{
  "schema": "world.workshop-evidence/v1",
  "as_of": "30 August 2026",
  "purpose": "Curated snapshot of preserved experiment outcomes; not runtime telemetry or automatic promotion.",
  "engine_contract": {
    "commit": "9b3fab5fb5e85991b642799cba3d37b32ad29253",
    "abi": "r3+machine-v1",
    "capabilities": "0xe07ff",
    "live_world_uses_cuda": false
  },
  "experiments": [
    {
      "id": "real-cohort-direct-r3",
      "title": "Camera-fixed real-agent trial · r3",
      "outcome": "rejected",
      "scope": "authoring workflow; no accepted artifact",
      "result": "Two equal-start agents made 36 model calls in 113 seconds ($1.99743525 usage-priced estimate). Both authored inspection cameras, but their first plans attached unsupported anchor lists to overlap criteria. A corrected plan arrived after the budget stop and was never executed. Both agents were removed; the original 42-object CPU station and earlier empty draft remain intact.",
      "next": "Query and criterion reference examples now pass local parser/server tests; the five bad saved inputs still fail and two corrected plans retain identical bytes. No new draft or invention was accepted. $1.08237225 remains from the original $5 envelope; no next trial is authorized.",
      "source_sha256": "fc4ae83e2492fc5628bcc2d6d43d65211ae0cf73bda2b307a2351a9ab392265e"
    },
    {
      "id": "standalone-camera-authoring-r1",
      "title": "Agents can define their inspection views",
      "outcome": "passed",
      "scope": "scripted authoring-to-render integration; not agent invention",
      "result": "The actual agent parser now carries explicit cameras through checking, draft submission, part batches and recovery. A fresh scripted portal passed the real server's geometry and front-view checks and produced two different 800×500 canonical images. Bad camera definitions and premature publication were rejected. No model or CUDA calls were made.",
      "next": "The camera-fixed r3 trial used authored cameras but ended without an accepted artifact. This scripted local fixture remains infrastructure evidence, not agent invention.",
      "source_sha256": "c8ec9e56164e661f270bb05c3cbb09369f19c528a0141407b3c7aa0d67d93cb3"
    },
    {
      "id": "real-cohort-direct-r2",
      "title": "Real-agent build attempt · r2",
      "outcome": "rejected",
      "scope": "authoring workflow; no accepted invention",
      "result": "Two equal-start agents made 36 model calls in 132 seconds ($1.9201925 usage-priced estimate). Both were stopped and removed when one reached 20 calls. The original 42-object CPU station remains; one empty draft is preserved. A local check passed a script, but the outgoing draft dropped its program and protected IDs. No work was accepted.",
      "next": "Script serialization and standalone camera authoring now have local regression and real-render proofs. A new bounded decision is still required before another paid attempt; no invention or communication advantage is claimed.",
      "source_sha256": "fcbb0d02375dfc63395689162fd83889d7c89b5e7f82f69ecb385672e24bfff4"
    },
    {
      "id": "local-canonical-camera-r2",
      "title": "Local agent vision",
      "outcome": "passed",
      "scope": "scripted visual-review infrastructure; not agent invention",
      "result": "The real canonical camera rendered a staged three-part portal, measured its completeness, connection and support, and allowed normal revision acceptance only after both required views existed. A forbidden asset request was blocked; worker failure required an explicit restart. Rendering uses local SwiftShader, not CUDA physics.",
      "next": "The bounded r2 agents used real local vision but accepted no artifact. A working camera and scripted portal prove infrastructure, not autonomous engineering.",
      "source_sha256": "e9f57b077391f64826c01c171f9c75cfa8d534b921d667ab2f7971680c17a4dc"
    },
    {
      "id": "real-cohort-direct-r1",
      "title": "Bounded real cohort · r1",
      "outcome": "rejected",
      "scope": "stopped before agent startup",
      "result": "The admission process exceeded its startup deadline before either agent or model call began: $0 model spend, no committed cohort, and no new artifact. The original 42-object station is intact. Cold macOS executable startup and invalid snapshot URI handling were diagnosed; corrected backups now match the live database content.",
      "next": "A fresh attempt needs a new decision and identity, with startup probing and valid snapshots. The two-agent, $5 total, 15-minute limits remain unchanged; no automatic retry.",
      "source_sha256": "4ab3aa9322a008e20097c376ec928d3f12c9b142b51e78c4034d3606ffc55099"
    },
    {
      "id": "daytona-world-humanoid-balance-20260830-r5",
      "title": "Humanoid balance · r5",
      "outcome": "rejected",
      "scope": "walking qualification",
      "result": "250 CUDA PPO updates completed. Earliest strict failure improved from 1.02 to 2.38 seconds; median first-episode survival was 2.72 seconds. Neither foot reached 4 cm swing clearance. All evaluated checkpoint episodes failed before 10 seconds.",
