For Minegishi: training documentation, reproducibility files and line plots 1/4

.gitattributes

@@ -54,3 +54,7 @@ loop_transformer_halting_2026-09-24/data/chain_g2_k8/tests.json filter=lfs diff=
 loop_transformer_halting_2026-09-24/data/chain_loop_k8/tests.json filter=lfs diff=lfs merge=lfs -text
 loop_transformer_halting_2026-09-24/data/chain_w43_k8/tests.json filter=lfs diff=lfs merge=lfs -text
 loop_transformer_halting_2026-09-24/data/chain_loop/tests.json filter=lfs diff=lfs merge=lfs -text
+for[[:space:]]Minegishi/figures/comparison_lines_d128.png filter=lfs diff=lfs merge=lfs -text
+for[[:space:]]Minegishi/figures/comparison_lines_d16.png filter=lfs diff=lfs merge=lfs -text
+for[[:space:]]Minegishi/figures/comparison_lines_d2.png filter=lfs diff=lfs merge=lfs -text
+for[[:space:]]Minegishi/figures/comparison_lines_d32.png filter=lfs diff=lfs merge=lfs -text

for Minegishi/README.md

@@ -0,0 +1,184 @@
+# Checkpoints for Minegishi: two wait-consistency methods
+
+This package contains **all 23 evaluation checkpoints from each of two completed runs** (46 checkpoints in total), plus the training configuration, frozen model/training code, synthetic data, evaluation records, and step-wise mechanism plots.
+
+Both runs use **RoPE base 10,000, seed 7, 200,000 optimizer updates, full shallow anchor supervision, and random entity-prefix augmentation**. They were initialized from scratch with the same named-parameter initialization. The difference is the additional wait-consistency objective. These are the original completed runs, not subsequent long-prefix continuation experiments.
+
+## Directory layout
+
+```text
+for Minegishi/
+  README.md
+  checkpoint_index.json
+  SHA256SUMS
+  pure_wait_cs_random_t/           # original run: R10k_PW_s7
+    ckpt_0001000.pt ... ckpt_0200000.pt
+    README.md
+    manifest.json
+    config.json
+    metrics.jsonl
+    train_log.jsonl
+    final_eval.json
+    checkpoints.json
+    train_command.json
+  wait_denoising_fixed_t0/         # original run: R10k_T0iso_s7
+    ... same files ...
+  figures/                        # final LINE plots, including all seven depths
+  analysis/                       # numerical probe summaries and plotting inputs
+  reproducibility/                # frozen scientific source and synthetic data
+    load_checkpoint.py
+    model/loop_gpt.py
+    train_halt.py
+    objectives.py
+    data/chain_loop_k8/
+    data/atomic_joint_2026-09-19/
+```
+
+Checkpoint steps are **1k, 2k, 5k, then every 10k from 10k through 200k**. Each named checkpoint was saved before its corresponding evaluation. All are included; none was selected by deep-test performance. The primary final checkpoint is `ckpt_0200000.pt`.
+
+These `.pt` files contain `model`, `update`, `rope_base`, and `max_pos`. They are model-weight snapshots for evaluation, **not full optimizer/RNG resume states**. The source training run used a separate `last.pt` for exact recovery; that mutable recovery file is not one of these 46 evaluation checkpoints. SHA-256 digests and sizes are listed in `checkpoint_index.json`, each method's `checkpoints.json`, and `SHA256SUMS`.
