Private CMF BANKING77 research release with measured evidence

.gitattributes

@@ -35,3 +35,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 cortiq-banking77.cmf filter=lfs diff=lfs merge=lfs -text
 training/seed.json filter=lfs diff=lfs merge=lfs -text
+evidence/banking77/local-test-repeat.jsonl filter=lfs diff=lfs merge=lfs -text

ARTICLE_RU.md

@@ -4,6 +4,8 @@
 
 Это не рассказ о «модели, победившей все LLM». Мы сравниваем обученного специалиста с универсальной моделью решений. Затем отдельно показываем, как добавлять навыки к другой, FCD-модели, не меняя её общее ядро. **Результат BANKING77 принадлежит affine-модели, а не FCD.**
 
+![Результаты BANKING77 и пакетного энкодирования](figures/banking77-results.png)
+
 ## Зачем вообще отдельная модель решений
 
 Агенту часто нужен не абзац текста, а одно действие: выбрать обработчик, определить категорию обращения, ответить да/нет, выставить оценку или отказаться от автоматического решения. Генерация JSON токен за токеном — не единственный способ сделать это.
@@ -66,6 +68,14 @@ Apple M4, 24 GiB RAM; прогретый энкодер, без кэша отв
 
 Decision-веса BANKING77 занимают **3,70 MiB**, но полный автономный CMF с энкодером — **468,70 MiB**. Ранее измеренный peak RSS полного CPU-eval — 2,12 GiB. В резидентном сервере дополнительно устранено удержание копии encoder-mapping для каждой топологии. Серверные RAM/VRAM и энергопотребление Jev неизвестны: сравнительного ресурсного превосходства над ним мы не заявляем.
 
+### Проверка настоящего Vulkan на NVIDIA
+
+На Linux-сервере с **RTX PRO 4000 Blackwell, 24 GiB** тот же affine scorer исполнен через Vulkan, без CPU/software fallback. Все 3080 предвычисленных embedding-векторов дали те же классы и те же решения порога: 2818 правильных, 2284 принятых, 2254 правильных среди принятых. Максимальное расхождение ошибки реконструкции с CPU — 1,43e-6.
+
+При batch=128: **2,190 мс Vulkan против 114,625 мс CPU** того же сервера, примерно 52×. При batch=1 GPU, наоборот, медленнее: 1,408 против 0,930 мс. Время включает отправку входа, submit, ожидание и чтение результата. Резидентные буферы scorer — 4 178 756 байт, **без энкодера**.
+
+На GPU одновременно работала чужая вычислительная задача; наблюдалась загрузка 94–100%. Мы её не останавливали. Эти цифры — воспроизводимый журнал работы под конкурирующей нагрузкой, а не предел производительности свободной карты. Энкодер на Vulkan **не портирован**, поэтому 52× нельзя переносить на обработку текста целиком.
+
 ## Вторая часть: добавить навык и не сломать старые
 
 Это **отдельный FCD-эксперимент**, не подмена affine-результатов. Формула реконструкции:
@@ -76,6 +86,8 @@ Decision-веса BANKING77 занимают **3,70 MiB**, но полный а
 
 Общие `U/S/V` заморожены. Новый навык получает собственный центр и вектор в свободном ортогональном подпространстве. У текущего ядра ранг 125 и ёмкость 24 вектора; после двух добавлений используются 10. Мы не обещаем бесконечное число ортогональных навыков в конечном пространстве. При исчерпании ёмкости текущий тренер отказывает, а не тайно переписывает ядро.
 
+Обучаемая добавка одного навыка — **24 коэффициента вектора + 384 координаты центра = 1632 байта F32-параметров**, плюс статистика и метаданные. Это не размер полного CMF: общий энкодер и ядро значительно больше, а текущий экспорт сохраняет полную новую версию файла.
+
 Важная неудача: первая новая топология перехватила **3 из 43** старых контрольных запросов, хотя старые ��еса не изменились. Заморозка весов сама по себе не доказывает отсутствие забывания.
 
 Исправление — индивидуальная граница допустимой ошибки реконструкции. Она выбирается **только по calibration**, без старых контролей и финального теста. Вне области поддержки новая топология исключается и из выбора, и из нормировки вероятностей. Поэтому она не понижает уверенность старого ответа одним своим присутствием.
@@ -113,3 +125,7 @@ Decision-веса BANKING77 занимают **3,70 MiB**, но полный а
 **Повторяемые решения исполняются локально; неизвестные не маскируются под уверенный ответ; новый навык допускается только после проверки того, что он не отнял старые.** Для конкретного приложения это может быть полезнее красивого, но недоказанного обещания заменить любую большую модель.
 
 Артефакты, протоколы, сырые результаты и ограничения должны публиковаться вместе. Linux/Vulkan-проверка и ссылка на закрытый HF-релиз приведены в сопровождающем отчёте; их показатели не подставляются задним числом в таблицу Apple M4.
+
+## Артефакт
+
+Закрытый исследовательский релиз BANKING77: [Hugging Face](https://huggingface.co/infosave/cortiq-decision-banking77-cmf-20260925). Репозиторий приватный; доступ имеет владелец и приглашённые им участники. Веса production-seed и производной FCD туда не загружены.

README.md

@@ -99,3 +99,24 @@ The separate private-production seed and its derived FCD weights, request histor
 SQLite registry, keys, SSH/WireGuard configuration, and web assets are **not included**.
 No Jev responses were used as training labels. See `LICENSES.md` and the accompanying
 Russian article for the distinct FCD experiment and exact scopes of each result.
+
+## Additional verified implementations
+
+* F32 Metal encoder: 3,080/3,080 same winners and same .85 gate as CPU.
+  On the same DEV batch=128, 671 vs 135 texts/s; single-request worker p50
+  6.14 vs 6.10 ms (no median single-request win). RSS after single-workload:
+  0.89 vs 1.66 GiB; not peak memory. See runtime-{metal,cpu}.json.
+* Linux / RTX PRO 4000 Blackwell / Vulkan scorer: all 3,080 winners and gates
+  agree, max reconstruction-error delta 1.43e-6. Batch=128 p50: 2.190 ms Vulkan
+  vs 114.625 ms CPU. This is scorer-only with precomputed embeddings, measured
+  under a pre-existing GPU job, not full-text speed or idle-card maximum.
+* Included isolated source builds successfully; supplied CMF loads and infers
+  offline; HTTP package smoke succeeds and serves no web UI.
+* The FCD experiment described in ARTICLE_RU.md is a different, local model:
+  its production-derived weights are deliberately not included in this repository.
+
+The working notebook/workstation had unrelated background applications. All scopes,
+client-validation incidents and limitations are retained in evidence, not removed
+in favor of the headline metrics. Article: [Russian version](ARTICLE_RU.md).
+
+Reproduce the full local test without paid API calls: `./reproduce.sh /path/to/new-output`. This evaluates the already-frozen model; it does not fit on test labels.

