{
  "index_schema_version": "mtlm-index/1",
  "generated_at": 1792742911,
  "artifacts_base": "https://github.com/javimosch/mtlm-heads/raw/master/lanes",
  "lanes": [
    {
      "name": "banking77",
      "version": "0.1.0",
      "file": "banking77-0.1.0.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/banking77-0.1.0.lanepack.tar.gz",
      "sha256": "2a1b5ad93558a3b5c0c13b334b899912a880cca3966613518c121db195adcb8b",
      "head_sha256": "603018f7d76f01fa7818b85a1dfb0bc7deaa341b9b9d6e39986c144e89c0eb22",
      "classes": 51,
      "taxonomy": [
        "Refund_not_showing_up",
        "activate_my_card",
        "apple_pay_or_google_pay",
        "automatic_top_up",
        "balance_not_updated_after_bank_transfer",
        "balance_not_updated_after_cheque_or_cash_deposit",
        "beneficiary_not_allowed",
        "cancel_transfer",
        "card_about_to_expire",
        "card_arrival",
        "card_linking",
        "card_payment_fee_charged",
        "card_payment_not_recognised",
        "card_payment_wrong_exchange_rate",
        "cash_withdrawal_charge",
        "cash_withdrawal_not_recognised",
        "change_pin",
        "country_support",
        "declined_card_payment",
        "declined_cash_withdrawal",
        "declined_transfer",
        "direct_debit_payment_not_recognised",
        "disposable_card_limits",
        "edit_personal_details",
        "exchange_charge",
        "exchange_via_app",
        "extra_charge_on_statement",
        "failed_transfer",
        "fiat_currency_support",
        "getting_spare_card",
        "lost_or_stolen_phone",
        "oos",
        "order_physical_card",
        "pending_card_payment",
        "pending_cash_withdrawal",
        "pending_top_up",
        "pending_transfer",
        "request_refund",
        "reverted_card_payment?",
        "supported_cards_and_currencies",
        "top_up_failed",
        "top_up_reverted",
        "transaction_charged_twice",
        "transfer_fee_charged",
        "transfer_not_received_by_recipient",
        "transfer_timing",
        "verify_top_up",
        "visa_or_mastercard",
        "why_verify_identity",
        "wrong_amount_of_cash_received",
        "wrong_exchange_rate_for_cash_withdrawal"
      ],
      "status": "certified",
      "holdout_acc": 0.733,
      "n_holdout": 2000,
      "oos": null,
      "live_verified": true,
      "operating": {
        "min_conf": 0.7,
        "note": "leak 1.6%/coverage 58% at 0.7; id-oos trained abstain",
        "qhat": 0.938,
        "qhat_note": "conformal \u03b1=0.10 on 1398-row cal split (test+id-oos mix): 46.4% auto, 9.1% err among auto, 88.3% set coverage \u2014 replaces the 0.7 floor where wired",
        "mondrian": "per-class qhat: auto collapses to 0.6% at \u03b1=0.05 \u2014 marginal auto rate is subsidized by easy classes; expect candidates[] on most traffic",
        "limitations": [
          "bank50 subset: only 50 of banking77's 77 classes \u2014 queries for absent classes (e.g. lost_or_stolen_card) correctly return oos",
          "idoos training set contains ~734 'card'-mentioning rows labeled oos that overlap real card classes (e.g. card-arrival phrasings) \u2014 contaminates card-* classes; retrain-with-relabeling would raise card-class recall",
          "clarification-first lane: per-class conformal automation near zero; candidates[] is the terminal UX"
        ]
      },
      "yanked": null,
      "superseded_by": null,
      "yank_reason": null,
      "conformal": null,
      "license": "Apache-2.0",
      "pricing": {
        "tier": "free"
      },
      "provenance": "BANKING77-OOS (jianguoz/Few-Shot-Intent-Detection, PolyAI-LDN/task-specific-datasets): 5905 train / 2062 id-oos train; m7router3s384 frozen trunk; emit tools/emit_banking77.py",
      "eval_sample": null,
      "runtime": "anvil-serve >= v0.3.0 (ANVIL_EXPERTS=<lane>:<head>, ANVIL_EXPERT_QHAT if provided)",
      "published_at": 1791756725
    },
    {
      "name": "banking77",
      "version": "0.1.1",
      "file": "banking77-0.1.1.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/banking77-0.1.1.lanepack.tar.gz",
      "sha256": "1d84e6fafa525e66a380e27e9b15d0cc067ac58b7fd0e47eb854c67c951d0ee2",
      "head_sha256": "603018f7d76f01fa7818b85a1dfb0bc7deaa341b9b9d6e39986c144e89c0eb22",
      "classes": 51,
      "taxonomy": [
        "Refund_not_showing_up",
        "activate_my_card",
        "apple_pay_or_google_pay",
        "automatic_top_up",
        "balance_not_updated_after_bank_transfer",
        "balance_not_updated_after_cheque_or_cash_deposit",
        "beneficiary_not_allowed",
        "cancel_transfer",
        "card_about_to_expire",
        "card_arrival",
        "card_linking",
        "card_payment_fee_charged",
        "card_payment_not_recognised",
        "card_payment_wrong_exchange_rate",
        "cash_withdrawal_charge",
        "cash_withdrawal_not_recognised",
        "change_pin",
        "country_support",
        "declined_card_payment",
        "declined_cash_withdrawal",
        "declined_transfer",
        "direct_debit_payment_not_recognised",
        "disposable_card_limits",
        "edit_personal_details",
        "exchange_charge",
        "exchange_via_app",
        "extra_charge_on_statement",
        "failed_transfer",
        "fiat_currency_support",
        "getting_spare_card",
        "lost_or_stolen_phone",
        "oos",
        "order_physical_card",
        "pending_card_payment",
        "pending_cash_withdrawal",
        "pending_top_up",
        "pending_transfer",
        "request_refund",
        "reverted_card_payment?",
        "supported_cards_and_currencies",
        "top_up_failed",
        "top_up_reverted",
        "transaction_charged_twice",
        "transfer_fee_charged",
        "transfer_not_received_by_recipient",
        "transfer_timing",
        "verify_top_up",
        "visa_or_mastercard",
        "why_verify_identity",
        "wrong_amount_of_cash_received",
        "wrong_exchange_rate_for_cash_withdrawal"
      ],
      "status": "certified",
      "holdout_acc": 0.733,
      "n_holdout": 2000,
      "oos": null,
      "live_verified": true,
      "operating": {
        "min_conf": 0.7,
        "note": "leak 1.6%/coverage 58% at 0.7; id-oos trained abstain",
        "qhat": 0.938,
        "qhat_note": "conformal \u03b1=0.10 on 1398-row cal split (test+id-oos mix): 46.4% auto, 9.1% err among auto, 88.3% set coverage \u2014 replaces the 0.7 floor where wired",
        "mondrian": "per-class qhat: auto collapses to 0.6% at \u03b1=0.05 \u2014 marginal auto rate is subsidized by easy classes; expect candidates[] on most traffic",
        "limitations": [
          "bank50 subset: only 50 of banking77's 77 classes \u2014 queries for absent classes (e.g. lost_or_stolen_card) correctly return oos",
          "idoos training set contains ~734 'card'-mentioning rows labeled oos that overlap real card classes (e.g. card-arrival phrasings) \u2014 contaminates card-* classes; retrain-with-relabeling would raise card-class recall",
          "clarification-first lane: per-class conformal automation near zero; candidates[] is the terminal UX"
        ]
      },
      "yanked": null,
      "superseded_by": null,
      "yank_reason": null,
      "conformal": null,
      "license": "Apache-2.0",
      "pricing": {
        "tier": "free"
      },
      "provenance": "BANKING77-OOS (jianguoz/Few-Shot-Intent-Detection, PolyAI-LDN/task-specific-datasets): 5905 train / 2062 id-oos train; m7router3s384 frozen trunk; emit tools/emit_banking77.py",
      "eval_sample": "eval_sample.jsonl",
      "runtime": "anvil-serve >= v0.3.0 (ANVIL_EXPERTS=<lane>:<head>, ANVIL_EXPERT_QHAT if provided)",
      "published_at": 1791756725
    },
    {
      "name": "civic",
      "version": "0.1.0",
      "file": "civic-0.1.0.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/civic-0.1.0.lanepack.tar.gz",
      "sha256": "1907a4c8731cb5e2b6bb825a7781999f1295fb43b79b60bc445c796fa2ec95a5",
      "head_sha256": "8a6393b45db910987ae8b8ccc789e13be427157c27536e7eb21e0780bae8269a",
      "classes": 13,
      "taxonomy": [
        "animals",
        "drains_flooding",
        "fly_tipping",
        "graffiti_flyposting",
        "other",
        "parking_vehicles",
        "parks_green",
        "pavement_footpath",
        "pothole",
        "signs_signals_markings",
        "street_cleaning_waste",
        "street_lighting",
        "trees_vegetation"
      ],
      "status": "certified",
      "holdout_acc": 0.9681,