+
+## Shared model and training recipe
+
+| Setting | Value |
+|---|---|
+| Architecture | Full-sequence causal, weight-shared loop Transformer |
+| One loop | One pass through the same **four** Transformer blocks |
+| Width / heads / MLP width | 768 / 12 / 3072 |
+| Parameters | 28,756,225, including an unused 769-parameter stop head |
+| Positional encoding | RoPE base 10,000; cache capacity 4,097; no position interpolation/scaling |
+| Attention | Full causal attention; no hard previous-token/local-window restriction |
+| Input/output embeddings | Tied token embedding and LM readout |
+| State update | Embed input once, then repeatedly apply the shared stack; no re-embedding between loops |
+| Blocks | Pre-LayerNorm, epsilon 1e-5; GELU with tanh approximation; dropout 0 |
+| Initialization | Seed 7; parameter-name-specific normal initialization (std 0.02); residual output projection weights zero |
+| Optimizer | AdamW, beta=(0.9,0.999), epsilon 1e-8, weight decay 0.01 |
+| Learning rate | Peak 1e-4; 2,000-step linear warm-up, then linear decay to zero at step 200,000 |
+| Gradient clipping | Global norm 1.0 |
+| Precision | FP32 parameters with TF32 training; strict FP32/TF32-disabled evaluation |
+| Recorded framework | PyTorch 2.7.1+cu128 |
+| Training implementation | CUDA graph capture/replay; one full H100 per run |
+| Training duration | Fixed 200,000 updates |
+| Additional consistency | Starts at update 20,000; 32 auxiliary rows per update, each of semantic depth 32 |
+
+The command-line field `b_train=8` is a legacy halting-budget argument: **it is not this experiment's batch size**. The actual supervised batch is 128 queries per update:
+
+- 32 atomic queries;
+- 16 two-query (`w2`) sources rendered as 32 independent atomic queries;
+- 64 depth-2 composition queries.
+
+Thus each update has 64 depth-1 and 64 depth-2 supervised queries. The supervised stream contains 25.6 million query presentations over 200k updates. The extra 32 consistency rows are separate from those 128 supervised rows.
+
+The synthetic task has 500 entities and 20 relation maps. Inputs are compact token sequences `[e0, r1, ..., rd]`; the target is the entity reached by composing the relations. The frozen data include 10,000 atomic facts, 40,000 w2 sources, and 80,000 depth-2 training examples. No depth>2 answer labels enter these two runs' training. Depth-32 auxiliary rows supply consistency supervision only.
+
+### Shallow CE and the full shallow anchor
+
+Every shallow batch is unrolled for **four complete loops**. Final-answer cross-entropy is averaged over loops 1,2,3 for depth-1 examples and loops 2,3,4 for depth-2 examples; these per-query losses are then averaged over the 128 queries.
+
+The full anchor adds hidden-state supervision throughout these four loops. At each position and loop:
+
+- the initial entity and random entity-prefix tokens target their original embeddings;
+- a relation that has not reached its execution step targets its original relation embedding;
+- a relation at or after its execution step targets the embedding of its correct intermediate entity.
+
+Targets are detached. For each loop, squared errors over all valid positions are divided by the total squared norm of the target embeddings; the four loop losses are averaged. The outer anchor coefficient is **1.0**. This anchor includes root, execute, wait, and hold targets; it uses the available shallow intermediate labels. Consequently, neither method should be described as “CE + wait CS alone.”
+
+For each supervised row, the chain is preceded by a uniformly sampled **0–128 random entity tokens**. These prefix tokens participate in the shallow root anchor. The prefix has no answer-CE target. The separate consistency rows have 33 positions (`entity + 32 relations`), without this added prefix.
+
+The total objective is `shallow CE + 1.0 * shallow anchor + 1.0 * selected CS` once CS is active. Before update 20k, the selected-CS contribution is zero.
+
+## Method 1: pure wait CS, random t
+
+Folder: [`pure_wait_cs_random_t`](pure_wait_cs_random_t); mode `wait_no_root`.
+
+For every auxiliary row, sample a root entity, 32 relations, and an integer `t` uniformly from **0,...,31**. Construct a detached state by placing independently sampled entity embeddings at positions 1,...,t, leaving the later positions as relation embeddings. These sampled entities are not the correct intermediate labels of the deep chain. Position `t+1` is the execution frontier; positions `p > t+1` are waiting positions.
+
+Apply the shared stack **once** and penalize only waiting positions that move away from their input embeddings. No noise is added. No root or hold term is included in this auxiliary CS loss; the separate shallow anchor still supervises those roles.
+
+For an output `y` and detached target `z`, the per-position discrepancy is:
+
+`rho(y,z) = ||y-z||² / (||y||² + ||z||²)`.
+
+The selected loss sums `rho` at waiting positions and divides by **32 rows × 32 non-frontier positions**. The denominator is retained from the full non-frontier objective; it is not renormalized by the number of selected waiting positions. A row sampled at t=31 contributes no waiting-position numerator.