SHA256SUMS

@@ -1,6 +1,6 @@
-444d778aff686ab91b9679e6293e1b8d71e22426078e1aa2ae364b0405ef2f5b  ARTICLE_RU.md
+fe02cf4034aade2ac4498c23a5ea0b608851e97bdabfb712fc77ec4316a235b6  ARTICLE_RU.md
 3bcff101bcd473a82cf9c5f6e342e57940e5b8f7aa6323c8602e6e793095a8de  LICENSES.md
-11aae3a04fcfb7ea6ccbb0a0b08670c80ddf9015c86e8df52ffa103d118d8c73  README.md
+e8362bcda89c2347289fa3e823443316380d5a2a8388fa01ff6837f6c4a16b27  README.md
 1d1708069621ea6f2f8c0bf8ee120a74dcdc8b3a8c3590c658de90e59a3317b1  contract.json
 4cdaf9ec19054f17cd34d420a56f4c8bb0eed12c2429732e9baf902ef52303df  cortiq-banking77.cmf
 ceeb47cab7a611169bedbe38b62441b0c58794a05295eeff751550b7640708a8  evidence/banking77/HARNESS_INCIDENTS.md
@@ -13,23 +13,48 @@ f073ec51d07d50a1f83f10c57a0e33cb3f4274c4d9c1ce48e6892422192e6335  evidence/banki
 415863de9f46d3e315613a281895cbbe844c0be52fcaf719ad502ca015b4d6b3  evidence/banking77/data/train.jsonl
 f4a7d35e026ca65a7fae5e1687cc990d60568c6785d4af2391b17e0279fa4abf  evidence/banking77/dataset-source.json
 707e285b0fec5dd3cfba0245f6400abb9577b0d6278ba1eb65d6a44a4eb87761  evidence/banking77/freeze.json
+89b41298347244c22bbab1175ff076f8ac28b9f42e802f106262151353eafa8a  evidence/banking77/jev-primary-combined.jsonl
+178c6a08537a2990b07b8975fbeb1e6ac9b445ef645b8a0d60d3c371b48407a4  evidence/banking77/jev-probability-diagnostic.json
+00aa8ca8498067c5bd46ae5651f2cd46d901cf12ebdc08b1bf03cb3948669c4f  evidence/banking77/local-test-repeat.jsonl
 b6add684980296431b2c66492655d7d7fb5599e0b52329fe36c5d9d3d81df09a  evidence/banking77/selective-test.json
 6371f55dd882864bcf76771170a331cfbd1ba183cb3bb65a430b85fffa7a5fa2  evidence/release/banking-release-manifest.json
+2acd7929f2308170163e0ee72ef63873bdbeadbbcaca6dac2cfa4a6366c2fb5f  evidence/release/build-cpu.txt
+f1b6afa213c7eee891dce27267b16ddd6fcf878056108dfaa9516f916a9555f0  evidence/release/clippy.txt
 ae9ef4a2695c27618f08ea72925dec58abc0789b476f657d7310652e25ce698a  evidence/release/fcd-incremental-acceptance.json
+ae9ef4a2695c27618f08ea72925dec58abc0789b476f657d7310652e25ce698a  evidence/release/fcd-incremental-console.txt
 4e2ee635fd044eb0d68b1fb75fea954d77e82e2c05d68c2b7089ed3d0707ea1c  evidence/release/fcd-incremental-protocol.json
 2caa042b7a049ed59b79c0e03a0baa94cdaf1b9a22196b256bf720174cf65f53  evidence/release/fcd-second-calibration.json
 c347d66ab4b4793c367d0a9a3d0ea7febed49af4693d63f780555d200105fcff  evidence/release/fcd-support-calibration.json
+371eeeabdf2aa68f3718605838efa25dcf9e12fdf0c80fbc46b39cd6494664b7  evidence/release/isolated-build.txt
+09857b047666c28321d6415071f006a8597b1b1bc2a467932f30f930bdb34dca  evidence/release/isolated-infer.json
+261a5fd9564d61b41103af78f2bbd2badd93a17e5eab43c810c814a2daf6f494  evidence/release/lifecycle-console.txt
+bff23271753c5aa8fd11c1fca76529a4a2f201b36947e479cfc40c8b2d07b317  evidence/release/live-release-status.json
 bb1d37c6e80790dcf80139da3a5595c9bc319edcc456de0896907103b125ef89  evidence/release/local-fcd-install.json
+a55a14bb89e934fd5f7eed6b894002e64b36d66a99aaa424a0fba22ab237bad7  evidence/release/metal-acceptance-console.txt
 f41f3899ab6e1a8d0c050a38474964b1cb83bdf531ad808d4fea6e3e8dc25bda  evidence/release/metal-acceptance.json
 f223fd7fa446639056ff4e074318ab8a7db6d533ba835ed032aaa447fdc05931  evidence/release/near-duplicate-audit.json
+7cf5774fd4f5a52e8806485b9ac38b8071e5aeb9f0a2ef104a584c0154546576  evidence/release/near-duplicate-console.txt
+98eb816bc72db0eb3af6b37bd18188d6f24beb170c157ca7952ea3b96b27b573  evidence/release/python-tests.txt
+3e2cb2f6cb7976660f837f5584e5f2404f7735074358e098f4a7653735310b49  evidence/release/release-acceptance-console.txt
 3e2cb2f6cb7976660f837f5584e5f2404f7735074358e098f4a7653735310b49  evidence/release/release-acceptance.json
+90c1a473501762deda5d0ed136903a01c77d31672c32f184e8497b0ec73dffe9  evidence/release/runtime-cpu-console-final.txt
+5fa0a31a8fdb7d737ababf42fa8668ef3ee4d992cc2cd3ddfcf894ab46780c6e  evidence/release/runtime-cpu-console.txt
 90c1a473501762deda5d0ed136903a01c77d31672c32f184e8497b0ec73dffe9  evidence/release/runtime-cpu.json
+d6bffcb3f85bf0b2b316cd3e5915de8ba8ee2821055ce3978b9a493296d7028b  evidence/release/runtime-metal-console-final.txt
+2c3c2e12a6ad428ed6612648720c108fb9c089b2009d021ce8e1984bf9864b9c  evidence/release/runtime-metal-console.txt
 d6bffcb3f85bf0b2b316cd3e5915de8ba8ee2821055ce3978b9a493296d7028b  evidence/release/runtime-metal.json
+e32b46c846623aa4f8c949bce53bef11ecb3c48d63e9ad72b6e9c9770073cdda  evidence/release/rust-tests.txt
+e33c6b7db7a23b489d146c6d96cf2215e0492934d84e9d4e0d197893d288d1b1  evidence/release/vulkan-acceptance.json
+eb51be7c3acf817fdca4f86b7a901e540c7b375d6307e757e6b89d41a27aeabc  evidence/release/vulkan-check-macos.txt
+5f4281c8a18d22a95823a1eea4ce6f3dd16d116f5cbe94524bf4df8d8a0f349a  evidence/release/vulkan-environment.txt
+9448fca2db97cbd2f6e5a6b1a13c225cff511f959b3b4e20babeaef0201d8995  evidence/release/vulkan-linux-tests.txt
+3a6b20fc5c8943b30b43a0084f8d307c21d856c0684183e390b5d89f7f7382ae  evidence/release/vulkan-shader-test.txt
 14b401166825d22d9a2a3771becb2632018c21d6fd4ed370246e83d45a8ea984  figures/banking77-results.png
 0bf59d136536dea5dcf80ce96c5988e46522ce69f3ae817666c7f5d10ded6618  licenses/Apache-2.0.txt
 7e7170e3cebf88a9f60c7b8421418323c09304da1af4d5e90f4da1dc1c8a2661  licenses/BANKING77-CC-BY-4.0.txt
 41ac81b48406cffd95db849913238063dacf58419b15c7c7c56b9d3a5755608b  licenses/encoder-source.json
 50fdf7871d31dbc665f837cd6217c961378916e9d2ec40251da520b13cd211f6  policy.json
+4379fd02656334b6356cd042c773361b045320f9fbaf4199beaa3b6dc3a8f931  reproduce.sh
 9ff38954ec8573a74ad6a04f03d2dbd6cbe510255c2d9bc0b1e640c2f8b7a15a  requirements-metal.txt
 5a81cde5d4f409222788b0a0da88fff0acdb160f45ab8839d9661cff6aa10424  requirements-training.txt
 61aa8a323944fa1ab678ba76e20375cf84b0b329ff826fc32751fba1d5afd975  source/Cargo.toml
@@ -47,7 +72,7 @@ b2523ec287c0411488aede69337ba1a160b511ec800608d9a7798ba0063fd9c2  source/crates/
 18b7aad159777c636d867861c989214a38e5b743d49c026444bbcd3bdbeb8675  source/tools/cortiq-decision/Cargo.toml
 0df7eb993c25ef4abe7cf41893c803c50dc6b24c34d6bf7ef1d05032627ef01f  source/tools/cortiq-decision/README.md
 76d2a79c9bfc0a8965d1f146d8caa40862b7e3d88d97cd9988801a4fe88d796e  source/tools/cortiq-decision/analyze_banking77.py
-0243bb8070d373977ff87f08d17dea4b57920b5bee354f7fbc79d7404a24bcf8  source/tools/cortiq-decision/decision_server.py
+8ac17986a6929ecf3c82251fd55f98b025d2d0416354f826c8cf6bff870531d5  source/tools/cortiq-decision/decision_server.py
 5dd0d8f2d406983ff54dacca15a275c2f388f49704e9def14acaa71d703df400  source/tools/cortiq-decision/decisions_protocol.py
 4f99186a22e3f1ef502e250cf6951b2d3863a28a6e01a6602c7e5972e4833092  source/tools/cortiq-decision/demo.py
 91acf11d529b8b432bb50679711b5d6e17020f09efa1562b5596aad56dbf38e1  source/tools/cortiq-decision/lifecycle.py
@@ -61,13 +86,19 @@ d8837fdb40a1d8d359056eba2ed3857707a4ea47b2b22aa446b6cc687c600410  source/tools/c
 8be3ba24299c1c5a6ebc931dd71025a9e0c255ccce1fc8b610914c64a6193f41  source/tools/cortiq-decision/src/encoder.rs
 d01420fc5f26e43ec2a073bc4c19e606bc4386203b99cfc2ad385515845b174d  source/tools/cortiq-decision/src/fcd.rs
 040cf88f426a040970712829e7e52a9dfd576b3863d6ceb88e0047a9b98e0bde  source/tools/cortiq-decision/src/lib.rs
-95a782bb884f0e8f71b2a52b9f145ef8ff3f832d02a506f7c722d4e4556c4032  source/tools/cortiq-decision/src/main.rs
+8f014af587ad6a3c19bcfd7b7dc0b842ea1a1849d1251c1e299f136403f19905  source/tools/cortiq-decision/src/main.rs
 6afd52189ca7f66500eef8c03f646dd617acbe92afa62cd793afe30299c30a93  source/tools/cortiq-decision/src/metal.rs
 f4db81c80be5e0eb80b19b24b4fd2de0d593a480c64fde5bc4883b3571aa84c0  source/tools/cortiq-decision/src/packed.rs
 44072e0dd4a534237316b966587cd115d679fe35beb3da7579c7171abf362433  source/tools/cortiq-decision/src/policy.rs
 66b024cb5fe6d66a413b4e0f3470b8895fa3010663615fc5f8dbdfaee570d4a0  source/tools/cortiq-decision/src/rpc.rs
 ab5fb6d464ba1cada047a00d65466166ee65f8d5190461c4ebcac1cf71b68009  source/tools/cortiq-decision/src/service.rs
-8ccdf017d50fe4a7aec1a96ebaddf33d797343055de152d2211c9cad0d3621e8  source/tools/cortiq-decision/src/vulkan.rs
+6761a362a318d2dc3327deb9c6982a9e6c30fa3fc383fe5d5350e8735ab3cbc5  source/tools/cortiq-decision/src/vulkan.rs
+1bd8074f6193b4439aa49d0a8c16fe0c7d6792aa0cfbdcf99712420a9efef836  source/tools/cortiq-decision/tests/contracts.rs
+0fd5efd944f72c3388b860bf4cec434e0eece11a12a96ce3c8188c3d36c83215  source/tools/cortiq-decision/tests/test_decisions_protocol.py
+7a894d7a95a008c7eae11be2e86be8cc3a806cdc40afec5006daa57036b8fc4b  source/tools/cortiq-decision/tests/test_harness.py
+8ab49885f48745fe2314cb5d807fa7f78e9650ee938a4af25dc2d3c0fc1a58b3  source/tools/cortiq-decision/tests/test_server.py
+c2aa55bac6e0aebb33fbfd0bd6a4f04885270669eaf83f9ac513030bd94d1969  source/tools/cortiq-decision/tests/test_support_gate.py
+9aa106c14aebc6e3378be2d63365cd4c91deee5f61fa30965a1f352e63bee09f  source/tools/cortiq-decision/tests/test_training.py
 20a7bf5a24176112cc71f61fd80f9c23ee74b43300cfbeae390ac0bbb8a46c89  source/tools/cortiq-decision/train.py
 362769573bd26a4189027f2b9ea27eaae680159848e5eba1c28b4c6226654d1e  training/seed.json
 7c7f64dc00cc5454792e643e21c4b4ed1acac74687015452cc3c1faf45c69758  training/training.json

