teleo-infrastructure/docs/reports/leo-oos-reasoning-benchmark-20260715/evaluated-protocol-manifest.json

17 lines
1 KiB
JSON

{
"schema": "livingip.leoEvaluatedOosProtocolManifest.v1",
"protocol_schema": "livingip.leoM3taversalOosProtocol.v2",
"protocol_id": "leo-m3taversal-oos-8c0fc263b5cd95b8",
"protocol_hash_sha256": "d68330b20d31677b231beb8450d5fdfd40efebf498cfc2f8ae2e63efd860bf43",
"protocol_file_sha256": "fac462ee83b1035215cca12262e27b70c70e8a8ed0a216aa70c69ef46cd6a185",
"harness_git_head": "93eddbcd4f27bd9e349f217767da2c9bf2b7baeb",
"trial_count": 3,
"prompt_count": 30,
"family_count": 10,
"unique_prompt_id_count": 30,
"unique_message_hash_count": 30,
"validation_pass": true,
"prompt_bodies_committed": false,
"retention_contract": "The exact post-evaluation protocol is retained as a private mode-0600 artifact and is bound by protocol_file_sha256. This manifest is not sufficient to recreate prompt bodies by itself.",
"claim_ceiling": "This manifest identifies the exact evaluated protocol without publishing the formerly blinded prompts. It does not independently authenticate live VPS execution."
}