File size: 2,010 Bytes
0b28a2c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
{
  "anti_leakage": {
    "eval_hidden_files": [
      "data/exact_benchmark_hidden.jsonl",
      "data/frontier_eval_hidden.jsonl",
      "fixtures/hidden"
    ],
    "never_mix_files": [
      "data/public_tasks.jsonl",
      "data/hidden_grader_records.jsonl"
    ],
    "recommended_storage": "separate access-controlled object stores and credentials for policy workers and reward workers"
  },
  "public_release_scope": {
    "all_included_prompts_and_fixtures_are_public": true,
    "omitted_hidden_paths_in_anti_leakage_policy": "production-policy documentation only; those assets are not distributed",
    "split_labels": "organizational only; they do not imply secrecy or holdout status"
  },
  "schema_version": "3.0",
  "stages": [
    {
      "episode_filter": {
        "episode_type": "exact_benchmark",
        "split": "train"
      },
      "name": "exact mathematical tools",
      "objective": "Learn exact computation, evidence formatting, and artifact discipline before open research.",
      "reward_mode": "exact_machine",
      "stage": 1
    },
    {
      "episode_filter": {
        "episode_type": "exact_benchmark",
        "split": "dev"
      },
      "name": "public exact validation",
      "objective": "Validate on the public dev split; this showcase is not held out.",
      "reward_mode": "exact_machine",
      "stage": 2
    },
    {
      "episode_filter": {
        "episode_type": "targeted_frontier_milestone"
      },
      "name": "targeted frontier milestones",
      "objective": "Train decomposition, gap control, adversarial checks, and tool use.",
      "reward_mode": "hybrid_expert_gated",
      "stage": 3
    },
    {
      "episode_filter": {
        "episode_type": "full_frontier_task",
        "split": "eval"
      },
      "name": "full frontier evaluation",
      "objective": "Evaluate long-horizon research behavior without treating unresolved conjectures as binary RLVR.",
      "reward_mode": "hybrid_expert_gated",
      "stage": 4
    }
  ]
}