evidence/banking77/jev-probability-diagnostic.json

@@ -0,0 +1,200 @@
+[
+  {
+    "index": 148,
+    "status": 200,
+    "model": "typesafe/jev-1.13-20260917",
+    "answer": {
+      "type": "choice",
+      "choice": "exchange_charge",
+      "probabilities": {
+        "card_linking": 0,
+        "lost_or_stolen_card": 0,
+        "get_disposable_virtual_card": 0,
+        "card_payment_not_recognised": 0,
+        "extra_charge_on_statement": 0.01,
+        "transfer_not_received_by_recipient": 0,
+        "request_refund": 0,
+        "Refund_not_showing_up": 0,
+        "terminate_account": 0,
+        "top_up_failed": 0,
+        "pending_cash_withdrawal": 0,
+        "declined_transfer": 0,
+        "change_pin": 0,
+        "verify_source_of_funds": 0,
+        "beneficiary_not_allowed": 0,
+        "age_limit": 0,
+        "top_up_limits": 0,
+        "automatic_top_up": 0,
+        "pending_card_payment": 0,
+        "cash_withdrawal_not_recognised": 0,
+        "direct_debit_payment_not_recognised": 0,
+        "exchange_rate": 0.05,
+        "country_support": 0,
+        "transfer_into_account": 0,
+        "order_physical_card": 0,
+        "contactless_not_working": 0,
+        "topping_up_by_card": 0,
+        "failed_transfer": 0,
+        "top_up_reverted": 0,
+        "card_about_to_expire": 0,
+        "reverted_card_payment?": 0,
+        "supported_cards_and_currencies": 0,
+        "transfer_timing": 0,
+        "pending_top_up": 0,
+        "verify_top_up": 0,
+        "card_acceptance": 0,
+        "fiat_currency_support": 0,
+        "balance_not_updated_after_bank_transfer": 0,
+        "card_payment_wrong_exchange_rate": 0.09,
+        "receiving_money": 0,
+        "getting_spare_card": 0,
+        "verify_my_identity": 0,
+        "exchange_via_app": 0.01,
+        "virtual_card_not_working": 0,
+        "transfer_fee_charged": 0,
+        "passcode_forgotten": 0,
+        "balance_not_updated_after_cheque_or_cash_deposit": 0,
+        "top_up_by_cash_or_cheque": 0,
+        "wrong_amount_of_cash_received": 0,
+        "apple_pay_or_google_pay": 0,
+        "lost_or_stolen_phone": 0,
+        "card_payment_fee_charged": 0,
+        "pending_transfer": 0,
+        "top_up_by_bank_transfer_charge": 0,
+        "wrong_exchange_rate_for_cash_withdrawal": 0.01,
+        "unable_to_verify_identity": 0,
+        "declined_card_payment": 0,
+        "exchange_charge": 0.83,
+        "activate_my_card": 0,
+        "card_arrival": 0,
+        "top_up_by_card_charge": 0,
+        "pin_blocked": 0,
+        "transaction_charged_twice": 0,
+        "compromised_card": 0,
+        "declined_cash_withdrawal": 0,
+        "card_delivery_estimate": 0,
+        "card_not_working": 0,
+        "getting_virtual_card": 0,
+        "get_physical_card": 0,
+        "card_swallowed": 0,
+        "edit_personal_details": 0,
+        "why_verify_identity": 0,
+        "cancel_transfer": 0,
+        "disposable_card_limits": 0,
+        "atm_support": 0,
+        "cash_withdrawal_charge": 0,
+        "visa_or_mastercard": 0
+      },
+      "confidence": 0.82
+    },
+    "usage": {
+      "input_tokens": 4387,
+      "output_tokens": 826,
+      "cost": 0.000184254
+    },
+    "probability_sum": 1.0,
+    "probability_min": 0,
+    "probability_max": 0.83,
+    "n_probabilities": 77,
+    "not_primary": "diagnostic retry, never replaces initial failure"
+  },
+  {
+    "index": 246,
+    "status": 200,
+    "model": "typesafe/jev-1.13-20260917",
+    "answer": {
+      "type": "choice",
+      "choice": "receiving_money",
+      "probabilities": {
+        "topping_up_by_card": 0,
+        "disposable_card_limits": 0,
+        "get_physical_card": 0,
+        "card_delivery_estimate": 0,
+        "getting_virtual_card": 0,
+        "transaction_charged_twice": 0,
+        "top_up_failed": 0,
+        "get_disposable_virtual_card": 0,
+        "verify_source_of_funds": 0,
+        "balance_not_updated_after_cheque_or_cash_deposit": 0,
+        "why_verify_identity": 0,
+        "country_support": 0,
+        "transfer_into_account": 0.06,
+        "beneficiary_not_allowed": 0,
+        "reverted_card_payment?": 0,
+        "cash_withdrawal_not_recognised": 0,
+        "lost_or_stolen_card": 0,
+        "card_payment_not_recognised": 0,
+        "cash_withdrawal_charge": 0,
+        "contactless_not_working": 0,
+        "pending_top_up": 0.01,
+        "card_acceptance": 0,
+        "pin_blocked": 0,
+        "declined_cash_withdrawal": 0,
+        "fiat_currency_support": 0,
+        "transfer_timing": 0,
+        "card_not_working": 0,
+        "lost_or_stolen_phone": 0,
+        "order_physical_card": 0,
+        "edit_personal_details": 0,
+        "failed_transfer": 0.01,
+        "atm_support": 0,
+        "change_pin": 0,
+        "receiving_money": 0.55,
+        "transfer_fee_charged": 0,
+        "Refund_not_showing_up": 0,
+        "apple_pay_or_google_pay": 0,
+        "card_linking": 0,
+        "wrong_amount_of_cash_received": 0,
+        "balance_not_updated_after_bank_transfer": 0,
+        "transfer_not_received_by_recipient": 0.01,
+        "activate_my_card": 0.01,
+        "pending_transfer": 0.24000000000000002,
+        "top_up_reverted": 0,
+        "pending_cash_withdrawal": 0,
+        "getting_spare_card": 0,
+        "direct_debit_payment_not_recognised": 0,
+        "unable_to_verify_identity": 0,
+        "verify_my_identity": 0,
+        "compromised_card": 0,
+        "extra_charge_on_statement": 0,
+        "automatic_top_up": 0,
+        "cancel_transfer": 0,
+        "age_limit": 0,
+        "virtual_card_not_working": 0,
+        "card_swallowed": 0,
+        "card_about_to_expire": 0,
+        "exchange_rate": 0,
+        "top_up_by_cash_or_cheque": 0,
+        "top_up_by_bank_transfer_charge": 0,
+        "card_arrival": 0,
+        "terminate_account": 0,
+        "card_payment_fee_charged": 0,
+        "supported_cards_and_currencies": 0,
+        "exchange_via_app": 0,
+        "top_up_by_card_charge": 0,
+        "wrong_exchange_rate_for_cash_withdrawal": 0,
+        "exchange_charge": 0,
+        "declined_transfer": 0.09999999999999999,
+        "request_refund": 0,
+        "pending_card_payment": 0,
+        "declined_card_payment": 0,
+        "visa_or_mastercard": 0,
+        "verify_top_up": 0,
+        "passcode_forgotten": 0,
+        "card_payment_wrong_exchange_rate": 0,
+        "top_up_limits": 0
+      },
+      "confidence": 0.53
+    },
+    "usage": {
+      "input_tokens": 4384,
+      "output_tokens": 827,
+      "cost": 0.000184128
+    },
+    "probability_sum": 0.99,
+    "probability_min": 0,
+    "probability_max": 0.55,
+    "n_probabilities": 77,
+    "not_primary": "diagnostic retry, never replaces initial failure"
+  }
+]

evidence/banking77/local-test-repeat.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:00aa8ca8498067c5bd46ae5651f2cd46d901cf12ebdc08b1bf03cb3948669c4f
+size 38296470

evidence/release/build-cpu.txt

@@ -0,0 +1,4 @@
+   Compiling cortiq-decision v0.1.0 (/Users/oleg/dev/cmfpublic/tools/cortiq-decision)
+    Finished `release` profile [optimized] target(s) in 20.34s
+warning: the following packages contain code that will be rejected by a future version of Rust: block v0.1.6
+note: to see what the problems were, use the option `--future-incompat-report`, or run `cargo report future-incompatibilities --id 1`

evidence/release/clippy.txt

@@ -0,0 +1,11 @@
+    Checking naga v0.19.2
+    Checking metal v0.27.0
+    Checking metal v0.29.0
+    Checking cortiq-core v0.6.8 (/Users/oleg/dev/cmfpublic/crates/cortiq-core)
+    Checking wgpu-hal v0.19.5
+    Checking wgpu-core v0.19.4
+    Checking wgpu v0.19.4
+    Checking cortiq-decision v0.1.0 (/Users/oleg/dev/cmfpublic/tools/cortiq-decision)
+    Finished `dev` profile [unoptimized + debuginfo] target(s) in 8.24s
+warning: the following packages contain code that will be rejected by a future version of Rust: block v0.1.6
+note: to see what the problems were, use the option `--future-incompat-report`, or run `cargo report future-incompatibilities --id 1`

evidence/release/fcd-incremental-console.txt

@@ -0,0 +1,33 @@
+{
+  "pass": true,
+  "scope": "two incremental topologies; finite held-out tests, not universal no-forgetting theorem",
+  "core_and_all_old_tasks_unchanged": true,
+  "core_sha256": "3c49b40c29c39dbb6fa5862bfb0d63e7345a0bd38fda144160b162a56f65d3f8",
+  "old_controls": 43,
+  "old_winner_or_abstention_changes": [],
+  "max_old_confidence_delta": 0.0,
+  "independent_new_task_test_n": 3080,
+  "new_skills": {
+    "lost_or_stolen_card": {
+      "tp": 16,
+      "fp": 0,
+      "test_positives": 40,
+      "precision": 1.0,
+      "recall": 0.4
+    },
+    "change_pin": {
+      "tp": 25,
+      "fp": 2,
+      "test_positives": 40,
+      "precision": 0.9259259259259259,
+      "recall": 0.625
+    }
+  },
+  "first_skill_regressions_after_second": [],
+  "mean_roundtrip_score_ms": 0.36394813555194844,
+  "task_capacity": 24,
+  "active_tasks": 10,
+  "artifact_sha256": "4beef0e577d29ef52fc0941b7813286aadfe0afe49959d927350988368091e4f",
+  "source_embeddings": "Metal F32; same-representation validation reported separately",
+  "production_route_replaced": false
+}