+
+## Method 2: fixed t=0, isotropic wait denoising
+
+Folder: [`wait_denoising_fixed_t0`](wait_denoising_fixed_t0); mode `t0_wait_denoise`.
+
+Each auxiliary row begins as `[e0,r1,...,r32]`. Time is fixed at **t=0**, so position 1 is the frontier and positions **2,...,32** are waiting. Noise is added only at those waiting positions; root position 0 and frontier position 1 are left clean.
+
+For each perturbed token embedding `z`, find the nearest **other** current entity/relation embedding, excluding special tokens and the token itself. Let that Euclidean distance be `delta`. Sample `alpha ~ Uniform[0,0.2)` and an isotropic direction by normalizing a Gaussian vector. The perturbation is `alpha * delta * direction`.
+
+The radius is therefore a fraction of the nearest-other-embedding distance, **not 20% of the token embedding's own norm**. The model runs one complete loop on the detached noisy state and is trained to return waiting positions to their **clean detached input embeddings**, using the same bounded discrepancy `rho`.
+
+The selected loss is **0.5 × sum(waiting-position rho) / (32 rows × 32)**. The outer CS weight remains 1.0. The internal factor 0.5 matches the expected selected-position mass of the random-t wait loss; it does not assert equal gradient magnitudes. The noise generator is an independent named Torch RNG stream. Noise is used in training only.
+
+## Evaluation and final results
+
+Training-time evaluations used the fixed steps listed above. The left-column plots show train-d2 fit (8,000 fixed training examples), validation d2 (1,000), and monitors at d4/8/16/32/64/128 (100 examples in each of three categories, 300 per depth).
+
+Formal final evaluation used the **200k checkpoint**, eight depths `2,3,4,8,16,32,64,128`, three categories `all_seen/one_new/random`, and 5,000 examples per cell: **120,000 examples per method**. Inference uses `R=d`, clean input, no added prefix, strict FP32, and the full-vocabulary argmax at the last input position.
+
+| Method | Formal errors / examples | Accuracy |
+|---|---:|---:|
+| Pure wait CS, random t | 0 / 120,000 | 100% |
+| Fixed t=0, wait isotropic denoising | 3 / 120,000 | 99.9975% |
+
+For the denoising run, the three errors occur in `all_seen|d128`, `one_new|d16`, and `random|d8` (one error each). Exact cell records are in each method's `final_eval.json`. These are single-seed results. This release does not make a completed-result claim about depths beyond 128.
+
+## Step-wise performance and mechanism plots
+
+![Line comparison at input depth 128](figures/comparison_lines_d128.png)
+
+- [All seven input depths, line plots (PDF)](figures/three_column_lines_all_depths.pdf)
+- [Main line plot, vector SVG](figures/comparison_lines_d128.svg)
+- [Fixed-loop error curves with three reference definitions (PDF)](figures/fixed_loop_error_curves.pdf)
+
+Each row is one method; all columns have the same **training-step** axis. All 23 measured checkpoints are shown without smoothing. The middle/right probe panels use a fixed subset of **32 examples per category, 96 chains per depth**, at d2/4/8/16/32/64/128. These small probes are distinct from the formal 120k-example assessment.
+
+- **Left:** train, validation, and depth-generalization final-answer accuracy.
+- **Middle:** root/hold/wait cumulative relative L2 drift at fixed loop counts R=2,4,8,16,32,64,128. Root/wait are compared with original input embeddings. Hold is compared with its own state immediately after its first expected execution; comparison with the correct entity embedding is also available in the supplemental error PDF. Empty roles are omitted, not treated as zero. A frontier token is excluded from retention error.
+- **Right, actual continuous rollout:** run from the original input without injecting gold states, and measure the frontier's full-vocabulary accuracy at position p after loop p. Earlier model errors propagate naturally.
+- **Right, one step from a clean correct state:** construct the correct intermediate entity embeddings through p-1, leave later relation embeddings intact, run one loop, and evaluate frontier p. This is a separate controlled one-step test; it does not replace the actual model trajectory.
