | 0.000 | 0.000 | 0.000 | 2 |
+
+## Precision versus coverage
+
+| Coverage | Confidence ≥ | Typed precision | Retained recall |
+|---:|---:|---:|---:|
+| 10% | 0.916 | 0.750 | 0.176 |
+| 19% | 0.784 | 0.500 | 0.235 |
+| 31% | 0.710 | 0.385 | 0.294 |
+| 40% | 0.542 | 0.294 | 0.294 |
+| 50% | 0.522 | 0.286 | 0.353 |
+| 60% | 0.507 | 0.280 | 0.412 |
+| 69% | 0.498 | 0.241 | 0.412 |
+| 81% | 0.447 | 0.235 | 0.471 |
+| 90% | 0.391 | 0.211 | 0.471 |
+| 100% | 0.307 | 0.214 | 0.529 |