      "next": "Learn sustained balance and actual stepping. The viewer intentionally retains the older diagnostic pilot; r5 has not replaced it.",
      "source_sha256": "47e53b3a41cf57959ea7629d58517803285c8a25ba13f2fd6d8e5e8b3935c775"
    },
    {
      "id": "daytona-world-inline-four-thermal-64step-20260830-r1",
      "title": "Inline-four & pump",
      "outcome": "rejected",
      "scope": "combined machine",
      "result": "Thermal cadence and temperature rise were observed, but joint-anchor error reached 0.188 m against a 0.01 m limit after 64 control steps. This run used CPU engine/fluid/thermal models coupled to CUDA rigid bodies.",
      "next": "Isolate closed-chain load conditioning before a longer integrated run. Isolated field-kernel parity is not full GPU multiphysics acceptance.",
      "source_sha256": "306b6544b2d4e7d3b16c80ba59ec27c22fb5ef648dc0cfdf8a2e8b139727091c"
    },
    {
      "id": "agent-society-foundry-ablation-r1",
      "title": "Agent invention cohort",
      "outcome": "rejected",
      "scope": "tooling lifecycle",
      "result": "Environment-only and direct-communication agents attempted pump instrumentation. No artifact passed acceptance. The ledger and rejected revisions are preserved; successful autonomous invention is not yet demonstrated.",
      "next": "Complete the inspect → revise → test → publish loop with a small cohort before scaling the society.",
      "source_sha256": "72862280d422978cc658ad04c88a9d0de9cae4a44fb6e471f887899f86a7fba1"
    },
    {
      "id": "live-cell-recovery-r3",
      "title": "Live-cell recovery · CPU",
      "outcome": "passed",
      "scope": "local subprocess and SQLite protocol",
      "result": "120 committed steps with ground contact. A killed worker's uncommitted step was discarded; a new generation restored full state and contact caches, rejected the late reply, and reproduced the next step bit-exactly.",
      "next": "Run the prepared native GPU recovery gate, then verify the bounded supervisor on a real sealed cell. This record remains CPU evidence only.",
      "source_sha256": "a2fe2d7e404f8ebf3a00803c8e63950b081bf473a7997987ec5b1a7e357ccf63"
    },
    {
      "id": "sealed-live-cell-local-r1",
      "title": "Sealed cell ownership",
      "outcome": "passed",
      "scope": "local integration tests; not GPU execution",
      "result": "493 Rust tests pass. Cell-owned objects reject competing CPU, script and editor writes; visible poses and checkpoints commit in one transaction. Browser batches are fenced and sealed regions replace native bodies in CPU prediction.",
      "next": "Execute and independently review the native E1 recovery gate. No production cell is attached; worker liveness, mixed-body coupling and walking are not claimed.",
      "source_sha256": "1f18e400e5d89772f2f01c527056c7dfce83a0ae144b7e1aecb155fd2411b064"
    },
    {
      "id": "bounded-cell-supervisor-local-r1",
      "title": "Bounded cell runner",
      "outcome": "passed",
      "scope": "local orchestration tests; not GPU execution",
      "result": "509 Rust tests pass, including 11 supervisor cases. The runner persists intent before dispatch, publishes only committed replies, and retains the last pose on timeout or failure. Uncertain cleanup blocks a second driver.",
      "next": "Two native E1 attempts ended before execution: spot capacity rejection, then on-demand startup timeout with cleanup verified. Native acceptance and production attachment remain pending.",
      "source_sha256": "4d5b18e307f4257f0736c7a17838c2f925a1847e82436934d4744fa079d165f5"
    },
    {
      "id": "live-domain-participation-r1",
      "title": "Construction-safe pump · CPU",
      "outcome": "passed",
      "scope": "local CPU calibration; not CUDA or agent invention",
      "result": "60 simulated seconds of loaded pumping, bypass and cutoff pass mass/energy gates. Adding 200 display objects leaves 216,000 physical scalars bit-exact. The 42-piece live station has field-driven gauges and corrected cylinder timing. 520 Rust tests pass.",
      "next": "Typed connections and local canonical rendering now work. Complete the real-agent improvement loop. The live station uses prescribed shaft motion, not a validated rigid crankshaft.",
      "source_sha256": "6c908f763fe7fbb1a9a95daf06b484250652e765c1214690827e3d7dfab4c292"
    },
    {
      "id": "stable-physics-ports-r1",
      "title": "Persistent machine connections",
      "outcome": "passed",
      "scope": "local authoring and CPU slice lowering; not CUDA execution",
      "result": "Pump ports retain integer object identities through scene edits and project persistence. A deleted endpoint faults instead of reconnecting to a different object; published IDs are not recycled after restart. Live slice export correctly maps all four pump/bypass connections and rejects a missing reservoir. 536 Rust tests pass.",
      "next": "Local canonical rendering is now available. Complete the agent build/inspect/test/publish workflow; connection readiness alone is not autonomous invention or an integrated CUDA machine.",
      "source_sha256": "df4e8717eb30e0b580c7a46e4fc943daed7693bb70b3d6afe4ce68188e978f83"
    }
  ]
}