+
+These plots show associations across training, not a causal proof of grokking. Correct entity readout also does not imply that the hidden vector exactly equals that entity's embedding or can already be reused robustly by later positions.
+
+## Download and load
+
+This is a custom PyTorch model, not an `AutoModelForCausalLM` checkpoint. Download the subfolder, then use the included loader:
+
+```python
+from huggingface_hub import snapshot_download
+
+snapshot = snapshot_download(
+    repo_id="Physicsru/compgen-checkpoints",
+    allow_patterns=["for Minegishi/**"],
+)
+```
+
+```python
+import sys
+from pathlib import Path
+import torch
+
+bundle = Path(snapshot) / "for Minegishi"
+sys.path.insert(0, str(bundle / "reproducibility"))
+from load_checkpoint import load_checkpoint, predict
+
+model, vocab, metadata = load_checkpoint(
+    bundle / "pure_wait_cs_random_t" / "ckpt_0200000.pt",
+    device="cuda",
+)
+ids = {token: i for i, token in enumerate(vocab)}
+query = torch.tensor([[ids["<e_0>"], ids["<r_0>"], ids["<r_1>"]]], device="cuda")
+prediction = predict(model, query, loops=2)
+print(vocab[prediction.item()])
+```
+
+Use an integer token sequence of length d+1 and normally run d complete loops. The output is a full-vocabulary token ID. Construct the model with **RoPE base 10,000** and the saved cache capacity, because RoPE buffers are non-persistent and must be recreated; silently using the older default base 100 changes the model. The loader checks metadata and loads all parameters strictly. The vocabulary and synthetic relation map are included.
+
+`reproducibility/train_halt.py`, `objectives.py`, `model/loop_gpt.py`, and the per-method `train_command.json` describe the exact training implementation. Commands use paths relative to this package root. Running training creates a new output directory; the original evaluation snapshots need not be modified. The archived trainer supports other historical arms as well: use the included command/config rather than defaults or its older general-purpose header documentation.
+
+## Provenance
+
+Prepared on 2026-09-25 for Minegishi from the two named TSUBAME runs. Checkpoint identities were cross-checked against their saved evaluation receipts and SHA-256 hashes. Model/training source hashes were checked against the deployed source. No training, checkpoint rewriting, score-based checkpoint selection, or further model adaptation was performed while preparing this package. The independent mechanism-probe job completed normally and made zero training updates.

for Minegishi/analysis/RESULTS_zh.md

@@ -0,0 +1,42 @@
+# 两种 wait CS：逐 checkpoint 的保持误差与 frontier 执行
+
+2026-09-25 补测完成。比较 RoPE=10000、seed7 的纯 wait CS（随机 t）与固定 t=0、wait 各向同性去噪；两项都保留完整浅层 anchor 与训练 prefix 0–128。每项全 23 个 checkpoint，共 46 个；没有重新训练或按表现筛选 checkpoint。
+
+左栏沿用原训练记录：train d2 固定 8000 条、validation d2 1000 条、每深度三类监控题各 100 条。中/右栏为本次独立推理：d2/4/8/16/32/64/128，每类固定前 32 条，共 96 条/深度；真实轨迹连续展开 128 个 loop，另对每个合法 frontier 位置测干净正确中间态的一步执行。FP32、TF32 关闭、无测试 prefix、无测试噪声；一个 loop 是四个共享 Transformer block。
+
+## 最主要的观察
+
+**原子操作的正确读出很早就出现，长链连续执行的改善明显滞后。** 在 5k、d128 的干净中间态对照中，纯 wait / 去噪的 frontier 平均准确率为 99.967% / 99.951%；真实展开中的 frontier 平均准确率两者均为 1.742%。这里对各 frontier 位置及固定样本求平均，同一条链的这些观测不是独立样本。正确读出不保证输出向量已接近 entity embedding，更不保证下一位置能稳定复用它。
+