      "n_holdout": 941,
      "oos": null,
      "live_verified": false,
      "operating": {
        "min_conf": 0.5,
        "note": "4% leak / 96% coverage at 0.5",
        "qhat": 0.95,
        "qhat_note": "selective-risk \u03b1=0.05 on 566-row eval / 565 cal (civic_holdout+civic_oos): 67.7% auto, ub95 bound 3.45%, eval err 2.09% \u2014 vs old floor-0.5's 96.8% auto at 10% unbounded err",
        "mondrian": "per-class qhat: auto 46.1% at \u03b1=0.05 \u2014 narrow 14-route lane sustains real class-conditional automation",
        "limitations": [
          "14-class narrow lane \u2014 genuine per-class automation viable (46%)",
          "EN civic corpus only"
        ]
      },
      "yanked": true,
      "superseded_by": "civic@0.2.1",
      "yank_reason": "superseded by 0.2.1 (oos class + conformal eval); migrate with: python3 tools/lane_install.py civic@0.2.1 --apply",
      "conformal": {
        "alpha_0.15": {
          "qhat": 0.0392,
          "auto": 0.8098,
          "err": 0.0013,
          "avgset": 0.81
        },
        "alpha_0.1": {
          "qhat": 0.0665,
          "auto": 0.8587,
          "err": 0.0012,
          "avgset": 0.859
        },
        "alpha_0.05": {
          "qhat": 0.2288,
          "auto": 0.9373,
          "err": 0.0079,
          "avgset": 0.937
        },
        "alpha_0.02": {
          "qhat": 0.8996,
          "auto": 0.9139,
          "err": 0.007,
          "avgset": 1.112
        }
      },
      "license": "CC-BY attribution (mySociety/FixMyStreet source reports); weights Apache-2.0",
      "pricing": {
        "tier": "free"
      },
      "provenance": "FixMyStreet Open311 API (public citizen reports, fetched 2026-10-11); 320 council labels -> 13 canonical routes; train 3751/holdout 941; tools/emit_civic.py",
      "eval_sample": null,
      "runtime": "anvil-serve >= v0.3.0 (ANVIL_EXPERTS=<lane>:<head>, ANVIL_EXPERT_QHAT if provided)",
      "published_at": 1791690380
    },
    {
      "name": "civic",
      "version": "0.2.1",
      "file": "civic-0.2.1.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/civic-0.2.1.lanepack.tar.gz",
      "sha256": "e9a656ff92a549d25c11f21e9f65da9f877620d139bfd6db733015c82f167da7",
      "head_sha256": "bb3ce7922c34aeb021bc248fce361175e0c3c84727fb8422b7cac42d75100b02",
      "classes": 14,
      "taxonomy": [
        "animals",
        "drains_flooding",
        "fly_tipping",
        "graffiti_flyposting",
        "oos",
        "other",
        "parking_vehicles",
        "parks_green",
        "pavement_footpath",
        "pothole",
        "signs_signals_markings",
        "street_cleaning_waste",
        "street_lighting",
        "trees_vegetation"
      ],
      "status": "demo",
      "holdout_acc": 0.866,
      "n_holdout": 941,
      "oos": {
        "oos_recall": 0.968,
        "id_oos_recall": 0.925,
        "far_oos_recall": 0.98,
        "in_scope_false_oos": 0.0021
      },
      "live_verified": true,
      "operating": {
        "min_conf": 0.5,
        "note": "4% leak / 96% coverage at 0.5",
        "qhat": 0.95,
        "qhat_note": "selective-risk \u03b1=0.05 on 566-row eval / 565 cal (civic_holdout+civic_oos): 67.7% auto, ub95 bound 3.45%, eval err 2.09% \u2014 vs old floor-0.5's 96.8% auto at 10% unbounded err",
        "mondrian": "per-class qhat: auto 46.1% at \u03b1=0.05 \u2014 narrow 14-route lane sustains real class-conditional automation",
        "limitations": [
          "14-class narrow lane \u2014 genuine per-class automation viable (46%)",
          "EN civic corpus only"
        ]
      },
      "yanked": null,
      "superseded_by": null,
      "yank_reason": null,
      "conformal": null,
      "license": "FixMyStreet Open311 corpus (public reports); OOS negatives: hand-written + CLINC150 (CC)",
      "pricing": {
        "tier": "free"
      },
      "provenance": "civic 0.1.0 + explicit oos class: 48 council-adjacent negatives + 200 CLINC train texts; live-verified on anvil 8437; supersedes yanked 0.2.0 (label-misaligned fit)",
      "eval_sample": null,
      "runtime": "anvil-serve >= v0.3.0 (ANVIL_EXPERTS=<lane>:<head>, ANVIL_EXPERT_QHAT if provided)",
      "published_at": 1791695826
    },
    {
      "name": "civic_fr",
      "version": "0.1.0",
      "file": "civic_fr-0.1.0.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/civic_fr-0.1.0.lanepack.tar.gz",
      "sha256": "1c14200b233ffc3d8700ee3a03d377b90c256c395db86cbcee3904146b6b0e13",
      "head_sha256": "7789fe17b855a823cecffba0e3da2d6fbbdc3dbbbbd751bc5d6a9d54f3ff7eb2",
      "classes": 11,
      "taxonomy": [
        "activites_commerciales",
        "arbres_vegetation",
        "autre",
        "eau",
        "eclairage",
        "graffiti_affiches",
        "mobilier_urbain",
        "objets_encombrants",
        "proprete_dechets",
        "stationnement_epaves",
        "voirie_trottoir"
      ],
      "status": "demo",
      "holdout_acc": 1.0,
      "n_holdout": 330,
      "oos": null,
      "live_verified": false,
      "operating": {
        "min_conf": null,
        "note": "confidence ~1.0 everywhere; gating cannot discriminate, argmax abstain carries it",
        "qhat": 0.746,
        "qhat_note": "v3w head (v2 corpus + 110 natural-phrasing rows x4 weight): conformal on 352-row cal split, alpha=0.02 -> qhat 0.746. Holdout: 96.7% auto, err_auto 0.31%, coverage 99.7%. Natural eval (22 phrasings): auto 86.4%, err 5.26%, coverage 95.5%. vs v2 head: natural_eval 7/22 -> 22/22 (route-or-candidate).",
        "mondrian": "per-class qhat: auto 89.2% at \u03b1=0.05, worst-class 78.6% (12 classes) \u2014 narrow taxonomy sustains real class-conditional automation, the standout lane",
        "golden_note": "RESOLVED 2026-10-13: confident-oos singleton on 'lampadaire en panne' was a corpus-narrowness defect, NOT a prompt mismatch (v2 features were already DECIDE_SYS). Fixed by training civic_fr_v3w_mt.head with 110 hand-written natural phrasings across all 11 classes; deployed to prod 8402, golden probe now eclairage@1.00. Bank77's 'lost my card'->oos is CORRECT scope rejection (bank50 subset lacks lost_or_stolen_card).",
        "limitations": [
          "synthetic corpus: in-scope numbers reflect template separability, not production skill",
          "natural_eval measures only 22 phrasings; wider natural coverage unmeasured",
          "best-calibrated lane under mondrian (89% per-class auto, 97.7% marginal)"
        ]
      },
      "yanked": true,
      "superseded_by": "civic_fr@0.1.1",
      "yank_reason": "superseded by 0.1.1; migrate with: python3 tools/lane_install.py civic_fr@0.1.1 --apply",
      "conformal": {
        "alpha_0.05": {
          "qhat": 0.0008,
          "auto": 0.979,
          "err": 0.0
        }
      },
      "license": "Taxonomy ODbL Mairie de Paris; synthetic utterances Apache-2.0",
      "pricing": {
        "tier": "free"
      },
      "provenance": "DansMaRue taxonomy (opendata.paris.fr) + tools/build_civic_fr.py template generation; no real free text exists publicly",
      "eval_sample": null,
      "runtime": "anvil-serve >= v0.3.0 (ANVIL_EXPERTS=<lane>:<head>, ANVIL_EXPERT_QHAT if provided)",
      "published_at": 1791691307
    },
    {
      "name": "civic_fr",
      "version": "0.1.1",
      "file": "civic_fr-0.1.1.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/civic_fr-0.1.1.lanepack.tar.gz",
      "sha256": "c0fbf627a8f86f489a23982bd81167335d7b4416529e1de44b53f2c88d2df66a",
      "head_sha256": "33c8a5dcedc57d1c121a3473496954feab05884df9811c1e9c5be6e45fc97100",
      "classes": 12,
      "taxonomy": [
        "activites_commerciales",
        "oos",
        "arbres_vegetation",
        "autre",
        "eau",
        "eclairage",
        "graffiti_affiches",
        "mobilier_urbain",
        "objets_encombrants",
        "proprete_dechets",
        "stationnement_epaves",
        "voirie_trottoir"
      ],
      "status": "demo",
      "holdout_acc": 0.997,
      "n_holdout": 330,
      "oos": {
        "oos_recall": 0.909,
        "id_oos_recall": 0.833,
        "far_oos_recall": 1.0,
        "in_scope_false_oos": 0.0
      },
      "live_verified": true,
      "operating": {
        "min_conf": null,
        "note": "confidence ~1.0 everywhere; gating cannot discriminate, argmax abstain carries it",
        "qhat": 0.746,
        "qhat_note": "v3w head (v2 corpus + 110 natural-phrasing rows x4 weight): conformal on 352-row cal split, alpha=0.02 -> qhat 0.746. Holdout: 96.7% auto, err_auto 0.31%, coverage 99.7%. Natural eval (22 phrasings): auto 86.4%, err 5.26%, coverage 95.5%. vs v2 head: natural_eval 7/22 -> 22/22 (route-or-candidate).",