evidence/release/isolated-build.txt

@@ -0,0 +1,242 @@
+   Compiling proc-macro2 v1.0.107
+   Compiling quote v1.0.47
+   Compiling unicode-ident v1.0.26
+   Compiling libc v0.2.189
+   Compiling version_check v0.9.5
+   Compiling autocfg v1.5.1
+   Compiling cfg-if v1.0.5
+   Compiling stable_deref_trait v1.2.1
+   Compiling find-msvc-tools v0.1.14
+   Compiling shlex v2.0.1
+   Compiling memchr v2.8.3
+   Compiling crossbeam-utils v0.8.23
+   Compiling crossbeam-epoch v0.9.21
+   Compiling cc v1.5.0
+   Compiling writeable v0.6.4
+   Compiling generic-array v0.14.7
+   Compiling num-traits v0.2.19
+   Compiling crossbeam-deque v0.8.8
+   Compiling litemap v0.8.3
+   Compiling serde_core v1.0.229
+   Compiling zerocopy v0.8.58
+   Compiling utf8_iter v1.0.4
+   Compiling icu_normalizer_data v2.3.0
+   Compiling rayon-core v1.13.0
+   Compiling icu_properties_data v2.3.0
+   Compiling syn v3.0.6
+   Compiling syn v2.0.119
+   Compiling zeroize v1.9.0
+   Compiling ring v0.17.14
+   Compiling num-integer v0.1.47
+   Compiling rustls-pki-types v1.15.1
+   Compiling typenum v1.20.1
+   Compiling either v1.18.0
+   Compiling serde v1.0.229
+   Compiling pkg-config v0.3.34
+   Compiling smallvec v1.16.2
+   Compiling simd-adler32 v0.3.10
+   Compiling paste v1.0.15
+   Compiling getrandom v0.2.17
+   Compiling thiserror v2.0.21
+   Compiling once_cell v1.21.4
+   Compiling rustix v1.1.5
+   Compiling untrusted v0.9.0
+   Compiling rayon v1.12.0
+   Compiling errno v0.3.14
+   Compiling ident_case v1.0.1
+   Compiling strsim v0.11.1
+   Compiling log v0.4.34
+   Compiling rustls v0.23.45
+   Compiling percent-encoding v2.3.2
+   Compiling bitflags v2.13.2
+   Compiling getrandom v0.3.4
+   Compiling fnv v1.0.7
+   Compiling crc32fast v1.5.2
+   Compiling adler2 v2.0.1
+   Compiling anyhow v1.0.104
+   Compiling form_urlencoded v1.2.2
+   Compiling crypto-common v0.1.7
+   Compiling block-buffer v0.10.4
+   Compiling webpki-roots v1.0.9
+   Compiling num-bigint v0.4.8
+   Compiling arrayvec v0.7.8
+   Compiling subtle v2.6.1
+   Compiling byteorder v1.5.0
+   Compiling darling_core v0.20.11
+   Compiling miniz_oxide v0.9.1
+   Compiling socks v0.3.4
+   Compiling synstructure v0.14.0
+   Compiling xattr v1.6.1
+   Compiling as-slice v0.2.1
+   Compiling webpki-roots v0.26.11
+   Compiling digest v0.10.7
+   Compiling itertools v0.14.0
+   Compiling filetime v0.2.29
+   Compiling num-rational v0.4.2
+   Compiling cpufeatures v0.2.17
+   Compiling base64 v0.22.1
+   Compiling zmij v1.0.23
+   Compiling av-scenechange v0.14.1
+   Compiling libm v0.2.16
+   Compiling built v0.8.1
+   Compiling rustversion v1.0.23
+   Compiling zerocopy-derive v0.8.58
+   Compiling equator-macro v0.4.2
+   Compiling zerofrom-derive v0.1.8
+   Compiling yoke-derive v0.8.3
+   Compiling zerovec-derive v0.11.6
+   Compiling displaydoc v0.2.7
+   Compiling bytemuck_derive v1.12.1
+   Compiling serde_derive v1.0.229
+   Compiling equator v0.4.2
+   Compiling thiserror-impl v2.0.21
+   Compiling rustls-webpki v0.103.15
+   Compiling darling_macro v0.20.11
+   Compiling rav1e v0.8.1
+   Compiling zerofrom v0.1.8
+   Compiling aligned-vec v0.6.4
+   Compiling v_frame v0.3.9
+   Compiling arg_enum_proc_macro v0.3.4
+   Compiling darling v0.20.11
+   Compiling profiling-procmacros v1.0.18
+   Compiling yoke v0.8.3
+   Compiling bytemuck v1.25.2
+   Compiling sha2 v0.10.9
+   Compiling tar v0.4.46
+   Compiling zerovec v0.11.8
+   Compiling zerotrie v0.2.5
+   Compiling flate2 v1.1.10
+   Compiling aligned v0.4.3
+   Compiling num-complex v0.4.6
+   Compiling onig_sys v69.9.3
+   Compiling no_std_io2 v0.9.4
+   Compiling nom v8.0.0
+   Compiling matrixmultiply v0.3.11
+   Compiling pulp v0.22.3
+   Compiling quick-error v2.0.1
+   Compiling tinystr v0.8.4
+   Compiling potential_utf v0.1.6
+   Compiling y4m v0.8.0
+   Compiling icu_collections v2.3.0
+   Compiling icu_locale_core v2.3.0
+   Compiling pastey v0.1.1
+   Compiling itoa v1.0.18
+   Compiling serde_json v1.0.151
+   Compiling icu_provider v2.3.1
+   Compiling bitstream-io v4.10.0
+   Compiling icu_properties v2.3.0
+   Compiling icu_normalizer v2.3.0
+   Compiling derive_builder_core v0.20.2
+   Compiling profiling v1.0.18
+   Compiling num-derive v0.4.2
+   Compiling tracing-attributes v0.1.31
+   Compiling idna_adapter v1.2.2
+   Compiling rand_core v0.9.5
+   Compiling idna v1.1.0
+   Compiling maybe-rayon v0.1.1
+   Compiling miniz_oxide v0.8.9
+   Compiling tracing-core v0.1.36
+   Compiling av1-grain v0.2.5
+   Compiling url v2.5.8
+   Compiling ppv-lite86 v0.2.21
+   Compiling half v2.7.1
+   Compiling simd_helpers v0.1.0
+   Compiling ureq v2.12.1
+   Compiling aho-corasick v1.1.5
+   Compiling ahash v0.8.12
+   Compiling pulp-wasm-simd-flag v0.1.1
+   Compiling reborrow v0.5.5
+   Compiling rawpointer v0.2.1
+   Compiling weezl v0.1.12
+   Compiling new_debug_unreachable v1.0.6
+   Compiling noop_proc_macro v0.3.0
+   Compiling regex-syntax v0.8.11
+   Compiling minimal-lexical v0.2.1
+   Compiling zune-core v0.5.3
+   Compiling esaxx-rs v0.1.10
+   Compiling pin-project-lite v0.2.17
+   Compiling utf8parse v0.2.2
+   Compiling imgref v1.12.3
+   Compiling zune-jpeg v0.5.15
+   Compiling loop9 v0.1.5
+   Compiling anstyle-parse v1.0.0
+   Compiling tracing v0.1.44
+   Compiling nom v7.1.3
+   Compiling ort-sys v2.0.0-rc.9
+   Compiling rand_chacha v0.9.0
+   Compiling derive_builder_macro v0.20.2
+   Compiling regex-automata v0.4.18
+   Compiling castaway v0.2.4
+   Compiling monostate-impl v0.1.18
+   Compiling avif-serialize v0.8.9
+   Compiling zune-inflate v0.2.54
+   Compiling fdeflate v0.3.7
+   Compiling pxfm v0.1.30
+   Compiling getrandom v0.4.3
+   Compiling rgb v0.8.53
+   Compiling colorchoice v1.0.5
+   Compiling ryu v1.0.23
+   Compiling bit_field v0.10.3
+   Compiling anstyle v1.0.14
+   Compiling anstyle-query v1.1.5
+   Compiling byteorder-lite v0.1.0
+   Compiling color_quant v1.1.0
+   Compiling core-foundation-sys v0.8.7
+   Compiling unicode-segmentation v1.13.3
+   Compiling lebe v0.5.3
+   Compiling fax v0.2.7
+   Compiling is_terminal_polyfill v1.70.2
+   Compiling macro_rules_attribute-proc_macro v0.2.3
+   Compiling static_assertions v1.1.0
+   Compiling base64 v0.13.1
+   Compiling pastey v0.2.3
+   Compiling spm_precompiled v0.1.4
+   Compiling regex v1.13.1
+   Compiling compact_str v0.9.1
+   Compiling macro_rules_attribute v0.2.3
+   Compiling anstream v1.0.0
+   Compiling moxcms v0.8.1
+   Compiling tiff v0.11.3
+   Compiling core-foundation v0.9.4
+   Compiling image-webp v0.2.4
+   Compiling gif v0.14.2
+   Compiling onig v6.5.3
+   Compiling exr v1.74.2
+   Compiling png v0.18.1
+   Compiling ravif v0.13.0
+   Compiling monostate v0.1.18
+   Compiling derive_builder v0.20.2
+   Compiling ndarray v0.16.1
+   Compiling rand v0.9.5
+   Compiling dary_heap v0.3.9
+   Compiling unicode-normalization-alignments v0.1.12
+   Compiling qoi v0.4.1
+   Compiling rayon-cond v0.4.0
+   Compiling foreign-types-macros v0.2.4
+   Compiling malloc_buf v0.0.6
+   Compiling bitflags v1.3.2
+   Compiling foreign-types-shared v0.3.1
+   Compiling unicode_categories v0.1.1
+   Compiling clap_lex v1.1.1
+   Compiling heck v0.5.0
+   Compiling clap_builder v4.6.7
+   Compiling tokenizers v0.21.4
+   Compiling foreign-types v0.5.0
+   Compiling clap_derive v4.6.7
+   Compiling core-graphics-types v0.1.3
+   Compiling objc v0.2.7
+   Compiling image v0.25.10
+   Compiling bincode v1.3.3
+   Compiling memmap2 v0.9.11
+   Compiling block v0.1.6
+   Compiling fastrand v2.5.0
+   Compiling metal v0.29.0
+   Compiling tempfile v3.27.0
+   Compiling cortiq-core v0.6.8 (/Users/oleg/dev/cmfpublic/artifacts/hf-private-cortiq-banking77-20260925/source/crates/cortiq-core)
+   Compiling clap v4.6.7
+   Compiling ort v2.0.0-rc.9
+   Compiling fastembed v4.9.1
+   Compiling cortiq-decision v0.1.0 (/Users/oleg/dev/cmfpublic/artifacts/hf-private-cortiq-banking77-20260925/source/tools/cortiq-decision)
+    Finished `release` profile [optimized] target(s) in 53.59s
+warning: the following packages contain code that will be rejected by a future version of Rust: block v0.1.6
+note: to see what the problems were, use the option `--future-incompat-report`, or run `cargo report future-incompatibilities --id 1`