+**泛化跃升与 wait 漂移下降相伴，但并非三个角色的误差都同步归零。** 以下准确率来自左栏原 300 题/深度的监控曲线；误差来自本次 96 条链。误差均为相对 L2，表内 wait 取 R=64，root/hold 取 R=128。
+
+| 方法与区间 | d128 最终答案准确率 | wait 累计误差 | root 累计误差 | hold 相对首次执行输出的误差 |
+|---|---:|---:|---:|---:|
+| 纯 wait：120k → 130k | 84.67% → 99.67% | 0.961 → 0.285 | 1.903 → 0.967 | 1.055 → 1.052 |
+| t=0 去噪：130k → 140k | 86.00% → 99.67% | 1.148 → 0.256 | 2.480 → 0.956 | 1.164 → 0.923 |
+
+这些是同一批 checkpoint 的观测关联，不是“wait 误差下降导致 grokking”的因果证明。10k 的记录间隔也无法确定跃升发生的精确训练步。
+
+## hold 的参照尤其重要
+
+root/wait 的累计误差参照原输入 embedding；hold 的累计误差参照该位置刚完成预期执行时的自身输出。hold 原来的 relation 应被替换成 entity，不能继续拿原 relation 当保持目标。以下均为 200k、d128：
+
+| 指标 | 纯 wait，随机 t | 固定 t=0，wait 去噪 |
+|---|---:|---:|
+| 真实 frontier 平均准确率（探针） | 100% | 100% |
+| root 在 R=128 的累计相对 L2 | 0.01593 | 0.00941 |
+| wait 在 R=64 的累计相对 L2 | 0.00546 | 0.01647 |
+| hold 在 R=128，相对首次执行输出 | 0.13077 | 0.12127 |
+| hold 在 R=128，相对正确 entity embedding | 0.01425 | 0.01311 |
+
+因此，hold 最终已经接近正确 entity embedding，但它相对首次执行输出仍有约 12–13% 的位移。不能把这一位移直接解释成失稳，或要求已泛化模型的所有隐状态严格不变。固定 loop 曲线的第三组页专门给出正确语义 embedding 参照，第一组是自身保持，第二组是单轮漂移。
+
+## 图表与审计
+
+- figures/comparison_d128.png：三栏主图；第一行纯 wait，第二行去噪。中央为 root/hold/wait 的训练 step × loop 误差热图；右侧分别为真实展开与干净中间态对照的 frontier 准确率。
+- figures/three_column_all_depths.pdf：上述七个输入深度的完整三栏图。
+- figures/fixed_loop_error_curves.pdf：R=2/4/8/16/32/64/128 的误差 vs 训练 step，包含三种参照 × 七个深度，共 21 页。
+- analysis.json：逐深度、逐 checkpoint 的数值摘要、阈值统计和原结果 SHA256。
+- results/：全部逐 loop 指标、计数、样本身份、模型/数据/checkpoint 哈希。
+
+10 项 CPU 测试通过；46 个 checkpoint、1932 个深度/类别/展开方式组合、158700 行 loop 记录已核对。空角色严格显示 NA；例如 d2 在 R≥2 没有 wait，不能记作零误差。探针的 96 条/深度不是正式 5000 条/类别评测，不能据此重报正式零错误结论。样本量与单 seed 限制仍然存在。

for Minegishi/analysis/protocol.json

@@ -0,0 +1,23 @@
+{
+  "depths": [
+    2,
+    4,
+    8,
+    16,
+    32,
+    64,
+    128
+  ],
+  "worlds": [
+    "all_seen",
+    "one_new",
+    "random"
+  ],
+  "samples_per_cell": 32,
+  "loop_horizon": 128,
+  "checkpoints": 46,
+  "precision": "FP32; TF32 off",
+  "training_updates": 0,
+  "checkpoint_selection": "all recorded evaluations",
+  "source_log_sha256": "2f74c7e4ae74f63d5ade4985a066cc9cf9806be4977e1e6806b321cd5dc4c6e6"
+}

for Minegishi/checkpoint_index.json

@@ -0,0 +1,511 @@
+{
+  "count": 46,
+  "checkpoints": [
+    {
+      "path": "pure_wait_cs_random_t/ckpt_0001000.pt",
+      "run_id": "R10k_PW_s7",
+      "update": 1000,
+      "sha256": "bcd81cb2fa52b7492682ba53d9a7d37826589dcab0bb45b5293a12352461a7a6",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "pure_wait_cs_random_t/ckpt_0002000.pt",
+      "run_id": "R10k_PW_s7",
+      "update": 2000,
+      "sha256": "1c30c6a7c423437787176a45a591f1a2697fba869350bcf8d4702e08dfabe6c8",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "pure_wait_cs_random_t/ckpt_0005000.pt",
+      "run_id": "R10k_PW_s7",
+      "update": 5000,
+      "sha256": "b7c914f3f4d4bfc5e015604ed4b7a3cc6fd1401f1ff6f71e28a0a2672773e2c7",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "pure_wait_cs_random_t/ckpt_0010000.pt",
+      "run_id": "R10k_PW_s7",
+      "update": 10000,
+      "sha256": "d655c3be6a5529390af03141f72ac94acb13c6bafefe7e5b86e425b3b02254eb",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "pure_wait_cs_random_t/ckpt_0020000.pt",
+      "run_id": "R10k_PW_s7",
+      "update": 20000,
