{"checkpoint":{"checkpoint_hash":"bb31c0922de639cb9220c4cc5fc35d79aec719eb6fcedb09159bdff8cfb8fd43","config_hash":"27f2c77c9d47a7d053c01ab65f8d43aad79463b27d882f2d85ec28bc062cb2b2","repo_id":"abdelstark/lensemble-phase3-consortium-checkpoint","repo_type":"model","revision":"828e210cba4870b2be4ab573a5f0dd4ee30bae29"},"claim_boundary":"Historical held-out SO-100 latent evidence only: these final-round val_pred/effective_rank values predate the outer-update direction correction and public-probe-v2 target-binding contract. They are retained for audit history and do not validate the corrected runtime; GitHub issue #335 tracks the replacement run. effective_rank is scale-invariant and blind to the historical magnitude collapse diagnostic (~7.5e-6 latent variance). The associated central ceiling diagnostic did not establish downstream usefulness. Closed-loop physical task-success is DEFERRED, not claimed: it requires the unvendored stable-worldmodel suite (#96), and it requires a genuinely useful non-collapsed checkpoint (#244). This report is a correction of the prior SO-100 overclaim, not a success story: skill_vs_identity is gameable, effective_rank is scale-invariant, and latent-MPC success on these checkpoints would be uninformative. This is engineering and training evidence, NOT a cryptographic proof of honest participant computation, and does not claim paper-scale LeWorldModel performance or SO-100 robotics task success.","evidence_status":"historical_pre_correctness_fix","generated_at":"2026-07-24T11:17:11.890600Z","held_out_data_ref":"hf://datasets/abdelstark/lensemble-phase3-so100-silos/phase3-so100-silo4.h5","held_out_latent_metrics":{"effective_rank":35.79561996459961,"held_out_windows":1216,"latent_dim":256,"measured_on":"hf://datasets/abdelstark/lensemble-phase3-so100-silos/phase3-so100-silo4.h5","note":"Historical SO-100 latent metrics: the archived run records final-round effective_rank and val_pred on the disjoint held-out split phase3-so100-silo4.h5 (#242). The values predate the public-probe-v2 target-binding contract and do not validate the corrected runtime. effective_rank is scale-invariant and blind to the historical magnitude collapse diagnostic (~7.5e-6 latent variance); the associated central ceiling diagnostic did not establish downstream usefulness.","round_index":9,"val_pred":32476.08585357666,"window_steps":4},"planner_budget":{"action_dim":6,"eval_episodes":20,"executed":false,"horizon":16,"note":"Non-toy latent-MPC budget a closed-loop run WOULD use over the SO-100 6-DoF action space; no planner is executed here because closed-loop task-success is blocked (see task_success blockers #96 and #244).","planner":"icem","planner_iterations":8,"planning_samples":512},"raw_data_in_report":false,"schema_version":2,"source_report_uri":"docs/evidence/phase3_consortium_run_report.json","superseded_reason":"These SO-100 downstream values predate the correction of the outer-update direction and the public-probe-v2 target-binding contract. They are retained for audit history only and do not validate the corrected runtime; GitHub issue #335 tracks the replacement run.","task_env_id":"so100-heldout://phase3-public-task-scale","task_success":{"blockers":[{"blocker_ref":"#96","reason":"Closed-loop physical task-success is blocked by the unvendored stable-worldmodel suite (#96, maintainer-gated): lensemble/eval/world.py::resolve_env raises EvaluationError for any stable-worldmodel:// id because the suite is not vendored. A recorded held-out dataset is open-loop (you cannot apply arbitrary planner actions to recorded frames), so a faithful physical task-success score requires the simulator."},{"blocker_ref":"#244","reason":"Latent-MPC planning success would be uninformative on the published federated checkpoints because they do not show downstream usefulness (#244): the held-out representation has magnitude collapse (~7.5e-6 latent variance), effective_rank is scale-invariant and blind to that collapse, and the associated central ceiling diagnostic did not establish that this checkpoint is a usable world model despite the recorded val_pred/effective_rank scalars."}],"status":"blocked","success_rate":null}}