        "mondrian": "per-class qhat: auto 89.2% at \u03b1=0.05, worst-class 78.6% (12 classes) \u2014 narrow taxonomy sustains real class-conditional automation, the standout lane",
        "golden_note": "RESOLVED 2026-10-13: confident-oos singleton on 'lampadaire en panne' was a corpus-narrowness defect, NOT a prompt mismatch (v2 features were already DECIDE_SYS). Fixed by training civic_fr_v3w_mt.head with 110 hand-written natural phrasings across all 11 classes; deployed to prod 8402, golden probe now eclairage@1.00. Bank77's 'lost my card'->oos is CORRECT scope rejection (bank50 subset lacks lost_or_stolen_card).",
        "limitations": [
          "synthetic corpus: in-scope numbers reflect template separability, not production skill",
          "natural_eval measures only 22 phrasings; wider natural coverage unmeasured",
          "best-calibrated lane under mondrian (89% per-class auto, 97.7% marginal)"
        ]
      },
      "yanked": null,
      "superseded_by": null,
      "yank_reason": null,
      "conformal": null,
      "license": "Taxonomy ODbL Mairie de Paris; synthetic utterances Apache-2.0",
      "pricing": {
        "tier": "free"
      },
      "provenance": "civic_fr 0.1.0 + explicit oos class (FR hard+easy negatives); live-verified",
      "eval_sample": null,
      "runtime": "anvil-serve >= v0.3.0 (ANVIL_EXPERTS=<lane>:<head>, ANVIL_EXPERT_QHAT if provided)",
      "published_at": 1791697925
    },
    {
      "name": "civic_fr",
      "version": "0.2.0",
      "file": "civic_fr-0.2.0.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/civic_fr-0.2.0.lanepack.tar.gz",
      "sha256": "edf31ea7c0b73e0f0dbefbd21f7d95cae842f6c9a74b850e3729d31bae18f721",
      "head_sha256": "6527ef96666ca3f3edd2a27f5b16cd532ff57f97829b4cb34255698c73d18afb",
      "classes": 12,
      "taxonomy": [
        "activites_commerciales",
        "oos",
        "arbres_vegetation",
        "autre",
        "eau",
        "eclairage",
        "graffiti_affiches",
        "mobilier_urbain",
        "objets_encombrants",
        "proprete_dechets",
        "stationnement_epaves",
        "voirie_trottoir"
      ],
      "status": "demo",
      "holdout_acc": 0.988,
      "n_holdout": 330,
      "oos": {
        "oos_recall": 0.682,
        "id_oos_recall": 0.417,
        "far_oos_recall": 1.0,
        "in_scope_false_oos": 0.0
      },
      "live_verified": true,
      "operating": {
        "min_conf": null,
        "note": "confidence ~1.0 everywhere; gating cannot discriminate, argmax abstain carries it",
        "qhat": 0.746,
        "qhat_note": "v3w head (v2 corpus + 110 natural-phrasing rows x4 weight): conformal on 352-row cal split, alpha=0.02 -> qhat 0.746. Holdout: 96.7% auto, err_auto 0.31%, coverage 99.7%. Natural eval (22 phrasings): auto 86.4%, err 5.26%, coverage 95.5%. vs v2 head: natural_eval 7/22 -> 22/22 (route-or-candidate).",
        "mondrian": "per-class qhat: auto 89.2% at \u03b1=0.05, worst-class 78.6% (12 classes) \u2014 narrow taxonomy sustains real class-conditional automation, the standout lane",
        "golden_note": "RESOLVED 2026-10-13: confident-oos singleton on 'lampadaire en panne' was a corpus-narrowness defect, NOT a prompt mismatch (v2 features were already DECIDE_SYS). Fixed by training civic_fr_v3w_mt.head with 110 hand-written natural phrasings across all 11 classes; deployed to prod 8402, golden probe now eclairage@1.00. Bank77's 'lost my card'->oos is CORRECT scope rejection (bank50 subset lacks lost_or_stolen_card).",
        "limitations": [
          "synthetic corpus: in-scope numbers reflect template separability, not production skill",
          "natural_eval measures only 22 phrasings; wider natural coverage unmeasured",
          "best-calibrated lane under mondrian (89% per-class auto, 97.7% marginal)"
        ]
      },
      "yanked": true,
      "superseded_by": "civic_fr@0.2.1",
      "yank_reason": "0.2.1 embeds eval_sample.jsonl (194 rows) making every published metric re-runnable via lane_repro.py; same v3w head, same evals",
      "conformal": {
        "alpha": 0.02,
        "qhat": 0.746,
        "cal_n": 352,
        "holdout": {
          "auto_pct": 96.7,
          "err_auto_pct": 0.31,
          "coverage_pct": 99.7
        },
        "natural": {
          "auto_pct": 86.4,
          "err_auto_pct": 5.26,
          "coverage_pct": 95.5
        }
      },
      "license": "Apache-2.0",
      "pricing": {
        "tier": "free"
      },
      "provenance": "Synthetic civic corpus + hand-written natural phrasings (Apache-2.0 compatible)",
      "eval_sample": null,
      "runtime": "anvil-serve >= v0.3.0 (ANVIL_EXPERTS=<lane>:<head>, ANVIL_EXPERT_QHAT if provided)",
      "published_at": 1791879815
    },
    {
      "name": "civic_fr",
      "version": "0.2.1",
      "file": "civic_fr-0.2.1.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/civic_fr-0.2.1.lanepack.tar.gz",
      "sha256": "12e6803c49b9a9760ddad68a57752391bdee940ae4b2159d390de6b5627ed5d8",
      "head_sha256": "6527ef96666ca3f3edd2a27f5b16cd532ff57f97829b4cb34255698c73d18afb",
      "classes": 12,
      "taxonomy": [
        "activites_commerciales",
        "oos",
        "arbres_vegetation",
        "autre",
        "eau",
        "eclairage",
        "graffiti_affiches",
        "mobilier_urbain",
        "objets_encombrants",
        "proprete_dechets",
        "stationnement_epaves",
        "voirie_trottoir"
      ],
      "status": "demo",
      "holdout_acc": 0.988,
      "n_holdout": 330,
      "oos": {
        "oos_recall": 0.682,
        "id_oos_recall": 0.417,
        "far_oos_recall": 1.0,
        "in_scope_false_oos": 0.0
      },
      "live_verified": true,
      "operating": {
        "min_conf": null,
        "note": "confidence ~1.0 everywhere; gating cannot discriminate, argmax abstain carries it",
        "qhat": 0.746,
        "qhat_note": "v3w head (v2 corpus + 110 natural-phrasing rows x4 weight): conformal on 352-row cal split, alpha=0.02 -> qhat 0.746. Holdout: 96.7% auto, err_auto 0.31%, coverage 99.7%. Natural eval (22 phrasings): auto 86.4%, err 5.26%, coverage 95.5%. vs v2 head: natural_eval 7/22 -> 22/22 (route-or-candidate).",
        "mondrian": "per-class qhat: auto 89.2% at \u03b1=0.05, worst-class 78.6% (12 classes) \u2014 narrow taxonomy sustains real class-conditional automation, the standout lane",
        "golden_note": "RESOLVED 2026-10-13: confident-oos singleton on 'lampadaire en panne' was a corpus-narrowness defect, NOT a prompt mismatch (v2 features were already DECIDE_SYS). Fixed by training civic_fr_v3w_mt.head with 110 hand-written natural phrasings across all 11 classes; deployed to prod 8402, golden probe now eclairage@1.00. Bank77's 'lost my card'->oos is CORRECT scope rejection (bank50 subset lacks lost_or_stolen_card).",
        "limitations": [
          "synthetic corpus: in-scope numbers reflect template separability, not production skill",
          "natural_eval measures only 22 phrasings; wider natural coverage unmeasured",
          "best-calibrated lane under mondrian (89% per-class auto, 97.7% marginal)"
        ]
      },
      "yanked": null,
      "superseded_by": null,
      "yank_reason": null,
      "conformal": {
        "alpha": 0.02,
        "qhat": 0.746,
        "cal_n": 352,
        "holdout": {
          "auto_pct": 96.7,
          "err_auto_pct": 0.31,
          "coverage_pct": 99.7
        },
        "natural": {
          "auto_pct": 86.4,
          "err_auto_pct": 5.26,
          "coverage_pct": 95.5
        }
      },
      "license": "Apache-2.0",
      "pricing": {
        "tier": "free"
      },
      "provenance": "Synthetic civic corpus + hand-written natural phrasings (Apache-2.0 compatible)",
      "eval_sample": "eval_sample.jsonl",
      "runtime": "anvil-serve >= v0.3.0 (ANVIL_EXPERTS=<lane>:<head>, ANVIL_EXPERT_QHAT if provided)",
      "published_at": 1791879815
    },
    {
      "name": "clinc150",
      "version": "0.1.2",
      "file": "clinc150-0.1.2.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/clinc150-0.1.2.lanepack.tar.gz",
      "sha256": "7257a15d2a2282cee754af35d1ba203cfa5d72bbf988979033f68f5b73969dc3",
      "head_sha256": "2110652d65a994984176da3d20596f094f5cb7550f70c117a76aac36bae733ec",
      "classes": 151,
      "taxonomy": [
        "accept_reservations",
        "account_blocked",
        "alarm",
        "application_status",
        "apr",
        "are_you_a_bot",