+      "sha256": "bdd1bd1b285737eb756ffb03362024fd2abd253d4a1bd4218488a355a30613e3",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "pure_wait_cs_random_t/ckpt_0030000.pt",
+      "run_id": "R10k_PW_s7",
+      "update": 30000,
+      "sha256": "a07c56565f7e8167c104ac7f64ecf7e45c42cd5e1bbbd0b5b9ece863c47e6bef",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "pure_wait_cs_random_t/ckpt_0040000.pt",
+      "run_id": "R10k_PW_s7",
+      "update": 40000,
+      "sha256": "b4a6ceac56316db92b9efc22b59536df5f611f4b879de3424a2a06b7ae2dc6a8",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "pure_wait_cs_random_t/ckpt_0050000.pt",
+      "run_id": "R10k_PW_s7",
+      "update": 50000,
+      "sha256": "4c7edff25cdb2da0877178bad0c447460931bde8b914dcb2a6a264585c18b6f9",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "pure_wait_cs_random_t/ckpt_0060000.pt",
+      "run_id": "R10k_PW_s7",
+      "update": 60000,
+      "sha256": "eace02f9b72eec0a5dbe58ea7a96c71db2f4ea04d0ffdb56e863219fa64d39c9",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "pure_wait_cs_random_t/ckpt_0070000.pt",
+      "run_id": "R10k_PW_s7",
+      "update": 70000,
+      "sha256": "4bc4e02c6fe8a79264411fb578f21de563d2079a0f8f6cbe7e1abe4a52999a1a",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "pure_wait_cs_random_t/ckpt_0080000.pt",
+      "run_id": "R10k_PW_s7",
+      "update": 80000,
+      "sha256": "492e306421ea28b94cf900033622633ca226089a8edc087934c248fe16012b1b",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "pure_wait_cs_random_t/ckpt_0090000.pt",
+      "run_id": "R10k_PW_s7",
+      "update": 90000,
+      "sha256": "8698c1108433fe3c38b6d519df122d76ab78a5be0c8e0f78b1b5fbd5af83c7cf",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "pure_wait_cs_random_t/ckpt_0100000.pt",
+      "run_id": "R10k_PW_s7",
+      "update": 100000,
+      "sha256": "728b00c9b1ebd5e4ddd9f714d8834baef48f008f3d66660775eae319183b58c9",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "pure_wait_cs_random_t/ckpt_0110000.pt",
+      "run_id": "R10k_PW_s7",
+      "update": 110000,
+      "sha256": "86802449c3b6ffd8b354eed64bf0be79fa712c834178333218edffb26debbd5b",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "pure_wait_cs_random_t/ckpt_0120000.pt",
+      "run_id": "R10k_PW_s7",
+      "update": 120000,
+      "sha256": "eadc632d68cf9d2e4457934b65f461b416affb35836874fcac74787a646d1ac9",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "pure_wait_cs_random_t/ckpt_0130000.pt",
+      "run_id": "R10k_PW_s7",
+      "update": 130000,
+      "sha256": "440f34d69aba2d802480318a5ab97cdf2550a9bb4b9dfc3f0ae454077abb4925",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "pure_wait_cs_random_t/ckpt_0140000.pt",
+      "run_id": "R10k_PW_s7",
+      "update": 140000,
+      "sha256": "9bc7ba9624294b4b91b133b2f18d87ae93292120ca5aa2913ccf95aa48192b49",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "pure_wait_cs_random_t/ckpt_0150000.pt",
+      "run_id": "R10k_PW_s7",
+      "update": 150000,
+      "sha256": "ab440a762c0db2d6d2bd00fca6e11a6275da8e6852e1636c15b1b60970eba247",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "pure_wait_cs_random_t/ckpt_0160000.pt",
+      "run_id": "R10k_PW_s7",
+      "update": 160000,
