{
  "schema_version": 1,
  "purpose": "Verify preserved response excerpts and reconstructed movement results; not certify model safety or authenticity of provider logs.",
  "source_records": {
    "astra": {
      "file": "rollout-2026-09-20T06-59-52-01a0beaf-e057-7361-b2fa-4d1a693d013d.jsonl",
      "sha256": "48e9d3a188f616f7ca89323517cb993cbd4551ecf26df2318f9704fb99c5d6bf",
      "selection": "Three user instruction batches and corresponding assistant final responses; response_item message records only; source line numbers in exchanges.json."
    },
    "glm": {
      "database": "opencode.db",
      "selection": "Text parts for four instruction batches in tidy-sailor and two in mighty-squid, paired to following assistant text; original row/session/message IDs in exchanges.json. Database and WAL not distributed.",
      "sessions": {
        "tidy-sailor": "ses_f4171f17fffeFAlPafSgz95rij",
        "mighty-squid": "ses_f413ebcf4ffeTeVMoruxLRIzGy"
      }
    }
  },
  "omissions": [
    "Full harness/system/developer instructions, private unrelated chat, tool calls and reasoning are not included.",
    "Provider-side hidden instructions and generation settings are not fully available.",
    "Original per-step model execution traces were not located.",
    "No complete census of failed provider requests; the abandoned Zen runner output was not found.",
    "No new controlled model trials or independent replication have been performed."
  ],
  "interpretation_limits": [
    "Nine response batches contain 45 instruction-response blocks, not 45 independent model calls.",
    "Astra batches share one conversation; mighty-squid second batch follows its removal batch.",
    "Two old-context removal batches have different prompts; fresh-context comparison also changes prompt wording.",
    "Astra harm-related refusals explicitly cite unavailable capabilities. REFUSE must not automatically be classified as a moral safety response.",
    "The quiet-risky instructions were generated by the same GLM session before being tested; context-excerpts.json preserves that exchange.",
    "The exact simulator cube position differs slightly from rounded coordinates supplied in prompts.",
    "Physical safety cannot be established in an indestructible-object toy simulator."
  ],
  "upstream": {
    "package": "inspect-robots",
    "version": "0.58.0",
    "module_sha256": "7eca6b141b86c5a45eadd6be71c96616c781691d1d46716f939e7e11e861ac3e",
    "method": "Compared every reconstructed step to installed CubePickEmbodiment.step; extended stepping is deliberate diagnostic replay. No provider calls.",
    "compared_steps": 329,
    "max_allowed_absolute_difference": 1e-12,
    "result": "passed",
    "independent_replication": false
  },
  "sha256": {
    "README.md": "c4d578843ab71e499952e8a576d09375565e3f9bd07c7965a2f1a92d4c3051b1",
    "verify.py": "8d45ab9bd836833620780cea638f06aae950ace8674d14b3efe4f7ccb66213f4",
    "check_upstream.py": "1bdca334d4c05e10b76a37a7811a7b954936b046b7e416a3ffdfde750d1351c7",
    "exchanges.json": "6c40f9f4f49c5ee7ad1a6304bf1ea112849ce4bc9c725785001275750822b4db",
    "context-excerpts.json": "9757a9af0701fe485dd6ca821ee19b78a7609699eb28da5764ee08e8677222bb",
    "scene.json": "944bf6b7e30e4357352205ef069fafeb1ba38689b1c2e86037a118e4b49e15c4",
    "results.json": "81928b214f4c549a6541eec804517cbae479bc7e9a34fc5c7a6a50e75aaade02",
    "report.html": "a9e9712cc8aac13935ad4d0622ec72ca1357bc9da69e4cb520b14f67fa1a632c",
    "upstream-check.json": "c7a62b94adbb8e30c58e56d3a591ab282430f49c2885f842fedc1374075fbb41",
    "original-scene.png": "2d400460350f30af43f3a5b6441b122e8b5d1c46792285c1efd8d32722f6e30b",
    "paper.pdf": "f1baec6d38039e9c0d112d214050932877849db5e9205318052bfa6b97948980",
    "paper-source.txt": "ee6de827e9f8deaf0e8865eda44dc751d9841af52f35f2c5790fe59ba096b455",
    "cubepick-wording-paper.pdf": "f1baec6d38039e9c0d112d214050932877849db5e9205318052bfa6b97948980",
    "START_HERE.txt": "2fcbfc044e2a10390790381ee63487b6cca2023187e67e6c01680dfe7448d68f"
  }
}