File size: 4,671 Bytes
700c7f4 | 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 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 | {
"id": "software-delivery",
"version": 1,
"basis": "The configurations behind the published report (docs/REPRODUCTION.md): every setting is read from the recorded episodes of the named console jobs in the episode table, not inferred from a sheet. Prompt, tool schema, profile budgets and scoring policy are the exact versioned artefacts those episodes recorded; model parameters are what the adapters send and are listed as recorded for comparison.",
"selection_evidence": {
"executable": {
"jobs": [
"rlj-20260911-182105-d16163",
"rlj-20260925-204349-0090bb",
"rlj-20261013-232626-6fc6bd"
],
"models": {
"gpt-6-astra": "rlj-20260911-182105-d16163",
"claude-fable-5-1": "rlj-20260911-182105-d16163",
"gemini-3.8-flash": "rlj-20260911-182105-d16163",
"claude-opus-5-5": "rlj-20260925-204349-0090bb",
"gemini-4-argon": "rlj-20261013-232626-6fc6bd"
}
},
"qna": {
"jobs": [
"rlj-20261004-175352-1f4a04",
"rlj-20261005-005743-894c05",
"rlj-20261014-113935-1586f8"
],
"models": {
"gpt-6-astra": "rlj-20261004-175352-1f4a04",
"claude-fable-5-1": "rlj-20261004-175352-1f4a04",
"gemini-3.8-flash": "rlj-20261004-175352-1f4a04",
"claude-opus-5-5": "rlj-20261005-005743-894c05",
"gemini-4-argon": "rlj-20261014-113935-1586f8"
}
}
},
"settings_evidence": "settings-evidence/software-delivery.json",
"families": {
"executable": {
"kind": "executable",
"prompt": "agent_prompt_v0.3",
"profile": "python_standard_2m_v0.1",
"scoring": "scoring_v0.2",
"pass_threshold": 0.95,
"sandbox_image": "python-standard",
"evaluator_image": "pdl-bench/swe:1.0.0",
"note": "Manifests pin python_standard_1m_v0.1; the published report ran the 2M profile as an operator override."
},
"qna": {
"kind": "qna",
"prompt": "agent_prompt_qa_v0.1",
"profile": "python_qa_2m_v0.1",
"scoring": "scoring_llm_v0.3",
"pass_threshold": 0.95,
"sandbox_image": "python-qa",
"evaluator_image": "poindexter-samples/sweqna:20261118",
"judges": [
"google:gemini-3.1-pro-preview",
"anthropic:claude-opus-5",
"openai:gpt-5.6-sol@https://<azure-resource>/openai/v1"
],
"note": "Manifests pin scoring_llm_v0.2; the published report graded under scoring_llm_v0.3 as an operator override. The judge panel is the reference panel; a researcher may name another panel with --judges, and the record says which was used."
}
},
"models": [
{
"id": "gpt-6-astra",
"provider": "openai",
"endpoint": "azure",
"model": "gpt-6-astra",
"attempts": 8,
"route_label": "openai-azure:swedencentral",
"recorded_model_parameters": {
"api_mode": "responses",
"sampling": "provider-default"
}
},
{
"id": "claude-fable-5-1",
"provider": "anthropic",
"endpoint": null,
"model": "claude-fable-5-1",
"attempts": 8,
"route_label": "anthropic:api",
"recorded_model_parameters": {
"max_tokens": 8192,
"prompt_cache": "ephemeral"
}
},
{
"id": "claude-opus-5-5",
"provider": "anthropic",
"endpoint": null,
"model": "claude-opus-5-5",
"attempts": 8,
"route_label": "anthropic:api",
"recorded_model_parameters": {
"max_tokens": 8192,
"prompt_cache": "ephemeral"
}
},
{
"id": "gemini-3.8-flash",
"provider": "gemini",
"endpoint": "vertex:global",
"model": "gemini-3.8-flash",
"attempts": 8,
"route_label": "google-vertex:global",
"recorded_model_parameters": {
"temperature": 1
}
},
{
"id": "gemini-4-argon",
"provider": "gemini",
"endpoint": "vertex:global",
"model": "gemini-4-argon",
"attempts": 8,
"route_label": "google-vertex:global",
"recorded_model_parameters": {
"temperature": 1
}
}
],
"baselines": [
{
"id": "gold",
"provider": "gold",
"attempts": 1,
"note": "reference patch via write_patch; expected reward 1.0"
},
{
"id": "noop",
"provider": "noop",
"attempts": 1,
"note": "submits at once; expected reward 0.0 under scoring_v0.2"
}
],
"evidence_note": "Earlier episodes under scoring_llm_v0.2 (QnA) or scoring_v0.1 with a different image digest (executable, August probes) appear in settings-evidence as evidence only and are not pooled with the published numbers."
}
|