evidence/release/isolated-infer.json

@@ -0,0 +1,563 @@
+{
+  "decision": {
+    "action": "local",
+    "choice": "change_pin",
+    "confidence": 0.9896446466445923,
+    "kind": "choice",
+    "reason": "validated_in_domain_confidence_gate",
+    "source": "cmf"
+  },
+  "decision_us": 341.625,
+  "encode_us": 3129.75,
+  "prediction": {
+    "confidence": 0.9896446466445923,
+    "is_novel": false,
+    "margin": 0.08189558982849121,
+    "novelty_score": 0.2757600247859955,
+    "raw_confidence": 0.9915828108787537,
+    "scores": [
+      {
+        "probability": 0.9896446466445923,
+        "reconstruction_error": 0.008488600142300129,
+        "score": 0.9915828108787537,
+        "task_id": 23,
+        "task_label": "change_pin"
+      },
+      {
+        "probability": 0.010215039364993572,
+        "reconstruction_error": 0.09927887469530106,
+        "score": 0.9096872210502625,
+        "task_id": 51,
+        "task_label": "pin_blocked"
+      },
+      {
+        "probability": 0.00014025272685103118,
+        "reconstruction_error": 0.18440523743629456,
+        "score": 0.8443056344985962,
+        "task_id": 40,
+        "task_label": "get_physical_card"
+      },
+      {
+        "probability": 7.914153776766852e-10,
+        "reconstruction_error": 0.424312561750412,
+        "score": 0.7020930647850037,
+        "task_id": 28,
+        "task_label": "declined_cash_withdrawal"
+      },
+      {
+        "probability": 7.645493682595372e-10,
+        "reconstruction_error": 0.4249981939792633,
+        "score": 0.7017552852630615,
+        "task_id": 16,
+        "task_label": "card_not_working"
+      },
+      {
+        "probability": 3.1089619268766244e-13,
+        "reconstruction_error": 0.5799899697303772,
+        "score": 0.6329154372215271,
+        "task_id": 2,
+        "task_label": "activate_my_card"
+      },
+      {
+        "probability": 9.239073536174841e-14,
+        "reconstruction_error": 0.6040783524513245,
+        "score": 0.6234109401702881,
+        "task_id": 70,
+        "task_label": "verify_my_identity"
+      },
+      {
+        "probability": 9.1371842817628e-14,
+        "reconstruction_error": 0.60429847240448,
+        "score": 0.6233254075050354,
+        "task_id": 11,
+        "task_label": "card_about_to_expire"
+      },
+      {
+        "probability": 7.583969801987223e-14,
+        "reconstruction_error": 0.6079971194267273,
+        "score": 0.6218916177749634,
+        "task_id": 44,
+        "task_label": "lost_or_stolen_phone"
+      },
+      {
+        "probability": 3.0816024238671316e-14,
+        "reconstruction_error": 0.6258750557899475,
+        "score": 0.6150534152984619,
+        "task_id": 15,
+        "task_label": "card_linking"
+      },
+      {
+        "probability": 2.236265405979824e-14,
+        "reconstruction_error": 0.6322402954101562,
+        "score": 0.6126548647880554,
+        "task_id": 46,
+        "task_label": "passcode_forgotten"
+      },
+      {
+        "probability": 9.084127662606096e-15,
+        "reconstruction_error": 0.650123655796051,
+        "score": 0.6060152053833008,
+        "task_id": 63,
+        "task_label": "topping_up_by_card"
+      },
+      {
+        "probability": 5.97310397283256e-15,
+        "reconstruction_error": 0.6584466099739075,
+        "score": 0.6029739379882812,
+        "task_id": 5,
+        "task_label": "atm_support"
+      },
+      {
+        "probability": 3.598285745428677e-15,
+        "reconstruction_error": 0.6685075163841248,
+        "score": 0.5993379950523376,
+        "task_id": 10,
+        "task_label": "cancel_transfer"
+      },
+      {
+        "probability": 3.241863810344724e-15,
+        "reconstruction_error": 0.6705782413482666,
+        "score": 0.598595142364502,
+        "task_id": 22,
+        "task_label": "cash_withdrawal_not_recognised"
+      },
+      {
+        "probability": 2.1305966058538395e-15,
+        "reconstruction_error": 0.6789107918739319,
+        "score": 0.5956242680549622,
+        "task_id": 74,
+        "task_label": "visa_or_mastercard"
+      },
+      {
+        "probability": 2.1219987979227348e-15,
+        "reconstruction_error": 0.6789910793304443,
+        "score": 0.5955957770347595,
+        "task_id": 32,
+        "task_label": "edit_personal_details"
+      },
+      {
+        "probability": 1.3099213122148433e-15,
+        "reconstruction_error": 0.6885672211647034,
+        "score": 0.5922181010246277,
+        "task_id": 58,
+        "task_label": "top_up_by_card_charge"
+      },
+      {
+        "probability": 9.428964304553978e-16,
+        "reconstruction_error": 0.6950936913490295,
+        "score": 0.5899379253387451,
+        "task_id": 30,
+        "task_label": "direct_debit_payment_not_recognised"
+      },
+      {
+        "probability": 8.49311624701092e-16,
+        "reconstruction_error": 0.69716876745224,
+        "score": 0.5892165899276733,
+        "task_id": 4,
+        "task_label": "apple_pay_or_google_pay"
+      },
+      {
+        "probability": 7.881772334912497e-16,
+        "reconstruction_error": 0.6986517310142517,
+        "score": 0.5887022018432617,
+        "task_id": 66,
+        "task_label": "transfer_into_account"
+      },
+      {
+        "probability": 7.268400308300992e-16,
+        "reconstruction_error": 0.7002600431442261,
+        "score": 0.5881453156471252,
+        "task_id": 57,
+        "task_label": "top_up_by_bank_transfer_charge"
+      },
+      {
+        "probability": 6.790677431818711e-16,
+        "reconstruction_error": 0.7016096711158752,
+        "score": 0.587678849697113,
+        "task_id": 20,
+        "task_label": "card_swallowed"
+      },
+      {
+        "probability": 6.648165197234359e-16,
+        "reconstruction_error": 0.7020307183265686,
+        "score": 0.5875334739685059,
+        "task_id": 64,
+        "task_label": "transaction_charged_twice"
+      },
+      {
+        "probability": 6.24594524948355e-16,
+        "reconstruction_error": 0.7032696008682251,
+        "score": 0.5871061086654663,
+        "task_id": 13,
+        "task_label": "card_arrival"
+      },
+      {
+        "probability": 6.086752289189801e-16,
+        "reconstruction_error": 0.7037821412086487,
+        "score": 0.5869295001029968,
+        "task_id": 72,
+        "task_label": "verify_top_up"
+      },
+      {
+        "probability": 5.222506573656109e-16,
+        "reconstruction_error": 0.7068220973014832,
+        "score": 0.5858840942382812,
+        "task_id": 25,
+        "task_label": "contactless_not_working"
+      },
+      {
+        "probability": 5.214762045540244e-16,
+        "reconstruction_error": 0.7068515419960022,
+        "score": 0.585874080657959,
+        "task_id": 24,
+        "task_label": "compromised_card"
+      },
+      {
+        "probability": 3.9380663418352797e-16,
+        "reconstruction_error": 0.712425947189331,
+        "score": 0.583966851234436,
+        "task_id": 49,
+        "task_label": "pending_top_up"
+      },
+      {
+        "probability": 3.234709770072214e-16,
+        "reconstruction_error": 0.7163317203521729,
+        "score": 0.5826379656791687,
+        "task_id": 60,
+        "task_label": "top_up_failed"
+      },
+      {
+        "probability": 3.193745403858424e-16,
+        "reconstruction_error": 0.7165846824645996,
+        "score": 0.5825520753860474,
+        "task_id": 55,
+        "task_label": "supported_cards_and_currencies"
+      },
+      {
+        "probability": 3.038462030011845e-16,
+        "reconstruction_error": 0.7175741791725159,
+        "score": 0.5822165012359619,
+        "task_id": 26,
+        "task_label": "country_support"
+      },
+      {
+        "probability": 1.8706714699177342e-16,
+        "reconstruction_error": 0.7272031903266907,
+        "score": 0.5789706707000732,
+        "task_id": 9,
+        "task_label": "beneficiary_not_allowed"
+      },
+      {
+        "probability": 1.7863763391937624e-16,
+        "reconstruction_error": 0.7281184792518616,
+        "score": 0.5786640644073486,
+        "task_id": 29,
+        "task_label": "declined_transfer"
+      },
+      {
+        "probability": 1.26695283957807e-16,
+        "reconstruction_error": 0.73493891954422,
+        "score": 0.5763891935348511,
+        "task_id": 59,
+        "task_label": "top_up_by_cash_or_cheque"
+      },
+      {
+        "probability": 1.2314581886211745e-16,
+        "reconstruction_error": 0.735503077507019,
+        "score": 0.5762017965316772,
+        "task_id": 45,
+        "task_label": "order_physical_card"
+      },
+      {
+        "probability": 9.471752041535631e-17,
+        "reconstruction_error": 0.7407134771347046,
+        "score": 0.5744770765304565,
+        "task_id": 35,
+        "task_label": "exchange_via_app"
+      },
+      {
+        "probability": 8.866139366116567e-17,
+        "reconstruction_error": 0.7420251369476318,
+        "score": 0.5740445256233215,
+        "task_id": 38,
+        "task_label": "fiat_currency_support"
+      },
+      {
+        "probability": 8.101340083150859e-17,
+        "reconstruction_error": 0.7438158988952637,
+        "score": 0.5734550356864929,
+        "task_id": 12,
+        "task_label": "card_acceptance"
+      },
+      {
+        "probability": 7.162347151093323e-17,
+        "reconstruction_error": 0.746261477470398,
+        "score": 0.5726519227027893,
+        "task_id": 41,
+        "task_label": "getting_spare_card"
+      },
+      {
+        "probability": 6.075173745847528e-17,
+        "reconstruction_error": 0.7495295405387878,
+        "score": 0.5715821981430054,
+        "task_id": 54,
+        "task_label": "reverted_card_payment?"
+      },
+      {
+        "probability": 5.729981350061799e-17,
+        "reconstruction_error": 0.7506908178329468,
+        "score": 0.5712031126022339,
+        "task_id": 53,
+        "task_label": "request_refund"
+      },
+      {
+        "probability": 4.684637806648321e-17,
+        "reconstruction_error": 0.7546893954277039,
+        "score": 0.5699014067649841,
+        "task_id": 52,
+        "task_label": "receiving_money"
+      },
+      {
+        "probability": 4.108285487345463e-17,
+        "reconstruction_error": 0.757295548915863,
+        "score": 0.5690562129020691,
+        "task_id": 36,
+        "task_label": "extra_charge_on_statement"
+      },
+      {
+        "probability": 3.7374270006411935e-17,
+        "reconstruction_error": 0.7591736316680908,
+        "score": 0.5684487223625183,
+        "task_id": 18,
+        "task_label": "card_payment_not_recognised"
+      },
+      {
+        "probability": 3.4640395068532325e-17,
+        "reconstruction_error": 0.7606815695762634,
+        "score": 0.5679618716239929,
+        "task_id": 75,
+        "task_label": "why_verify_identity"
+      },
+      {
+        "probability": 3.1090205362626185e-17,
+        "reconstruction_error": 0.7628280520439148,
+        "score": 0.5672702789306641,
+        "task_id": 65,
+        "task_label": "transfer_fee_charged"
+      },
+      {
+        "probability": 2.6817489935267226e-17,
+        "reconstruction_error": 0.7657629251480103,
+        "score": 0.5663274526596069,
+        "task_id": 43,
+        "task_label": "lost_or_stolen_card"
+      },
+      {
+        "probability": 2.41980273109935e-17,
+        "reconstruction_error": 0.7678033113479614,
+        "score": 0.5656737685203552,
+        "task_id": 71,
+        "task_label": "verify_source_of_funds"
+      },
+      {
+        "probability": 2.2321369568069945e-17,
+        "reconstruction_error": 0.7694058418273926,
+        "score": 0.5651614665985107,
+        "task_id": 14,
+        "task_label": "card_delivery_estimate"
+      },
+      {
+        "probability": 1.9210611290776476e-17,
+        "reconstruction_error": 0.7723851799964905,
+        "score": 0.5642114877700806,
+        "task_id": 27,
+        "task_label": "declined_card_payment"
+      },
+      {
+        "probability": 1.79934944287435e-17,
+        "reconstruction_error": 0.7736845016479492,
+        "score": 0.5637981295585632,
+        "task_id": 6,
+        "task_label": "automatic_top_up"
+      },
+      {
+        "probability": 1.0062334514352361e-17,
+        "reconstruction_error": 0.7852223515510559,
+        "score": 0.5601543188095093,
+        "task_id": 69,
+        "task_label": "unable_to_verify_identity"
+      },
+      {
+        "probability": 9.012251060592832e-18,
+        "reconstruction_error": 0.787410318851471,
+        "score": 0.5594686269760132,
+        "task_id": 31,
+        "task_label": "disposable_card_limits"
+      },
+      {
+        "probability": 5.8391593474695095e-18,
+        "reconstruction_error": 0.7960257530212402,
+        "score": 0.5567848682403564,
+        "task_id": 68,
+        "task_label": "transfer_timing"
+      },
+      {
+        "probability": 5.441096462061535e-18,
+        "reconstruction_error": 0.7974274158477783,
+        "score": 0.5563507080078125,
+        "task_id": 1,
+        "task_label": "Refund_not_showing_up"
+      },
+      {
+        "probability": 5.384089655786218e-18,
+        "reconstruction_error": 0.7976365089416504,
+        "score": 0.5562859773635864,
+        "task_id": 21,
+        "task_label": "cash_withdrawal_charge"
+      },
+      {
+        "probability": 3.9270064572806905e-18,
+        "reconstruction_error": 0.8039010167121887,
+        "score": 0.5543541312217712,
+        "task_id": 77,
+        "task_label": "wrong_exchange_rate_for_cash_withdrawal"
+      },
+      {
+        "probability": 3.763549777975924e-18,
+        "reconstruction_error": 0.8047450184822083,
+        "score": 0.5540949106216431,
+        "task_id": 62,
+        "task_label": "top_up_reverted"
+      },
+      {
+        "probability": 3.416622164060212e-18,
+        "reconstruction_error": 0.8066648244857788,
+        "score": 0.553506076335907,
+        "task_id": 67,
+        "task_label": "transfer_not_received_by_recipient"
+      },
+      {
+        "probability": 2.2661367683199148e-18,
+        "reconstruction_error": 0.8148153424263,
+        "score": 0.5510202646255493,
+        "task_id": 17,
+        "task_label": "card_payment_fee_charged"
+      },
+      {
+        "probability": 2.0517483039535742e-18,
+        "reconstruction_error": 0.8167883157730103,
+        "score": 0.5504218339920044,
+        "task_id": 76,
+        "task_label": "wrong_amount_of_cash_received"
+      },
+      {
+        "probability": 1.8268895941886905e-18,
+        "reconstruction_error": 0.8190926313400269,
+        "score": 0.5497246384620667,
+        "task_id": 39,
+        "task_label": "get_disposable_virtual_card"
+      },
+      {
+        "probability": 1.237229758627202e-18,
+        "reconstruction_error": 0.8268294930458069,
+        "score": 0.5473964810371399,
+        "task_id": 48,
+        "task_label": "pending_cash_withdrawal"
+      },
+      {
+        "probability": 1.2356354713940826e-18,
+        "reconstruction_error": 0.8268551230430603,
+        "score": 0.5473887920379639,
+        "task_id": 3,
+        "task_label": "age_limit"
+      },
+      {
+        "probability": 1.1729103655441865e-18,
+        "reconstruction_error": 0.8278892636299133,
+        "score": 0.5470790863037109,
+        "task_id": 7,
+        "task_label": "balance_not_updated_after_bank_transfer"
+      },
+      {
+        "probability": 8.206558118812094e-19,
+        "reconstruction_error": 0.8349790573120117,
+        "score": 0.5449653267860413,
+        "task_id": 8,
+        "task_label": "balance_not_updated_after_cheque_or_cash_deposit"
+      },
+      {
+        "probability": 5.197417222011962e-19,
+        "reconstruction_error": 0.8440465927124023,
+        "score": 0.5422856211662292,
+        "task_id": 56,
+        "task_label": "terminate_account"
+      },
+      {
+        "probability": 3.4778196224150877e-19,
+        "reconstruction_error": 0.8520220518112183,
+        "score": 0.5399503707885742,
+        "task_id": 19,
+        "task_label": "card_payment_wrong_exchange_rate"
+      },
+      {
+        "probability": 1.4447798950201726e-19,
+        "reconstruction_error": 0.8694605231285095,
+        "score": 0.5349136590957642,
+        "task_id": 73,
+        "task_label": "virtual_card_not_working"
+      },
+      {
+        "probability": 1.3004934883028408e-19,
+        "reconstruction_error": 0.8715491890907288,
+        "score": 0.5343167185783386,
+        "task_id": 61,
+        "task_label": "top_up_limits"
+      },
+      {
+        "probability": 9.606922809335173e-20,
+        "reconstruction_error": 0.8775610327720642,
+        "score": 0.5326058268547058,
+        "task_id": 47,
+        "task_label": "pending_card_payment"
+      },
+      {
+        "probability": 7.423343060545055e-20,
+        "reconstruction_error": 0.88267982006073,
+        "score": 0.5311577916145325,
+        "task_id": 37,
+        "task_label": "failed_transfer"
+      },
+      {
+        "probability": 3.906830901755063e-20,
+        "reconstruction_error": 0.8954225182533264,
+        "score": 0.5275868773460388,
+        "task_id": 33,
+        "task_label": "exchange_charge"
+      },
+      {
+        "probability": 2.227829226359017e-20,
+        "reconstruction_error": 0.9065731167793274,
+        "score": 0.5245012640953064,
+        "task_id": 50,
+        "task_label": "pending_transfer"
+      },
+      {
+        "probability": 1.988301066197745e-21,
+        "reconstruction_error": 0.9545407891273499,
+        "score": 0.5116291642189026,
+        "task_id": 34,
+        "task_label": "exchange_rate"
+      },
+      {
+        "probability": 1.2283238770365118e-24,
+        "reconstruction_error": 1.1012307405471802,
+        "score": 0.47591158747673035,
+        "task_id": 42,
+        "task_label": "getting_virtual_card"
+      }
+    ],
+    "task_id": 23,
+    "task_label": "change_pin"
+  },
+  "representation_id": "e5991f3eff5987b1ad5228b158682be638c7944fcd891bc30b461c4079f387d2"
+}