+      "sha256": "78c238c7e4be4041be7fcc366e65aa37753d7ffec652411eece3bbbb900738cb",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "pure_wait_cs_random_t/ckpt_0170000.pt",
+      "run_id": "R10k_PW_s7",
+      "update": 170000,
+      "sha256": "743d16c538289bf6cf065f710c248e02d12fab22c5a206c164e57affd08168b7",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "pure_wait_cs_random_t/ckpt_0180000.pt",
+      "run_id": "R10k_PW_s7",
+      "update": 180000,
+      "sha256": "b0ff1c593de460f07f9e33c600e09bb719e84a5904b1da1471849d880df9d660",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "pure_wait_cs_random_t/ckpt_0190000.pt",
+      "run_id": "R10k_PW_s7",
+      "update": 190000,
+      "sha256": "add8fde1083bbebf1970cdd9c9535b47cca3834ec5031a08e59fd844374d2658",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "pure_wait_cs_random_t/ckpt_0200000.pt",
+      "run_id": "R10k_PW_s7",
+      "update": 200000,
+      "sha256": "1f2b047da97b5628695fdb9446f4a75c2f6c1c225e350244866707a1dc482111",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "wait_denoising_fixed_t0/ckpt_0001000.pt",
+      "run_id": "R10k_T0iso_s7",
+      "update": 1000,
+      "sha256": "b04cbfdb2a1e7a54e1c3857ef6d70d87542c02cd8f03baf4e874d8a998915e97",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "wait_denoising_fixed_t0/ckpt_0002000.pt",
+      "run_id": "R10k_T0iso_s7",
+      "update": 2000,
+      "sha256": "09775249c367ccf150c3fd773f92b295743a30dc7db1bc910b9b1f29ba125f68",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "wait_denoising_fixed_t0/ckpt_0005000.pt",
+      "run_id": "R10k_T0iso_s7",
+      "update": 5000,
+      "sha256": "cbb6360b83c05c86ba437a856901801e36acf74776beefffe41db5a319305cd6",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "wait_denoising_fixed_t0/ckpt_0010000.pt",
+      "run_id": "R10k_T0iso_s7",
+      "update": 10000,
+      "sha256": "aefc1a2e8d13155cf68e5cdbe2b56648b666bede651114928563fe53aee388d8",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "wait_denoising_fixed_t0/ckpt_0020000.pt",
+      "run_id": "R10k_T0iso_s7",
+      "update": 20000,
+      "sha256": "107b8652838126379d513366c4e684c2de800ed5582deae281f48b691966e505",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "wait_denoising_fixed_t0/ckpt_0030000.pt",
+      "run_id": "R10k_T0iso_s7",
+      "update": 30000,
+      "sha256": "584e9f1b1fb3f1ce52fa7c95b5a8dc85c8696f4e9aa829c583dc415bd1144f79",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "wait_denoising_fixed_t0/ckpt_0040000.pt",
+      "run_id": "R10k_T0iso_s7",
+      "update": 40000,
+      "sha256": "259bdfe027e50d18508c6e332add29ebca312fce38c21f6fce9bba2be6b86ce5",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "wait_denoising_fixed_t0/ckpt_0050000.pt",
+      "run_id": "R10k_T0iso_s7",
+      "update": 50000,
+      "sha256": "f8e4cd617629f536541f238f8e7df5e3b89f15db47c4c3025b5b7757a58a51fa",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "wait_denoising_fixed_t0/ckpt_0060000.pt",
+      "run_id": "R10k_T0iso_s7",
+      "update": 60000,
+      "sha256": "f6dd0d27bcb48a93839be15dbccbf37b20db15b5bb5144ba8ef5eda581a0176b",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "wait_denoising_fixed_t0/ckpt_0070000.pt",
+      "run_id": "R10k_T0iso_s7",
+      "update": 70000,