        "balance",
        "bill_balance",
        "bill_due",
        "book_flight",
        "book_hotel",
        "calculator",
        "calendar",
        "calendar_update",
        "calories",
        "cancel",
        "cancel_reservation",
        "car_rental",
        "card_declined",
        "carry_on",
        "change_accent",
        "change_ai_name",
        "change_language",
        "change_speed",
        "change_user_name",
        "change_volume",
        "confirm_reservation",
        "cook_time",
        "credit_limit",
        "credit_limit_change",
        "credit_score",
        "current_location",
        "damaged_card",
        "date",
        "definition",
        "direct_deposit",
        "directions",
        "distance",
        "do_you_have_pets",
        "exchange_rate",
        "expiration_date",
        "find_phone",
        "flight_status",
        "flip_coin",
        "food_last",
        "freeze_account",
        "fun_fact",
        "gas",
        "gas_type",
        "goodbye",
        "greeting",
        "how_busy",
        "how_old_are_you",
        "improve_credit_score",
        "income",
        "ingredient_substitution",
        "ingredients_list",
        "insurance",
        "insurance_change",
        "interest_rate",
        "international_fees",
        "international_visa",
        "jump_start",
        "last_maintenance",
        "lost_luggage",
        "make_call",
        "maybe",
        "meal_suggestion",
        "meaning_of_life",
        "measurement_conversion",
        "meeting_schedule",
        "min_payment",
        "mpg",
        "new_card",
        "next_holiday",
        "next_song",
        "no",
        "nutrition_info",
        "oil_change_how",
        "oil_change_when",
        "oos",
        "order",
        "order_checks",
        "order_status",
        "pay_bill",
        "payday",
        "pin_change",
        "play_music",
        "plug_type",
        "pto_balance",
        "pto_request",
        "pto_request_status",
        "pto_used",
        "recipe",
        "redeem_rewards",
        "reminder",
        "reminder_update",
        "repeat",
        "replacement_card_duration",
        "report_fraud",
        "report_lost_card",
        "reset_settings",
        "restaurant_reservation",
        "restaurant_reviews",
        "restaurant_suggestion",
        "rewards_balance",
        "roll_dice",
        "rollover_401k",
        "routing",
        "schedule_maintenance",
        "schedule_meeting",
        "share_location",
        "shopping_list",
        "shopping_list_update",
        "smart_home",
        "spelling",
        "spending_history",
        "sync_device",
        "taxes",
        "tell_joke",
        "text",
        "thank_you",
        "time",
        "timer",
        "timezone",
        "tire_change",
        "tire_pressure",
        "todo_list",
        "todo_list_update",
        "traffic",
        "transactions",
        "transfer",
        "translate",
        "travel_alert",
        "travel_notification",
        "travel_suggestion",
        "uber",
        "update_playlist",
        "user_name",
        "vaccines",
        "w2",
        "weather",
        "what_are_your_hobbies",
        "what_can_i_ask_you",
        "what_is_your_name",
        "what_song",
        "where_are_you_from",
        "whisper_mode",
        "who_do_you_work_for",
        "who_made_you",
        "yes"
      ],
      "status": "certified",
      "holdout_acc": 0.738,
      "n_holdout": 5500,
      "oos": null,
      "live_verified": false,
      "operating": {
        "min_conf": 0.7,
        "note": "7.8% leak / 69% coverage at 0.7; 15-20% argmax abstain is weak, the floor is the gate",
        "qhat": 0.95,
        "qhat_note": "selective-risk \u03b1=0.05 on 2766-row cal split (clinc_holdout+hn_eval): auto 46.5% (v1) / 46.2% (v2), ub95 bound ~4.5-4.8%, eval err ~3.9%. Both hosted aliases (clinc, clinc2) wired at qhat 0.95. Caveat: cal split nearly all in-scope \u2014 bound covers routing errors, not OOS recall.",
        "mondrian": "per-class qhat: auto 4.7% at \u03b1=0.05 \u2014 wide taxonomies route through clarification sets, not singletons; confusion is local though \u2014 72.6% of multi-sets fall inside one co-occurrence cluster (tools/set_cluster.py, qhat 0.95): a coarse\u2192fine hierarchy could automate at parent level for most abstains; bank77 contrast: 100% one cluster, no hierarchical rescue",
        "limitations": [
          "151 classes: mondrian auto ~4-5%; candidates[] clarification is the primary interaction mode",
          "cal split had ~32 oos rows \u2014 the bound covers routing errors, not oos recall"
        ]
      },
      "yanked": true,
      "superseded_by": "clinc150@0.1.4",
      "yank_reason": "superseded by 0.1.4 (hard-negative retrain: OOS abstain 15.5\u219220.4%, hard-negative recall 0\u219259.4%); migrate with: python3 tools/lane_install.py clinc150@0.1.4 --apply",
      "conformal": {
        "alpha_015": {
          "auto": 0.399,
          "err": 0.018
        },
        "selective_risk_alpha_02": {
          "auto": 0.38,
          "err_bound_95": 0.0187
        }
      },
      "license": "Apache-2.0 (model weights + pack); upstream CLINC150 dataset is Apache-2.0",
      "pricing": {
        "tier": "free"
      },
      "provenance": "Trained on the public CLINC150 benchmark (train split only), holdout eval on untouched test split. v0.1.1: labels.txt repaired; v0.1.2: manifest status field added.",
      "eval_sample": null,
      "runtime": "anvil-serve >= v0.3.0 (ANVIL_EXPERTS=<lane>:<head>, ANVIL_EXPERT_QHAT if provided)",
      "published_at": 1791651859
    },
    {
      "name": "clinc150",
      "version": "0.1.3",
      "file": "clinc150-0.1.3.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/clinc150-0.1.3.lanepack.tar.gz",
      "sha256": "d46f14e1f6e6f3095ff3871d33c488290ab34a0b7802c89d85f359534a8fb6de",
      "head_sha256": "2110652d65a994984176da3d20596f094f5cb7550f70c117a76aac36bae733ec",
      "classes": 151,
      "taxonomy": [
        "accept_reservations",
        "account_blocked",
        "alarm",
        "application_status",
        "apr",
        "are_you_a_bot",
        "balance",
        "bill_balance",
        "bill_due",
        "book_flight",
        "book_hotel",
        "calculator",
        "calendar",
        "calendar_update",
        "calories",
        "cancel",
        "cancel_reservation",
        "car_rental",
        "card_declined",
        "carry_on",
        "change_accent",
        "change_ai_name",
        "change_language",
        "change_speed",
        "change_user_name",
        "change_volume",
        "confirm_reservation",
        "cook_time",
        "credit_limit",
        "credit_limit_change",
        "credit_score",
        "current_location",
        "damaged_card",
        "date",
        "definition",
        "direct_deposit",
        "directions",
        "distance",
        "do_you_have_pets",
        "exchange_rate",
        "expiration_date",
        "find_phone",
        "flight_status",
        "flip_coin",
        "food_last",
        "freeze_account",
        "fun_fact",
        "gas",
        "gas_type",
        "goodbye",
        "greeting",
        "how_busy",
        "how_old_are_you",
        "improve_credit_score",
        "income",
        "ingredient_substitution",
        "ingredients_list",
        "insurance",
        "insurance_change",
        "interest_rate",
        "international_fees",
        "international_visa",
        "jump_start",
        "last_maintenance",
        "lost_luggage",
        "make_call",
        "maybe",
        "meal_suggestion",
        "meaning_of_life",
        "measurement_conversion",
        "meeting_schedule",
        "min_payment",
        "mpg",
        "new_card",
        "next_holiday",
        "next_song",
        "no",
        "nutrition_info",
        "oil_change_how",
        "oil_change_when",
        "oos",
        "order",
        "order_checks",
        "order_status",
        "pay_bill",
        "payday",
        "pin_change",
        "play_music",
        "plug_type",
        "pto_balance",
        "pto_request",
        "pto_request_status",
        "pto_used",
        "recipe",
        "redeem_rewards",
        "reminder",
        "reminder_update",
        "repeat",
        "replacement_card_duration",
        "report_fraud",
        "report_lost_card",
        "reset_settings",
        "restaurant_reservation",
        "restaurant_reviews",