evidence/release/lifecycle-console.txt

@@ -0,0 +1,38 @@
+{
+  "job_status": "accepted_candidate",
+  "result": {
+    "baseline_generation": "33122a0fb5964d529e40feea484addb3",
+    "contract_id": "17876d9ff9eafc7bae3d9b4ca90f91e4b261c80c5851ef2f6f08ec664ee4061e",
+    "entry": {
+      "artifact_sha256": "e0238e945abc6a7d18d8138636ca05669590b23ecef88a600751e6b06ab957e3",
+      "backend": "affine",
+      "contract": {
+        "criteria": {
+          "false": "A request to solve arithmetic",
+          "true": "A request to translate text"
+        },
+        "input_template": "text-v1",
+        "instructions": "Functional fixture: is the text a request to translate rather than calculate?",
+        "type": "binary"
+      },
+      "legacy_adapter": false,
+      "max_novelty": 1.0,
+      "min_confidence": 0.0,
+      "min_margin": 0.0,
+      "model": "/Users/oleg/dev/cmfpublic/tmp/decision-lifecycle-fixture-6e4eb06d/candidates/13008a306bc649ca9a6be0573874fa88/model.cmf",
+      "name": "SYNTHETIC binary lifecycle fixture",
+      "policy_file": "/Users/oleg/dev/cmfpublic/tmp/decision-lifecycle-fixture-6e4eb06d/candidates/13008a306bc649ca9a6be0573874fa88/trained/policy.json",
+      "tenant": "default"
+    },
+    "metrics": {
+      "accuracy": 1.0,
+      "coverage": 1.0,
+      "old_controls_checked": 0,
+      "route_controls_n": 43,
+      "selective_accuracy": 1.0
+    },
+    "passed": true,
+    "report_path": "/Users/oleg/dev/cmfpublic/tmp/decision-lifecycle-fixture-6e4eb06d/candidates/13008a306bc649ca9a6be0573874fa88/acceptance.json",
+    "report_sha256": "b43e607837aadf6bc70e45dbe5692b396ea44a152f60b8f50044e93bf765c477"
+  }
+}

evidence/release/live-release-status.json

@@ -0,0 +1,31 @@
+{
+  "models": [
+    {
+      "name": "task-routing",
+      "backend": "affine",
+      "type": "choice"
+    },
+    {
+      "name": "agent-tool-selection",
+      "backend": "affine",
+      "type": "choice"
+    },
+    {
+      "name": "banking-support-intent",
+      "backend": "affine",
+      "type": "choice"
+    },
+    {
+      "name": "fcd-incremental-skills",
+      "backend": "fcd",
+      "type": "choice"
+    }
+  ],
+  "generation": "b38a26d1be634f908cfbd6b48ed9ddf8",
+  "oracle_configured": true,
+  "oracle_model": "typesafe/jev-1.13",
+  "persisted_controls": 123,
+  "ui_file_absent": true,
+  "isolation_build_passed": true,
+  "current_article": "/Users/oleg/dev/cmfpublic/docs/decision/ARTICLE_RU.md"
+}

evidence/release/metal-acceptance-console.txt

@@ -0,0 +1,117 @@
+huggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...
+To disable this warning, you can either:
+	- Avoid using `tokenizers` before the fork if possible
+	- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)
+test 0
+test 500
+test 1000
+test 1500
+test 2000
+test 2500
+test 3000
+huggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...
+To disable this warning, you can either:
+	- Avoid using `tokenizers` before the fork if possible
+	- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)
+huggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...
+To disable this warning, you can either:
+	- Avoid using `tokenizers` before the fork if possible
+	- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)
+{
+  "backend": "mlx-metal-f32 (encoder); Rust packed CPU (decision)",
+  "test_n": 3080,
+  "correct": 2818,
+  "accuracy": 0.9149350649350649,
+  "winner_changes": [],
+  "gate_changes_at_085": [],
+  "novelty_changes": [],
+  "max_confidence_delta": 3.337860107421875e-06,
+  "probe_max_abs_embedding_delta": 1.993030309677124e-07,
+  "accepted": 2284,
+  "accepted_correct": 2254,
+  "selective_accuracy": 0.9868651488616462,
+  "warm_wall_ms": {
+    "p50": 7.799104000241641,
+    "p95": 11.332041650030076,
+    "p99": 12.547094820120039
+  },
+  "batch_encoder_dev": [
+    {
+      "batch": 1,
+      "median_ms": 4.802416499842366,
+      "texts_per_second": 208.22850330304004
+    },
+    {
+      "batch": 8,
+      "median_ms": 12.841916500292427,
+      "texts_per_second": 622.9599764036644
+    },
+    {
+      "batch": 32,
+      "median_ms": 43.0315840003459,
+      "texts_per_second": 743.639834400304
+    },
+    {
+      "batch": 128,
+      "median_ms": 173.48837500003356,
+      "texts_per_second": 737.8015962163184
+    }
+  ],
+  "memory_samples": [
+    {
+      "index": 0,
+      "python_rss": 1545519104,
+      "rust_rss": 895680512,
+      "metal_active": 470025172,
+      "metal_cache": 380062036
+    },
+    {
+      "index": 500,
+      "python_rss": 1551564800,
+      "rust_rss": 895680512,
+      "metal_active": 470025708,
+      "metal_cache": 389108908
+    },
+    {
+      "index": 1000,
+      "python_rss": 1556037632,
+      "rust_rss": 895680512,
+      "metal_active": 470026088,
+      "metal_cache": 412656268
+    },
+    {
+      "index": 1500,
+      "python_rss": 1556578304,
+      "rust_rss": 895680512,
+      "metal_active": 470026112,
+      "metal_cache": 412656268
+    },
+    {
+      "index": 2000,
+      "python_rss": 1558413312,
+      "rust_rss": 895680512,
+      "metal_active": 470026208,
+      "metal_cache": 413737612
+    },
+    {
+      "index": 2500,
+      "python_rss": 1027850240,
+      "rust_rss": 895680512,
+      "metal_active": 470026256,
+      "metal_cache": 431434444
+    },
+    {
+      "index": 3000,
+      "python_rss": 1028276224,
+      "rust_rss": 895680512,
+      "metal_active": 470026280,
+      "metal_cache": 431434444
+    }
+  ],
+  "python_peak_rss_bytes": 1559003136,
+  "no_result_cache": true,
+  "compiler_warmed_on_dev_only": true,
+  "artifact_sha256": "bc0393616aed93b26862a04fd61f78cb79245ad122dcbe3b3e3fd2620f64f9d3",
+  "platform": "macOS-27.2-arm64-arm-64bit",
+  "pass": true
+}

evidence/release/near-duplicate-console.txt

@@ -0,0 +1,11 @@
+{
+  "audit": "post-hoc character 4-gram Jaccard vs fitted TRAIN only",
+  "threshold": 0.8,
+  "test_n": 3080,
+  "train_n": 6996,
+  "flagged": 95,
+  "posthoc_unflagged_n": 2985,
+  "posthoc_unflagged_cmf_accuracy": 0.91356783919598,
+  "no_retraining_no_test_change": true,
+  "semantic_duplicates_and_encoder_pretraining_contamination_not_ruled_out": true
+}

evidence/release/python-tests.txt

@@ -0,0 +1,5 @@
+.................................
+----------------------------------------------------------------------
+Ran 33 tests in 0.042s
+
+OK

evidence/release/release-acceptance-console.txt

@@ -0,0 +1,19 @@
+{
+  "passed": true,
+  "test_n": 3080,
+  "old_controls_n": 43,
+  "old_changes": [],
+  "release_prediction_mismatches": {
+    "banking": 0,
+    "fcd": 0
+  },
+  "artifact_sha256": {
+    "banking": "4cdaf9ec19054f17cd34d420a56f4c8bb0eed12c2429732e9baf902ef52303df",
+    "fcd": "62270619d5c54474ee841abd411349ffd2d0a882b95ef61bb8594b2495759c74"
+  },
+  "input_artifact_sha256": {
+    "banking": "bc0393616aed93b26862a04fd61f78cb79245ad122dcbe3b3e3fd2620f64f9d3",
+    "fcd": "4beef0e577d29ef52fc0941b7813286aadfe0afe49959d927350988368091e4f"
+  },
+  "scope": "contract-bound export is decision-identical to evaluated models; packaging is not retraining"
+}

evidence/release/runtime-cpu-console-final.txt

@@ -0,0 +1,35 @@
+{
+  "mode": "cpu",
+  "dataset": "first 128 DEV, 3 passes, not quality evaluation",
+  "scope": "Python Worker roundtrip includes JSON IPC, encoding, decision, no HTTP",
+  "p50_ms": 6.097750499975518,
+  "p95_ms": 12.997524949696526,
+  "p99_ms": 19.843087499630208,
+  "memory_after_single_workload": {
+    "python_rss": 50249728,
+    "rust_rss": 1736048640
+  },
+  "batch_encoder": [
+    {
+      "batch": 1,
+      "p50_ms": 4.170291500486201,
+      "texts_per_sec": 239.7913910534583
+    },
+    {
+      "batch": 8,
+      "p50_ms": 73.517583499779,
+      "texts_per_sec": 108.81750486295635
+    },
+    {
+      "batch": 32,
+      "p50_ms": 275.66524950088933,
+      "texts_per_sec": 116.08282167570331
+    },
+    {
+      "batch": 128,
+      "p50_ms": 945.4263545003414,
+      "texts_per_sec": 135.38865231617973
+    }
+  ],
+  "discard_duplicate_cmf_encoder_mappings": true
+}