+      "sha256": "afe2b3f44168be01e54532886d2d1439bcb167f0f437c826c45e14668ce1caa0",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "wait_denoising_fixed_t0/ckpt_0080000.pt",
+      "run_id": "R10k_T0iso_s7",
+      "update": 80000,
+      "sha256": "2800b11b460d0e5de02d7009c25a0e4d93b555c7c63f10a59e89739383043118",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "wait_denoising_fixed_t0/ckpt_0090000.pt",
+      "run_id": "R10k_T0iso_s7",
+      "update": 90000,
+      "sha256": "61ba4ff1b03eed8710610004e9b1e68db3886813e2fbd68c83d58da444a9dafe",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "wait_denoising_fixed_t0/ckpt_0100000.pt",
+      "run_id": "R10k_T0iso_s7",
+      "update": 100000,
+      "sha256": "e600093705a1477f6a092135438a4346dd1a4bc5c97430f2d179fdbb7b73a8d7",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "wait_denoising_fixed_t0/ckpt_0110000.pt",
+      "run_id": "R10k_T0iso_s7",
+      "update": 110000,
+      "sha256": "1791c9e8a4be8f84d2fdd713b598439dd7ca3332ae0b9efdc20c53dbd86c07e6",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "wait_denoising_fixed_t0/ckpt_0120000.pt",
+      "run_id": "R10k_T0iso_s7",
+      "update": 120000,
+      "sha256": "c6767cfb80c10f38ca4bac059d5240211929dbe5a5cadb4b48ebb07a196566bf",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "wait_denoising_fixed_t0/ckpt_0130000.pt",
+      "run_id": "R10k_T0iso_s7",
+      "update": 130000,
+      "sha256": "1bf915b7a6cb858a51b534b9afd289421e0fdbb3c6f534831f0c4254143fb890",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "wait_denoising_fixed_t0/ckpt_0140000.pt",
+      "run_id": "R10k_T0iso_s7",
+      "update": 140000,
+      "sha256": "a00e672829198b3cba32d4241aa914b067cbd82f4f46f1a1219075f23ce22f8a",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "wait_denoising_fixed_t0/ckpt_0150000.pt",
+      "run_id": "R10k_T0iso_s7",
+      "update": 150000,
+      "sha256": "c7b347a167d12485afddc0943dc92ce8326e3eb5a099ce5e4aafb4c2c0b81d8c",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "wait_denoising_fixed_t0/ckpt_0160000.pt",
+      "run_id": "R10k_T0iso_s7",
+      "update": 160000,
+      "sha256": "bbd8ee69c630dd68e2ae25cb2eacb4bd1d3ffcadadd37a7bbc9938cc791ecb3a",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "wait_denoising_fixed_t0/ckpt_0170000.pt",
+      "run_id": "R10k_T0iso_s7",
+      "update": 170000,
+      "sha256": "76a4c1cd41d687ef06a1ccf6b67b2c56399c1ea2bf1103c51c144b3eddb6e20f",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "wait_denoising_fixed_t0/ckpt_0180000.pt",
+      "run_id": "R10k_T0iso_s7",
+      "update": 180000,
+      "sha256": "d767ae2ced030111845d93e7f35587d7ecfc7d8ce43622d47df0d9c93638b191",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "wait_denoising_fixed_t0/ckpt_0190000.pt",
+      "run_id": "R10k_T0iso_s7",
+      "update": 190000,
+      "sha256": "a44ac51a1decb6919d48d1d02fcef7f4e525759760b913fffa12e064754d2c29",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    },
+    {
+      "path": "wait_denoising_fixed_t0/ckpt_0200000.pt",
+      "run_id": "R10k_T0iso_s7",
+      "update": 200000,
+      "sha256": "6be182b9f912f741dfb0ba491879812e0813cc8adb264c76fa0cf5a9bbc64e22",
+      "bytes": 115042564,
+      "rope_base": 10000,
+      "seed": 7,
+      "max_pos": 4097,
+      "kind": "evaluation_weights"
+    }
+  ]
+}