        "restaurant_suggestion",
        "rewards_balance",
        "roll_dice",
        "rollover_401k",
        "routing",
        "schedule_maintenance",
        "schedule_meeting",
        "share_location",
        "shopping_list",
        "shopping_list_update",
        "smart_home",
        "spelling",
        "spending_history",
        "sync_device",
        "taxes",
        "tell_joke",
        "text",
        "thank_you",
        "time",
        "timer",
        "timezone",
        "tire_change",
        "tire_pressure",
        "todo_list",
        "todo_list_update",
        "traffic",
        "transactions",
        "transfer",
        "translate",
        "travel_alert",
        "travel_notification",
        "travel_suggestion",
        "uber",
        "update_playlist",
        "user_name",
        "vaccines",
        "w2",
        "weather",
        "what_are_your_hobbies",
        "what_can_i_ask_you",
        "what_is_your_name",
        "what_song",
        "where_are_you_from",
        "whisper_mode",
        "who_do_you_work_for",
        "who_made_you",
        "yes"
      ],
      "status": "certified",
      "holdout_acc": 0.738,
      "n_holdout": 5500,
      "oos": {
        "oos_recall": 0.15,
        "oos_leak_curve": {
          "0.3": 0.275,
          "0.5": 0.125,
          "0.7": 0.075,
          "0.9": 0.025
        },
        "coverage_curve": {
          "0.3": 0.92,
          "0.5": 0.86,
          "0.7": 0.7,
          "0.9": 0.38
        },
        "oos_eval_note": "argmax abstain share 15% on held-out CLINC oos rows (n=40); leaks carry low conf (median 0.23 vs in-scope 0.86) \u2014 conf>=0.5 gate cuts leak to 12.5%, conf>=0.7 to 7.5%. bench/oos_eval_catalog.json"
      },
      "live_verified": true,
      "operating": {
        "min_conf": 0.7,
        "note": "7.8% leak / 69% coverage at 0.7; 15-20% argmax abstain is weak, the floor is the gate",
        "qhat": 0.95,
        "qhat_note": "selective-risk \u03b1=0.05 on 2766-row cal split (clinc_holdout+hn_eval): auto 46.5% (v1) / 46.2% (v2), ub95 bound ~4.5-4.8%, eval err ~3.9%. Both hosted aliases (clinc, clinc2) wired at qhat 0.95. Caveat: cal split nearly all in-scope \u2014 bound covers routing errors, not OOS recall.",
        "mondrian": "per-class qhat: auto 4.7% at \u03b1=0.05 \u2014 wide taxonomies route through clarification sets, not singletons; confusion is local though \u2014 72.6% of multi-sets fall inside one co-occurrence cluster (tools/set_cluster.py, qhat 0.95): a coarse\u2192fine hierarchy could automate at parent level for most abstains; bank77 contrast: 100% one cluster, no hierarchical rescue",
        "limitations": [
          "151 classes: mondrian auto ~4-5%; candidates[] clarification is the primary interaction mode",
          "cal split had ~32 oos rows \u2014 the bound covers routing errors, not oos recall"
        ]
      },
      "yanked": true,
      "superseded_by": "clinc150@0.1.4",
      "yank_reason": "superseded by 0.1.4 (hard-negative retrain: OOS abstain 15.5\u219220.4%, hard-negative recall 0\u219259.4%); migrate with: python3 tools/lane_install.py clinc150@0.1.4 --apply",
      "conformal": {
        "alpha_015": {
          "auto": 0.399,
          "err": 0.018
        },
        "selective_risk_alpha_02": {
          "auto": 0.38,
          "err_bound_95": 0.0187
        }
      },
      "license": "/tmp/clinc012/LICENSE.txt",
      "pricing": {
        "tier": "free"
      },
      "provenance": "Trained on the public CLINC150 benchmark (train split only), holdout eval on untouched test split. v0.1.1: labels.txt repaired; v0.1.2: manifest status field added. v0.1.3: eval.json gains measured oos_recall + leak/coverage curves (bench/oos_eval_catalog); head unchanged.",
      "eval_sample": null,
      "runtime": "anvil-serve >= v0.3.0 (ANVIL_EXPERTS=<lane>:<head>, ANVIL_EXPERT_QHAT if provided)",
      "published_at": 1791739486
    },
    {
      "name": "clinc150",
      "version": "0.1.4",
      "file": "clinc150-0.1.4.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/clinc150-0.1.4.lanepack.tar.gz",
      "sha256": "c5862061595c081e23d4a42ce28dc20a9db550e0f4e33aa54f25c597e9a3652e",
      "head_sha256": "b868762377aa03a7eeb0b882add8b84f810b81a2a689eeffdbb9f8e42e463250",
      "classes": 151,
      "taxonomy": [
        "accept_reservations",
        "account_blocked",
        "alarm",
        "application_status",
        "apr",
        "are_you_a_bot",
        "balance",
        "bill_balance",
        "bill_due",
        "book_flight",
        "book_hotel",
        "calculator",
        "calendar",
        "calendar_update",
        "calories",
        "cancel",
        "cancel_reservation",
        "car_rental",
        "card_declined",
        "carry_on",
        "change_accent",
        "change_ai_name",
        "change_language",
        "change_speed",
        "change_user_name",
        "change_volume",
        "confirm_reservation",
        "cook_time",
        "credit_limit",
        "credit_limit_change",
        "credit_score",
        "current_location",
        "damaged_card",
        "date",
        "definition",
        "direct_deposit",
        "directions",
        "distance",
        "do_you_have_pets",
        "exchange_rate",
        "expiration_date",
        "find_phone",
        "flight_status",
        "flip_coin",
        "food_last",
        "freeze_account",
        "fun_fact",
        "gas",
        "gas_type",
        "goodbye",
        "greeting",
        "how_busy",
        "how_old_are_you",
        "improve_credit_score",
        "income",
        "ingredient_substitution",
        "ingredients_list",
        "insurance",
        "insurance_change",
        "interest_rate",
        "international_fees",
        "international_visa",
        "jump_start",
        "last_maintenance",
        "lost_luggage",
        "make_call",
        "maybe",
        "meal_suggestion",
        "meaning_of_life",
        "measurement_conversion",
        "meeting_schedule",
        "min_payment",
        "mpg",
        "new_card",
        "next_holiday",
        "next_song",
        "no",
        "nutrition_info",
        "oil_change_how",
        "oil_change_when",
        "oos",
        "order",
        "order_checks",
        "order_status",
        "pay_bill",
        "payday",
        "pin_change",
        "play_music",
        "plug_type",
        "pto_balance",
        "pto_request",
        "pto_request_status",
        "pto_used",
        "recipe",
        "redeem_rewards",
        "reminder",
        "reminder_update",
        "repeat",
        "replacement_card_duration",
        "report_fraud",
        "report_lost_card",
        "reset_settings",
        "restaurant_reservation",
        "restaurant_reviews",
        "restaurant_suggestion",
        "rewards_balance",
        "roll_dice",
        "rollover_401k",
        "routing",
        "schedule_maintenance",
        "schedule_meeting",
        "share_location",
        "shopping_list",
        "shopping_list_update",
        "smart_home",
        "spelling",
        "spending_history",
        "sync_device",
        "taxes",
        "tell_joke",
        "text",
        "thank_you",
        "time",
        "timer",
        "timezone",
        "tire_change",
        "tire_pressure",
        "todo_list",
        "todo_list_update",
        "traffic",
        "transactions",
        "transfer",
        "translate",
        "travel_alert",
        "travel_notification",
        "travel_suggestion",
        "uber",
        "update_playlist",
        "user_name",
        "vaccines",
        "w2",
        "weather",
        "what_are_your_hobbies",
        "what_can_i_ask_you",
        "what_is_your_name",
        "what_song",
        "where_are_you_from",
        "whisper_mode",
        "who_do_you_work_for",
        "who_made_you",
        "yes"
      ],
      "status": "certified",
      "holdout_acc": 0.7411,
      "n_holdout": 5500,
      "oos": {
        "oos_recall": 0.204,
        "oos_leak_curve": {
          "0.5": 0.183,
          "0.7": 0.078,
          "0.9": 0.019
        },
        "coverage_curve": {
          "0.5": 0.812,
          "0.7": 0.687,
          "0.9": 0.452
        },
        "oos_eval_note": "argmax abstain 20.4% on held-out CLINC oos (n=1000, was 15.5% in 0.1.3) and 59% on 200-row hard-negative set (was 0%). Leaks carry low conf - conf>=0.7 gate gives 7.8% leak / 68.7% coverage. Retrained with +200 intent-adjacent hard negatives (pizza delivery, dentist, wifi, legal...)"