evidence/release/runtime-cpu-console.txt

@@ -0,0 +1,35 @@
+{
+  "mode": "cpu",
+  "dataset": "first 128 DEV, 3 passes, not quality evaluation",
+  "scope": "Python Worker roundtrip includes JSON IPC, encoding, decision, no HTTP",
+  "p50_ms": 6.574583499059372,
+  "p95_ms": 12.957724899570163,
+  "p99_ms": 16.551455000044378,
+  "memory_after_single_workload": {
+    "python_rss": 529547264,
+    "rust_rss": 1730215936
+  },
+  "batch_encoder": [
+    {
+      "batch": 1,
+      "p50_ms": 5.3011249992778176,
+      "texts_per_sec": 188.63920396825804
+    },
+    {
+      "batch": 8,
+      "p50_ms": 77.20885399976396,
+      "texts_per_sec": 103.6150594855929
+    },
+    {
+      "batch": 32,
+      "p50_ms": 296.38806250022753,
+      "texts_per_sec": 107.96656157491307
+    },
+    {
+      "batch": 128,
+      "p50_ms": 1050.3822495002169,
+      "texts_per_sec": 121.86039897466257
+    }
+  ],
+  "discard_duplicate_cmf_encoder_mappings": true
+}

evidence/release/runtime-metal-console-final.txt

@@ -0,0 +1,41 @@
+{
+  "mode": "metal",
+  "dataset": "first 128 DEV, 3 passes, not quality evaluation",
+  "scope": "Python Worker roundtrip includes JSON IPC, encoding, decision, no HTTP",
+  "p50_ms": 6.139041500318854,
+  "p95_ms": 11.730420550338748,
+  "p99_ms": 12.415499999715397,
+  "memory_after_single_workload": {
+    "python_rss": 934805504,
+    "rust_rss": 18153472
+  },
+  "batch_encoder": [
+    {
+      "batch": 1,
+      "p50_ms": 4.852812000535778,
+      "texts_per_sec": 206.0660911425364
+    },
+    {
+      "batch": 8,
+      "p50_ms": 16.870395499609003,
+      "texts_per_sec": 474.2034648912298
+    },
+    {
+      "batch": 32,
+      "p50_ms": 48.76266700011911,
+      "texts_per_sec": 656.2397417664181
+    },
+    {
+      "batch": 128,
+      "p50_ms": 190.6259585002772,
+      "texts_per_sec": 671.4720335416116
+    }
+  ],
+  "discard_duplicate_cmf_encoder_mappings": true,
+  "metal": {
+    "active": 470025164,
+    "cache": 71008256,
+    "peak": 2090107852,
+    "cache_limit": 67108864
+  }
+}

evidence/release/runtime-metal-console.txt

@@ -0,0 +1,41 @@
+{
+  "mode": "metal",
+  "dataset": "first 128 DEV, 3 passes, not quality evaluation",
+  "scope": "Python Worker roundtrip includes JSON IPC, encoding, decision, no HTTP",
+  "p50_ms": 6.0803545002272585,
+  "p95_ms": 11.632976849159604,
+  "p99_ms": 14.136444029682028,
+  "memory_after_single_workload": {
+    "python_rss": 1426259968,
+    "rust_rss": 18186240
+  },
+  "batch_encoder": [
+    {
+      "batch": 1,
+      "p50_ms": 2.4183749992516823,
+      "texts_per_sec": 413.5008012857518
+    },
+    {
+      "batch": 8,
+      "p50_ms": 12.943416500093008,
+      "texts_per_sec": 618.07483363782
+    },
+    {
+      "batch": 32,
+      "p50_ms": 54.572499499954574,
+      "texts_per_sec": 586.3759273116423
+    },
+    {
+      "batch": 128,
+      "p50_ms": 213.30472950012336,
+      "texts_per_sec": 600.0804590688927
+    }
+  ],
+  "discard_duplicate_cmf_encoder_mappings": true,
+  "metal": {
+    "active": 470025016,
+    "cache": 71008256,
+    "peak": 2090107704,
+    "cache_limit": 67108864
+  }
+}

evidence/release/rust-tests.txt

@@ -0,0 +1,40 @@
+   Compiling cortiq-decision v0.1.0 (/Users/oleg/dev/cmfpublic/tools/cortiq-decision)
+    Finished `release` profile [optimized] target(s) in 8.33s
+warning: the following packages contain code that will be rejected by a future version of Rust: block v0.1.6
+note: to see what the problems were, use the option `--future-incompat-report`, or run `cargo report future-incompatibilities --id 1`
+     Running unittests src/lib.rs (tools/cortiq-decision/target/release/deps/cortiq_decision-1874dc25976d455a)
+
+running 4 tests
+test fcd::tests::support_bound_rejects_invalid_calibration ... ok
+test fcd::tests::reconstruct_and_select_without_materializing_dense_matrix ... ok
+test fcd::tests::unsupported_task_does_not_dilute_old_confidence ... ok
+test fcd::tests::core_and_orthogonality_fail_closed ... ok
+
+test result: ok. 4 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s
+
+     Running unittests src/main.rs (tools/cortiq-decision/target/release/deps/cortiq_decision-e59f30d0f1db229c)
+
+running 0 tests
+
+test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s
+
+     Running tests/contracts.rs (tools/cortiq-decision/target/release/deps/contracts-ef6555562fe91b1e)
+
+running 8 tests
+test empty_active_set_abstains_and_stable_ties ... ok
+test rejects_invalid_shapes_numbers_and_calibration ... ok
+test packed_exact_nonorthogonal_variable_rank_and_tail ... ok
+test roundtrip_f32_flags_identity_and_legacy_reader ... ok
+test policy_binds_model_and_abstains_without_gate ... ok
+test exporter_does_not_overwrite ... ok
+test corrupted_payload_and_missing_feature_refused ... ok
+test metal_handles_dimension_tail_variable_rank_and_batch ... ok
+
+test result: ok. 8 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.06s
+
+   Doc-tests cortiq_decision
+
+running 0 tests
+
+test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s
+

evidence/release/vulkan-acceptance.json

@@ -0,0 +1,44 @@
+{
+  "accepted": 2284,
+  "accepted_correct": 2254,
+  "adapter": "NVIDIA RTX PRO 4000 Blackwell / NVIDIA 595.91.07 / Vulkan",
+  "correct": 2818,
+  "gate_changes": 0,
+  "max_abs_error": 1.430511474609375e-6,
+  "model_sha256": "4cdaf9ec19054f17cd34d420a56f4c8bb0eed12c2429732e9baf902ef52303df",
+  "pass": true,
+  "resident_buffer_bytes": 4178756,
+  "scope": "resident decision scorer only; precomputed same-representation embeddings; upload/submit/wait/readback included; encoder NOT measured",
+  "test_n": 3080,
+  "timing": [
+    {
+      "batch": 1,
+      "cpu_p50_us": 929.503,
+      "cpu_p95_us": 1018.4110000000001,
+      "vulkan_p50_us": 1407.564,
+      "vulkan_p95_us": 3020.487
+    },
+    {
+      "batch": 8,
+      "cpu_p50_us": 7172.128,
+      "cpu_p95_us": 7534.2339999999995,
+      "vulkan_p50_us": 1303.231,
+      "vulkan_p95_us": 3771.9860000000003
+    },
+    {
+      "batch": 32,
+      "cpu_p50_us": 28570.895999999997,
+      "cpu_p95_us": 29261.018,
+      "vulkan_p50_us": 1501.221,
+      "vulkan_p95_us": 2035.341
+    },
+    {
+      "batch": 128,
+      "cpu_p50_us": 114625.38799999999,
+      "cpu_p95_us": 116469.27,
+      "vulkan_p50_us": 2189.832,
+      "vulkan_p95_us": 3748.502
+    }
+  ],
+  "winner_changes": 0
+}

evidence/release/vulkan-check-macos.txt

@@ -0,0 +1,143 @@
+    Updating crates.io index
+     Locking 72 packages to latest Rust 1.96.0 compatible versions
+      Adding allocator-api2 v0.2.21
+      Adding android_system_properties v0.1.6
+      Adding ash v0.37.3+1.3.251
+      Adding bit-set v0.5.3
+      Adding bit-vec v0.6.3
+      Adding cfg_aliases v0.1.1
+      Adding codespan-reporting v0.11.1
+      Adding com v0.6.0
+      Adding com_macros v0.6.0
+      Adding com_macros_support v0.6.0
+      Adding d3d12 v0.19.0
+      Adding equivalent v1.0.2
+      Adding futures-core v0.3.34
+      Adding futures-task v0.3.34
+      Adding futures-util v0.3.34
+      Adding gl_generator v0.14.0
+      Adding glow v0.13.1
+      Adding glutin_wgl_sys v0.5.0
+      Adding gpu-alloc v0.6.2
+      Adding gpu-alloc-types v0.3.1
+      Adding gpu-allocator v0.25.0
+      Adding gpu-descriptor v0.2.4
+      Adding gpu-descriptor-types v0.1.2
+      Adding hashbrown v0.14.5
+      Adding hashbrown v0.17.1
+      Adding hassle-rs v0.11.0
+      Adding hexf-parse v0.2.1
+      Adding indexmap v2.14.2
+      Adding jni-sys v0.3.1
+      Adding jni-sys v0.4.1
+      Adding jni-sys-macros v0.4.1
+      Adding js-sys v0.3.106
+      Adding khronos-egl v6.0.0
+      Adding khronos_api v3.1.0
+      Adding libloading v0.7.4
+      Adding libloading v0.8.9
+      Adding lock_api v0.4.14
+      Adding metal v0.27.0
+      Adding naga v0.19.2
+      Adding ndk-sys v0.5.0+25.2.9519653
+      Adding objc_exception v0.1.2
+      Adding parking_lot v0.12.5
+      Adding parking_lot_core v0.9.12
+      Adding pollster v0.3.0 (available: v1.0.1)
+      Adding presser v0.3.1
+      Adding range-alloc v0.1.5
+      Adding raw-window-handle v0.6.2
+      Adding redox_syscall v0.5.18
+      Adding renderdoc-sys v1.1.0
+      Adding rustc-hash v1.1.0
+      Adding scopeguard v1.2.0
+      Adding slab v0.4.12
+      Adding slotmap v1.1.1
+      Adding spirv v0.3.0+sdk-1.3.268.0
+      Adding syn v1.0.109
+      Adding termcolor v1.4.1
+      Adding thiserror v1.0.69
+      Adding thiserror-impl v1.0.69
+      Adding tokio v1.53.1
+      Adding unicode-width v0.1.14
+      Adding unicode-xid v0.2.6
+      Adding wasm-bindgen-futures v0.4.79
+      Adding web-sys v0.3.106
+      Adding wgpu v0.19.4 (available: v30.0.1)
+      Adding wgpu-core v0.19.4
+      Adding wgpu-hal v0.19.5
+      Adding wgpu-types v0.19.2
+      Adding widestring v1.2.1
+      Adding winapi-util v0.1.11
+      Adding windows v0.52.0
+      Adding windows-core v0.52.0
+      Adding xml-rs v0.8.29
+ Downloading crates ...
+  Downloaded pollster v0.3.0
+   Compiling cfg_aliases v0.1.1
+   Compiling objc_exception v0.1.2
+   Compiling num-traits v0.2.19
+   Compiling parking_lot_core v0.9.12
+   Compiling thiserror v1.0.69
+   Compiling syn v3.0.6
+   Compiling serde_core v1.0.229
+   Compiling thiserror-impl v1.0.69
+    Checking bit-vec v0.6.3
+    Checking scopeguard v1.2.0
+    Checking hashbrown v0.17.1
+    Checking termcolor v1.4.1
+    Checking unicode-width v0.1.14
+    Checking equivalent v1.0.2
+    Checking bit-set v0.5.3
+    Checking codespan-reporting v0.11.1
+    Checking lock_api v0.4.14
+    Checking indexmap v2.14.2
+   Compiling wgpu-hal v0.19.5
+    Checking unicode-xid v0.2.6
+    Checking hexf-parse v0.2.1
+    Checking rustc-hash v1.1.0
+   Compiling serde v1.0.229
+   Compiling wgpu-core v0.19.4
+    Checking libloading v0.8.9
+    Checking wgpu-types v0.19.2
+    Checking profiling v1.0.18
+    Checking raw-window-handle v0.6.2
+   Compiling wgpu v0.19.4
+    Checking parking_lot v0.12.5
+    Checking pollster v0.3.0
+    Checking naga v0.19.2
+    Checking objc v0.2.7
+   Compiling foreign-types-macros v0.2.4
+   Compiling serde_derive v1.0.229
+   Compiling thiserror-impl v2.0.21
+   Compiling bytemuck_derive v1.12.1
+   Compiling clap_derive v4.6.7
+    Checking foreign-types v0.5.0
+    Checking serde_json v1.0.151
+    Checking bytemuck v1.25.2
+    Checking metal v0.27.0
+    Checking thiserror v2.0.21
+    Checking metal v0.29.0
+    Checking clap v4.6.7
+    Checking bincode v1.3.3
+    Checking cortiq-core v0.6.8 (/Users/oleg/dev/cmfpublic/crates/cortiq-core)
+    Checking cortiq-decision v0.1.0 (/Users/oleg/dev/cmfpublic/tools/cortiq-decision)
+error[E0596]: cannot borrow `packed` as mutable, as it is not declared as mutable
+  --> src/vulkan.rs:57:63
+   |
+57 | ...count();let packed=Packed::new(&m.seed)?;let mut gpu=VulkanScorer::new(&m.seed,128)?;
+   |                ^^^^^^ not mutable
+...
+62 | ...tual.chunks_exact(nt)).zip(batch){let mut expected=vec![0.;nt];packed.errors_into(x,&mut expected)?;for(a,b)in e.iter().zip(&expe...
+   |                                                                   ------ cannot borrow as mutable
+...
+67 | ...let start=Instant::now();for(x,o)in xs.chunks_exact(d).zip(out.chunks_exact_mut(nt)){packed.errors_into(x,o)?;}std::hint::black_b...
+   |                                                                                         ------ cannot borrow as mutable
+   |
+help: consider changing this to be mutable
+   |
+57 |     let d=m.seed.input_dim;let nt=m.seed.active().count();let mut packed=Packed::new(&m.seed)?;let mut gpu=VulkanScorer::new(&m.seed,128)?;
+   |                                                               +++
+
+For more information about this error, try `rustc --explain E0596`.
+error: could not compile `cortiq-decision` (lib) due to 1 previous error