      },
      "live_verified": true,
      "operating": {
        "min_conf": 0.7,
        "note": "7.8% leak / 69% coverage at 0.7; 15-20% argmax abstain is weak, the floor is the gate",
        "qhat": 0.95,
        "qhat_note": "selective-risk \u03b1=0.05 on 2766-row cal split (clinc_holdout+hn_eval): auto 46.5% (v1) / 46.2% (v2), ub95 bound ~4.5-4.8%, eval err ~3.9%. Both hosted aliases (clinc, clinc2) wired at qhat 0.95. Caveat: cal split nearly all in-scope \u2014 bound covers routing errors, not OOS recall.",
        "mondrian": "per-class qhat: auto 4.7% at \u03b1=0.05 \u2014 wide taxonomies route through clarification sets, not singletons; confusion is local though \u2014 72.6% of multi-sets fall inside one co-occurrence cluster (tools/set_cluster.py, qhat 0.95): a coarse\u2192fine hierarchy could automate at parent level for most abstains; bank77 contrast: 100% one cluster, no hierarchical rescue",
        "limitations": [
          "151 classes: mondrian auto ~4-5%; candidates[] clarification is the primary interaction mode",
          "cal split had ~32 oos rows \u2014 the bound covers routing errors, not oos recall"
        ]
      },
      "yanked": null,
      "superseded_by": null,
      "yank_reason": null,
      "conformal": {
        "alpha_015": {
          "auto": 0.399,
          "err": 0.018
        },
        "selective_risk_alpha_02": {
          "auto": 0.38,
          "err_bound_95": 0.0187
        }
      },
      "license": "/tmp/clincp/LICENSE.txt",
      "pricing": {
        "tier": "free"
      },
      "provenance": "Trained on the public CLINC150 benchmark (train split only) + 200 hand-written intent-adjacent hard negatives; holdout eval on untouched test split. v0.1.4: retrained with hard-negative OOS corpus - argmax abstain 15.5%->20.4%, hard-negative recall 0%->59.4%, in-scope 86.7%->86.0%. Head bytes differ from 0.1.3.",
      "eval_sample": null,
      "runtime": "anvil-serve >= v0.3.0 (ANVIL_EXPERTS=<lane>:<head>, ANVIL_EXPERT_QHAT if provided)",
      "published_at": 1791745818
    },
    {
      "name": "disposable_email",
      "version": "0.1.0",
      "file": "disposable_email-0.1.0.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/disposable_email-0.1.0.lanepack.tar.gz",
      "sha256": "4365543f3a8f698f41af66aa3bb409c3bc057d979b4a295f00b91ad00784957d",
      "head_sha256": null,
      "classes": 2,
      "taxonomy": [
        "legit",
        "disposable"
      ],
      "status": "certified",
      "holdout_acc": null,
      "n_holdout": null,
      "oos": null,
      "live_verified": false,
      "operating": null,
      "yanked": null,
      "superseded_by": null,
      "yank_reason": null,
      "conformal": null,
      "license": "Apache-2.0",
      "pricing": {
        "tier": "free"
      },
      "provenance": "disposable-email-domains public blocklist (martenson/disposable-email-domains), deterministic domains rule",
      "eval_sample": "eval_sample.jsonl",
      "runtime": "anvil-serve rules lanes",
      "published_at": 1792741597
    },
    {
      "name": "eq_tone",
      "version": "0.1.1",
      "file": "eq_tone-0.1.1.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/eq_tone-0.1.1.lanepack.tar.gz",
      "sha256": "df3401cae3d9a27172b39b4b06316b517224e7fa67207cd5e1f096cb82f183e5",
      "head_sha256": "f47c29e3709d7998863e6dee5432584217dfc45f0722dcd46b215c0513ebd79a",
      "classes": 5,
      "taxonomy": [
        "confused",
        "distressed",
        "frustrated",
        "neutral",
        "positive"
      ],
      "status": "demo",
      "holdout_acc": 0.5137,
      "n_holdout": 4995,
      "oos": null,
      "live_verified": false,
      "operating": {
        "min_conf": 0.9,
        "note": "HEAD CALIBRATION WARNING (eq_test 4995-row eval, head=eq_tone_v2): confidence nearly uninformative \u2014 floor 0.5 \u2192 62.7% auto at 39.4% error; floor 0.9 \u2192 5.6% auto at 1.4% err. No selective-risk conformal point meets \u03b1\u22640.05; best is \u03b1=0.10 at qhat 0.10 \u2192 5.6% auto \u22648.0% bound. Hosted minconf 0.9. eq_tone_v3 retrain (matched DECIDE_SYS features) evaluated and REJECTED \u2014 no improvement; the ~50% ceiling is the trunk's, not a data artifact.",
        "mondrian": "per-class auto 7.4% at \u03b1=0.05, worst-class 0.9% (5 classes) \u2014 the head cannot discriminate within classes either",
        "limitations": [
          "trunk ceiling, not a prompt artifact: eq_tone_v3 (retrain under matched DECIDE_SYS features, label-smoothing sweep) evaluated 2026-10-13 and REJECTED \u2014 selective curve no better than v2 (0.9 floor: 5.2% auto @ 4.6% err vs v2's 5.6% @ 1.4%). ~50% accuracy on 6-way tone is what the ~7M trunk can do; lane stays triage-only",
          "minconf 0.9 in production \u2014 automates only ~5.6% of traffic at ~1.4% err; below that, calibration gives no useful bound"
        ]
      },
      "yanked": true,
      "superseded_by": "eq_tone@0.2.0",
      "yank_reason": "superseded \u2014 v0.2.0 adds an explicit oos class (87.9% non-tone recall); migrate with: lane_install eq_tone@0.2.0 --apply",
      "conformal": {
        "alpha_0.15": {
          "qhat": 0.8667,
          "auto": 0.158,
          "err": 0.17
        },
        "alpha_0.10": {
          "qhat": 0.891,
          "auto": 0.125,
          "err": 0.143
        },
        "alpha_0.05": {
          "qhat": 0.9248,
          "auto": 0.081,
          "err": 0.116
        },
        "alpha_0.02": {
          "qhat": 0.9559,
          "auto": 0.049,
          "err": 0.086
        }
      },
      "license": "Apache-2.0 (weights); GoEmotions dataset Apache-2.0; Reddit source text remains with its authors",
      "pricing": {
        "tier": "free"
      },
      "provenance": "GoEmotions (google-research, Reddit comments, Apache-2.0), 28 labels merged to 5 tone superclasses; trained on frozen m7router3s384 trunk via LR head. v0.1.1: manifest status field added.",
      "eval_sample": null,
      "runtime": "anvil-serve >= v0.3.0 (ANVIL_EXPERTS=<lane>:<head>, ANVIL_EXPERT_QHAT if provided)",
      "published_at": 1791651859
    },
    {
      "name": "eq_tone",
      "version": "0.2.0",
      "file": "eq_tone-0.2.0.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/eq_tone-0.2.0.lanepack.tar.gz",
      "sha256": "2e2c608b1935731c0553a35b85b1d3f859bbf83dd343f1cdfad00102ba2ddf00",
      "head_sha256": "c9ca011880b679f65f8584c7faf902d4c8d62c07b506fe94ff66c16a4540589f",
      "classes": 6,
      "taxonomy": [
        "confused",
        "distressed",
        "frustrated",
        "neutral",
        "oos",
        "positive"
      ],
      "status": "demo",
      "holdout_acc": 0.5029,
      "n_holdout": 4995,
      "oos": {
        "oos_recall": 0.8789
      },
      "live_verified": true,
      "operating": {
        "min_conf": 0.9,
        "note": "HEAD CALIBRATION WARNING (eq_test 4995-row eval, head=eq_tone_v2): confidence nearly uninformative \u2014 floor 0.5 \u2192 62.7% auto at 39.4% error; floor 0.9 \u2192 5.6% auto at 1.4% err. No selective-risk conformal point meets \u03b1\u22640.05; best is \u03b1=0.10 at qhat 0.10 \u2192 5.6% auto \u22648.0% bound. Hosted minconf 0.9. eq_tone_v3 retrain (matched DECIDE_SYS features) evaluated and REJECTED \u2014 no improvement; the ~50% ceiling is the trunk's, not a data artifact.",
        "mondrian": "per-class auto 7.4% at \u03b1=0.05, worst-class 0.9% (5 classes) \u2014 the head cannot discriminate within classes either",
        "limitations": [
          "trunk ceiling, not a prompt artifact: eq_tone_v3 (retrain under matched DECIDE_SYS features, label-smoothing sweep) evaluated 2026-10-13 and REJECTED \u2014 selective curve no better than v2 (0.9 floor: 5.2% auto @ 4.6% err vs v2's 5.6% @ 1.4%). ~50% accuracy on 6-way tone is what the ~7M trunk can do; lane stays triage-only",
          "minconf 0.9 in production \u2014 automates only ~5.6% of traffic at ~1.4% err; below that, calibration gives no useful bound"
        ]
      },
      "yanked": null,
      "superseded_by": null,
      "yank_reason": null,
      "conformal": null,
      "license": "Apache-2.0 (weights); GoEmotions dataset Apache-2.0; Reddit source text remains with its authors",
      "pricing": {
        "tier": "free"
      },
      "provenance": "GoEmotions (google-research, Reddit comments, Apache-2.0), 28 labels merged to 5 tone superclasses + explicit oos class (hard negatives: clinc hn, banking77). Trained on frozen m7router3s384 trunk via LR head. v0.2.0: oos class added, live-verified.",
      "eval_sample": null,
      "runtime": "anvil-serve >= v0.3.0 (ANVIL_EXPERTS=<lane>:<head>, ANVIL_EXPERT_QHAT if provided)",
      "published_at": 1791783913
    },
    {
      "name": "extremism",
      "version": "0.1.0",
      "file": "extremism-0.1.0.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/extremism-0.1.0.lanepack.tar.gz",
      "sha256": "eb1e976d8617d23c3a92d9be876f4581eda887b7f1b885d564231308743fda17",
      "head_sha256": null,
      "classes": 2,
      "taxonomy": [
        "clean",
        "extremist"
      ],
      "status": "certified",
      "holdout_acc": null,
      "n_holdout": null,
      "oos": null,
      "live_verified": false,
      "operating": null,
      "yanked": null,
      "superseded_by": null,
      "yank_reason": null,
      "conformal": null,
      "license": "Apache-2.0",
      "pricing": {
        "tier": "free"
      },
      "provenance": "hand-authored call-to-violence patterns + curated term list; covers explicit phrasing not coded/dogwhistle language",
      "eval_sample": "eval_sample.jsonl",
      "runtime": "anvil-serve rules lanes",
      "published_at": 1792742071
    },
    {
      "name": "harassment",
      "version": "0.1.0",
      "file": "harassment-0.1.0.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/harassment-0.1.0.lanepack.tar.gz",
      "sha256": "21d02c85773566d9a1ce04503f0f07ab35853a80dfcb52539699c99e515a8a79",
      "head_sha256": null,
      "classes": 2,
      "taxonomy": [
        "clean",
        "harassment"
      ],
      "status": "certified",
      "holdout_acc": null,
      "n_holdout": null,
      "oos": null,
      "live_verified": false,
      "operating": null,
      "yanked": null,
      "superseded_by": null,
      "yank_reason": null,
      "conformal": null,
      "license": "Apache-2.0",
      "pricing": {
        "tier": "free"
      },
      "provenance": "hand-authored directed-abuse patterns (you-anchored insult + threat formulas); distinct from profanity (vocabulary) \u2014 harassment is target-directed",
      "eval_sample": "eval_sample.jsonl",
      "runtime": "anvil-serve rules lanes",
      "published_at": 1792742797
    },
    {
      "name": "helpdesk",
      "version": "0.1.1",
      "file": "helpdesk-0.1.1.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/helpdesk-0.1.1.lanepack.tar.gz",
      "sha256": "a1846154c250211b9ff26bd3c22f1cb52b66bbff21f5f546de713cbce7059647",
      "head_sha256": "478cae39914d191d49e2290591de70743fbbe557a286a8753bd05d54eb868280",
      "classes": 5,
      "taxonomy": [
        "file_ticket",
        "provision_vm",
        "reset_password",
        "ticket_status",
        "escalate"
      ],
      "status": "demo",
      "holdout_acc": null,
      "n_holdout": null,
      "oos": null,
      "live_verified": false,
      "operating": {
        "min_conf": null,
        "note": "argmax abstains 100% of held-out OOS; no floor needed",
        "qhat": 0.95,
        "qhat_note": "selective-risk on 160-row cal split (holdout_helpdesk 300 + oos_eval 20, halved): helpdesk_v2_mt \u2192 98.1% auto, ub95 1.89%, eval err 0%; helpdesk_pool_L-3_mean \u2192 93.8% auto, ub95 3.10%, eval err 0%. Both hosted aliases (helpdesk2, it_helpdesk) wired at qhat 0.95.",
        "mondrian": "per-class qhat \u03b1=0.05: v2 96.2% auto (worst-class 88%); pool head 90.6% but worst-class 0% \u2014 one route never reaches singletons under class-conditional coverage",
        "limitations": [
          "7-class taxonomy incl. oos \u2014 out-of-taxonomy queries correctly resolve to oos/escalate"
        ]
      },
      "yanked": true,
      "superseded_by": "helpdesk@0.2.0",
      "yank_reason": "demo pack superseded \u2014 0.2.0 adds the oos class, live-verified eval, and measured abstention metrics; migrate with: python3 tools/lane_install.py helpdesk@0.2.0 --apply",
      "conformal": null,
      "license": "Apache-2.0",
      "pricing": {
        "tier": "free"
      },
      "provenance": "Synthetic IT-helpdesk demo corpus (public, generated). v0.1.1: manifest status field added.",
      "eval_sample": null,
      "runtime": "anvil-serve >= v0.3.0 (ANVIL_EXPERTS=<lane>:<head>, ANVIL_EXPERT_QHAT if provided)",
      "published_at": 1791651859
    },
    {
      "name": "helpdesk",
      "version": "0.2.0",
      "file": "helpdesk-0.2.0.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/helpdesk-0.2.0.lanepack.tar.gz",
      "sha256": "dc573f8b22114b973d65dd084fbe732a8dfc82223ccb7e3d7d8cf7c2187c0599",
      "head_sha256": "faa30df566b09a46a23fd6a24391b540ec09f69fa7fd79e4369a7765b4779821",
      "classes": 7,
      "taxonomy": [
        "chat",
        "escalate",
        "file_ticket",
        "oos",
        "provision_vm",
        "reset_password",
        "ticket_status"
      ],
      "status": "demo",
      "holdout_acc": 1.0,
      "n_holdout": null,
      "oos": {
        "oos_recall": 0.9,
        "id_oos_recall": 0.9,
        "far_oos_recall": 0.9,
        "in_scope_false_oos": 0.0
      },
      "live_verified": true,
      "operating": {
        "min_conf": null,
        "note": "argmax abstains 100% of held-out OOS; no floor needed",
        "qhat": 0.95,
        "qhat_note": "selective-risk on 160-row cal split (holdout_helpdesk 300 + oos_eval 20, halved): helpdesk_v2_mt \u2192 98.1% auto, ub95 1.89%, eval err 0%; helpdesk_pool_L-3_mean \u2192 93.8% auto, ub95 3.10%, eval err 0%. Both hosted aliases (helpdesk2, it_helpdesk) wired at qhat 0.95.",
        "mondrian": "per-class qhat \u03b1=0.05: v2 96.2% auto (worst-class 88%); pool head 90.6% but worst-class 0% \u2014 one route never reaches singletons under class-conditional coverage",
        "limitations": [
          "7-class taxonomy incl. oos \u2014 out-of-taxonomy queries correctly resolve to oos/escalate"
        ]
      },
      "yanked": null,
      "superseded_by": null,
      "yank_reason": null,
      "conformal": null,
      "license": "/tmp/hd_old/LICENSE.txt",
      "pricing": {
        "tier": "free"
      },
      "provenance": "Synthetic IT-helpdesk corpus (2000 rows, public) + 80 hand-written OOS negatives (facilities/HR-adjacent + generic). 7 routes incl explicit oos. Live-verified + parity-checked on anvil-serve-mixed.",
      "eval_sample": null,
      "runtime": "anvil-serve >= v0.3.0 (ANVIL_EXPERTS=<lane>:<head>, ANVIL_EXPERT_QHAT if provided)",
      "published_at": 1791701170
    },
    {
      "name": "link_safety",
      "version": "0.1.0",
      "file": "link_safety-0.1.0.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/link_safety-0.1.0.lanepack.tar.gz",
      "sha256": "debf0eadd2b2a57c3210fa8db9ea43b2eb6c78daf808c503269067e1d073e1bf",
      "head_sha256": null,
      "classes": 5,
      "taxonomy": [
        "safe",
        "link_unsafe",
        "link_adult",
        "link_gambling",
        "link_phishing"
      ],
      "status": "certified",
      "holdout_acc": null,
      "n_holdout": null,
      "oos": null,
      "live_verified": false,
      "operating": null,
      "yanked": null,
      "superseded_by": null,
      "yank_reason": null,
      "conformal": null,
      "license": "Apache-2.0",
      "pricing": {
        "tier": "free"
      },
      "provenance": "hand-written patterns + curated demo blocklist; runtime: anvil-serve rules lanes (machin-anvil 249de6a+)",
      "eval_sample": "eval_sample.jsonl",
      "runtime": "anvil-serve >= rules-lanes build",
      "published_at": 1792514065
    },
    {
      "name": "mod_policy",
      "version": "0.1.0",
      "file": "mod_policy-0.1.0.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/mod_policy-0.1.0.lanepack.tar.gz",
      "sha256": "6c16cb900b03a880bdc14156b620b628fe38bfc56dc082349dd21472456521c9",
      "head_sha256": null,
      "classes": 4,
      "taxonomy": [
        "publish",
        "review",
        "remove",
        "unchecked"
      ],
      "status": "certified",
      "holdout_acc": null,
      "n_holdout": null,
      "oos": null,
      "live_verified": false,
      "operating": null,
      "yanked": null,
      "superseded_by": null,
      "yank_reason": null,
      "conformal": null,
      "license": "Apache-2.0",
      "pricing": {
        "tier": "free"
      },
      "provenance": "hand-authored threshold policy on SightEngine-style categories (nudity/violence/drugs/self_harm/scam); runtime: anvil-serve rules lanes with score-input kinds (machin-anvil d2e2b36)",
      "eval_sample": "eval_sample.jsonl",
      "runtime": "anvil-serve >= score-policy build (POST /v1/route {state,expert,scores})",
      "published_at": 1792696581
    },
    {
      "name": "mod_text",
      "version": "0.1.0",
      "file": "mod_text-0.1.0.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/mod_text-0.1.0.lanepack.tar.gz",
      "sha256": "b7f3a8676be7931a11a615d250a49e9451ff5800bc94d5aea5641bb10c7f7c75",
      "head_sha256": null,
      "classes": 13,
      "taxonomy": [
        "clean",
        "pii_email",
        "pii_phone",
        "pii_iban",
        "pii_card",
        "pii_crypto",
        "profane",
        "scam",
        "suspicious",
        "link_unsafe",
        "link_adult",
        "link_gambling",
        "link_phishing"
      ],
      "status": "certified",
      "holdout_acc": null,
      "n_holdout": null,
      "oos": null,
      "live_verified": false,
      "operating": null,
      "yanked": null,
      "superseded_by": null,
      "yank_reason": null,
      "conformal": null,
      "license": "Apache-2.0",
      "pricing": {
        "tier": "free"
      },
      "provenance": "composite of pii_rules+scam_detect+profanity+link_safety packs \u2014 pattern coverage, not semantic moderation",
      "eval_sample": "eval_sample.jsonl",
      "runtime": "anvil-serve rules lanes",
      "published_at": 1792741998
    },
    {
      "name": "mod_text",
      "version": "0.2.0",
      "file": "mod_text-0.2.0.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/mod_text-0.2.0.lanepack.tar.gz",
      "sha256": "40f96d987167ed7aaca0a5e9bf8bededa889195d864a2a17e0f51f850c080e20",
      "head_sha256": null,
      "classes": 16,
      "taxonomy": [
        "clean",
        "self_harm",
        "extremist",
        "scam",
        "suspicious",
        "pii_email",
        "pii_phone",
        "pii_iban",
        "pii_card",
        "pii_crypto",
        "link_unsafe",
        "link_adult",
        "link_gambling",
        "link_phishing",
        "spam",
        "profane"
      ],
      "status": "certified",
      "holdout_acc": null,
      "n_holdout": null,
      "oos": null,
      "live_verified": false,
      "operating": null,
      "yanked": null,
      "superseded_by": null,
      "yank_reason": null,
      "conformal": null,
      "license": "Apache-2.0",
      "pricing": {
        "tier": "free"
      },
      "provenance": "composite of self_harm+extremism+scam_detect+pii_rules+link_safety+spam+profanity \u2014 deterministic pattern coverage",
      "eval_sample": "eval_sample.jsonl",
      "runtime": "anvil-serve rules lanes (phrases kind)",
      "published_at": 1792742554
    },
    {
      "name": "mod_text",
      "version": "0.3.0",
      "file": "mod_text-0.3.0.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/mod_text-0.3.0.lanepack.tar.gz",
      "sha256": "1de23754017c465784341a2e40c8162f5c5e0eaea5afe70c23e6fcc3a7659c53",
      "head_sha256": null,
      "classes": 17,
      "taxonomy": [
        "clean",
        "self_harm",
        "extremist",
        "scam",
        "suspicious",
        "harassment",
        "pii_email",
        "pii_phone",
        "pii_iban",
        "pii_card",
        "pii_crypto",
        "link_unsafe",
        "link_adult",
        "link_gambling",
        "link_phishing",
        "spam",
        "profane"
      ],
      "status": "certified",
      "holdout_acc": null,
      "n_holdout": null,
      "oos": null,
      "live_verified": false,
      "operating": null,
      "yanked": null,
      "superseded_by": null,
      "yank_reason": null,
      "conformal": null,
      "license": "Apache-2.0",
      "pricing": {
        "tier": "free"
      },
      "provenance": "composite of self_harm+extremism+scam_detect+harassment+pii_rules+link_safety+spam+profanity",
      "eval_sample": "eval_sample.jsonl",
      "runtime": "anvil-serve rules lanes",
      "published_at": 1792742911
    },
    {
      "name": "pii",
      "version": "0.1.0",
      "file": "pii-0.1.0.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/pii-0.1.0.lanepack.tar.gz",
      "sha256": "b61687a5de48082372def1bf64860b5c6f8ebfc80aa15b8f08e6b3413b70a3e0",
      "head_sha256": "1ad9e4e992656441afd0d1fd4b1feec0e162e134217628d3f38d30b6f7f5375e",
      "classes": 8,
      "taxonomy": [
        "none",
        "pii_email",
        "pii_phone",
        "pii_iban",
        "pii_card",
        "pii_ssn",
        "handoff",
        "oos"
      ],
      "status": "certified",
      "holdout_acc": null,
      "n_holdout": null,
      "oos": null,
      "live_verified": true,
      "operating": null,
      "yanked": null,
      "superseded_by": null,
      "yank_reason": null,
      "conformal": null,
      "license": "Apache-2.0",
      "pricing": {
        "tier": "free"
      },
      "provenance": "synthetic injection into civil_comments safe carriers + textdetox multilingual bases; obfuscated variants (at dot com, spelled digits); m7router3s384 frozen trunk",
      "eval_sample": "eval_sample.jsonl",
      "runtime": "anvil-serve >= v0.3.0 (ANVIL_EXPERTS=pii:<head>, ANVIL_EXPERT_QHAT if provided)",
      "published_at": 1792443979
    },
    {
      "name": "pii_rules",
      "version": "0.1.0",
      "file": "pii_rules-0.1.0.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/pii_rules-0.1.0.lanepack.tar.gz",
      "sha256": "27bf5293eb23db2bf01ccaccdc284ba7e2469344c1df7222ef9ec1212434417f",
      "head_sha256": null,
      "classes": 6,
      "taxonomy": [
        "none",
        "pii_email",
        "pii_phone",
        "pii_card",
        "pii_iban",
        "link_adult"
      ],
      "status": "certified",
      "holdout_acc": null,
      "n_holdout": null,
      "oos": null,
      "live_verified": false,
      "operating": null,
      "yanked": null,
      "superseded_by": null,
      "yank_reason": null,
      "conformal": null,
      "license": "Apache-2.0",
      "pricing": {
        "tier": "free"
      },
      "provenance": "hand-written rule spec + HaGeZi-style demo domain list; runtime: anvil-serve with rules-lane support (machin-anvil 63175e3)",
      "eval_sample": "eval_sample.jsonl",
      "runtime": "anvil-serve >= rules-lanes build (ANVIL_EXPERTS=pii_rules:pii_rules.rules.json)",
      "published_at": 1792448818
    },
    {
      "name": "profanity",
      "version": "0.1.0",
      "file": "profanity-0.1.0.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/profanity-0.1.0.lanepack.tar.gz",
      "sha256": "3d68dd51810cc19a6c6e87c627ea4632c0d02e7e93cecf4eecf6816f29aa795d",
      "head_sha256": null,
      "classes": 2,
      "taxonomy": [
        "clean",
        "profane"
      ],
      "status": "certified",
      "holdout_acc": null,
      "n_holdout": null,
      "oos": null,
      "live_verified": false,
      "operating": null,
      "yanked": null,
      "superseded_by": null,
      "yank_reason": null,
      "conformal": null,
      "license": "Apache-2.0",
      "pricing": {
        "tier": "free"
      },
      "provenance": "LDNOOBW public profanity list (en), single-token word matching \u2014 multiword list entries don't fire; anvil 'words' kind, machin-anvil ad68792+",
      "eval_sample": "eval_sample.jsonl",
      "runtime": "anvil-serve >= words-kind build",
      "published_at": 1792741329
    },
    {
      "name": "scam_detect",
      "version": "0.1.0",
      "file": "scam_detect-0.1.0.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/scam_detect-0.1.0.lanepack.tar.gz",
      "sha256": "b8cf0aa43d5a3d04e849fe7e00790d93e4f2ef775a67ec883892562ac83b8e5a",
      "head_sha256": null,
      "classes": 3,
      "taxonomy": [
        "clean",
        "suspicious",
        "scam"
      ],
      "status": "certified",
      "holdout_acc": null,
      "n_holdout": null,
      "oos": null,
      "live_verified": false,
      "operating": null,
      "yanked": null,
      "superseded_by": null,
      "yank_reason": null,
      "conformal": null,
      "license": "Apache-2.0",
      "pricing": {
        "tier": "free"
      },
      "provenance": "hand-authored patterns over classic fraud phrasing (419/advance-fee/impersonation formulas); honest scope: pattern detection, NOT semantic fraud analysis \u2014 novel scam phrasing misses",
      "eval_sample": "eval_sample.jsonl",
      "runtime": "anvil-serve rules lanes",
      "published_at": 1792741473
    },
    {
      "name": "self_harm",
      "version": "0.1.0",
      "file": "self_harm-0.1.0.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/self_harm-0.1.0.lanepack.tar.gz",
      "sha256": "6592193ede19d1b737909769d0efbc2246825e10d8d1cf490d2d5a5f267a2129",
      "head_sha256": null,
      "classes": 2,
      "taxonomy": [
        "clean",
        "self_harm"
      ],
      "status": "certified",
      "holdout_acc": null,
      "n_holdout": null,
      "oos": null,
      "live_verified": false,
      "operating": null,
      "yanked": null,
      "superseded_by": null,
      "yank_reason": null,
      "conformal": null,
      "license": "Apache-2.0",
      "pricing": {
        "tier": "free"
      },
      "provenance": "hand-authored anchored distress patterns; coverage is explicit phrasing \u2014 always a triage signal, never an autonomous verdict on someone's welfare",
      "eval_sample": "eval_sample.jsonl",
      "runtime": "anvil-serve rules lanes",
      "published_at": 1792742209
    },
    {
      "name": "spam",
      "version": "0.1.0",
      "file": "spam-0.1.0.lanepack.tar.gz",
      "url": "https://github.com/javimosch/mtlm-heads/raw/master/lanes/spam-0.1.0.lanepack.tar.gz",
      "sha256": "9f3e150eb21a1ad65c30462d92c097f3aa0634cc890cec5052852c67a9acb8a4",
      "head_sha256": null,
      "classes": 2,
      "taxonomy": [
        "clean",
        "spam"
      ],
      "status": "certified",
      "holdout_acc": null,
      "n_holdout": null,
      "oos": null,
      "live_verified": false,
      "operating": null,
      "yanked": null,
      "superseded_by": null,
      "yank_reason": null,
      "conformal": null,
      "license": "Apache-2.0",
      "pricing": {
        "tier": "free"
      },
      "provenance": "hand-authored spam patterns (pharma/casino/SEO/promo formulas) + classic spam phrase list",
      "eval_sample": "eval_sample.jsonl",
      "runtime": "anvil-serve rules lanes",
      "published_at": 1792742483
    }
  ]
}