evidence/release/vulkan-environment.txt

@@ -0,0 +1,9 @@
+GPU workload at acceptance: pre-existing unrelated job, NOT stopped
+name, utilization.gpu [%], memory.used [MiB], memory.total [MiB], driver_version
+NVIDIA RTX PRO 4000 Blackwell, 88 %, 12516 MiB, 24467 MiB, 595.91.07
+Architecture:                            x86_64
+CPU(s):                                  28
+Model name:                              Intel(R) Xeon(R) CPU E5-2690 v4 @ 2.60GHz
+4cdaf9ec19054f17cd34d420a56f4c8bb0eed12c2429732e9baf902ef52303df  /data/cmf-decision-release-20260925/cortiq-banking77.cmf
+Filesystem      Size  Used Avail Use% Mounted on
+/dev/nvme1n1p1  938G  557G  334G  63% /data

evidence/release/vulkan-linux-tests.txt

@@ -0,0 +1,114 @@
+   Compiling proc-macro2 v1.0.107
+   Compiling quote v1.0.47
+   Compiling unicode-ident v1.0.26
+   Compiling cfg-if v1.0.5
+   Compiling version_check v0.9.5
+   Compiling bitflags v2.13.2
+   Compiling libc v0.2.189
+   Compiling once_cell v1.21.4
+   Compiling cfg_aliases v0.1.1
+   Compiling zerocopy v0.8.58
+   Compiling autocfg v1.5.1
+   Compiling parking_lot_core v0.9.12
+   Compiling thiserror v1.0.69
+   Compiling serde_core v1.0.229
+   Compiling typenum v1.20.1
+   Compiling equivalent v1.0.2
+   Compiling termcolor v1.4.1
+   Compiling allocator-api2 v0.2.21
+   Compiling khronos-egl v6.0.0
+   Compiling ash v0.37.3+1.3.251
+   Compiling hashbrown v0.17.1
+   Compiling unicode-width v0.1.14
+   Compiling scopeguard v1.2.0
+   Compiling smallvec v1.16.2
+   Compiling bit-vec v0.6.3
+   Compiling rustc-hash v1.1.0
+   Compiling log v0.4.34
+   Compiling unicode-xid v0.2.6
+   Compiling libloading v0.8.9
+   Compiling wgpu-hal v0.19.5
+   Compiling libloading v0.7.4
+   Compiling hexf-parse v0.2.1
+   Compiling lock_api v0.4.14
+   Compiling zmij v1.0.23
+   Compiling utf8parse v0.2.2
+   Compiling serde v1.0.229
+   Compiling wgpu-core v0.19.4
+   Compiling spirv v0.3.0+sdk-1.3.268.0
+   Compiling gpu-alloc-types v0.3.1
+   Compiling gpu-descriptor-types v0.1.2
+   Compiling anstyle-parse v1.0.0
+   Compiling bit-set v0.5.3
+   Compiling wgpu-types v0.19.2
+   Compiling getrandom v0.4.3
+   Compiling anstyle v1.0.14
+   Compiling rustix v1.1.5
+   Compiling colorchoice v1.0.5
+   Compiling thiserror v2.0.21
+   Compiling codespan-reporting v0.11.1
+   Compiling ahash v0.8.12
+   Compiling generic-array v0.14.7
+   Compiling arrayvec v0.7.8
+   Compiling profiling v1.0.18
+   Compiling is_terminal_polyfill v1.70.2
+   Compiling glow v0.13.1
+   Compiling anstyle-query v1.1.5
+   Compiling raw-window-handle v0.6.2
+   Compiling renderdoc-sys v1.1.0
+   Compiling serde_json v1.0.151
+   Compiling gpu-alloc v0.6.2
+   Compiling anstream v1.0.0
+   Compiling tracing-core v0.1.36
+   Compiling wgpu v0.19.4
+   Compiling anyhow v1.0.104
+   Compiling strsim v0.11.1
+   Compiling heck v0.5.0
+   Compiling num-traits v0.2.19
+   Compiling memchr v2.8.3
+   Compiling clap_lex v1.1.1
+   Compiling linux-raw-sys v0.12.1
+   Compiling pin-project-lite v0.2.17
+   Compiling itoa v1.0.18
+   Compiling cpufeatures v0.2.17
+   Compiling fastrand v2.5.0
+   Compiling static_assertions v1.1.0
+   Compiling pollster v0.3.0
+   Compiling clap_builder v4.6.7
+   Compiling indexmap v2.14.2
+   Compiling syn v3.0.6
+   Compiling syn v2.0.119
+   Compiling crypto-common v0.1.7
+   Compiling block-buffer v0.10.4
+   Compiling digest v0.10.7
+   Compiling memmap2 v0.9.11
+   Compiling sha2 v0.10.9
+   Compiling parking_lot v0.12.5
+   Compiling tempfile v3.27.0
+   Compiling serde_derive v1.0.229
+   Compiling thiserror-impl v2.0.21
+   Compiling bytemuck_derive v1.12.1
+   Compiling clap_derive v4.6.7
+   Compiling hashbrown v0.14.5
+   Compiling bytemuck v1.25.2
+   Compiling thiserror-impl v1.0.69
+   Compiling tracing-attributes v0.1.31
+   Compiling clap v4.6.7
+   Compiling gpu-descriptor v0.2.4
+   Compiling tracing v0.1.44
+   Compiling naga v0.19.2
+   Compiling bincode v1.3.3
+   Compiling cortiq-core v0.6.8 (/data/cmf-decision-release-20260925/source/crates/cortiq-core)
+   Compiling cortiq-decision v0.1.0 (/data/cmf-decision-release-20260925/source/tools/cortiq-decision)
+    Finished `test` profile [unoptimized + debuginfo] target(s) in 31.45s
+     Running unittests src/lib.rs (/data/cmf-decision-release-20260925/target/debug/deps/cortiq_decision-98c66b2e57c892e3)
+
+running 5 tests
+test fcd::tests::unsupported_task_does_not_dilute_old_confidence ... ok
+test fcd::tests::reconstruct_and_select_without_materializing_dense_matrix ... ok
+test fcd::tests::support_bound_rejects_invalid_calibration ... ok
+test fcd::tests::core_and_orthogonality_fail_closed ... ok
+test vulkan::tests::shader_parses_and_has_valid_uniform_barriers ... ok
+
+test result: ok. 5 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s
+

evidence/release/vulkan-shader-test.txt

@@ -0,0 +1,54 @@
+   Compiling objc_exception v0.1.2
+   Compiling equivalent v1.0.2
+   Compiling scopeguard v1.2.0
+   Compiling foreign-types v0.5.0
+   Compiling bit-vec v0.6.3
+   Compiling termcolor v1.4.1
+   Compiling unicode-width v0.1.14
+   Compiling hashbrown v0.17.1
+   Compiling num-traits v0.2.19
+   Compiling parking_lot_core v0.9.12
+   Compiling objc v0.2.7
+   Compiling lock_api v0.4.14
+   Compiling thiserror v1.0.69
+   Compiling serde_core v1.0.229
+   Compiling hexf-parse v0.2.1
+   Compiling bit-set v0.5.3
+   Compiling rustc-hash v1.1.0
+   Compiling unicode-xid v0.2.6
+   Compiling codespan-reporting v0.11.1
+   Compiling parking_lot v0.12.5
+   Compiling wgpu-types v0.19.2
+   Compiling libloading v0.8.9
+   Compiling metal v0.27.0
+   Compiling raw-window-handle v0.6.2
+   Compiling profiling v1.0.18
+   Compiling bytemuck v1.25.2
+   Compiling thiserror v2.0.21
+   Compiling clap v4.6.7
+   Compiling metal v0.29.0
+   Compiling pollster v0.3.0
+   Compiling indexmap v2.14.2
+   Compiling naga v0.19.2
+   Compiling serde v1.0.229
+   Compiling serde_json v1.0.151
+   Compiling bincode v1.3.3
+   Compiling cortiq-core v0.6.8 (/Users/oleg/dev/cmfpublic/crates/cortiq-core)
+   Compiling wgpu-hal v0.19.5
+   Compiling wgpu-core v0.19.4
+   Compiling wgpu v0.19.4
+   Compiling cortiq-decision v0.1.0 (/Users/oleg/dev/cmfpublic/tools/cortiq-decision)
+    Finished `test` profile [unoptimized + debuginfo] target(s) in 17.13s
+warning: the following packages contain code that will be rejected by a future version of Rust: block v0.1.6
+note: to see what the problems were, use the option `--future-incompat-report`, or run `cargo report future-incompatibilities --id 1`
+     Running unittests src/lib.rs (tools/cortiq-decision/target/debug/deps/cortiq_decision-ae9ff32075835df0)
+
+running 5 tests
+test fcd::tests::core_and_orthogonality_fail_closed ... ok
+test fcd::tests::reconstruct_and_select_without_materializing_dense_matrix ... ok
+test fcd::tests::support_bound_rejects_invalid_calibration ... ok
+test fcd::tests::unsupported_task_does_not_dilute_old_confidence ... ok
+test vulkan::tests::shader_parses_and_has_valid_uniform_barriers ... ok
+
+test result: ok. 5 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.01s
+

reproduce.sh

@@ -0,0 +1,15 @@
+#!/bin/sh
+# Offline evaluation only. No API calls, credentials, or test-based fitting.
+set -eu
+HERE=$(CDPATH= cd -- "$(dirname -- "$0")" && pwd)
+cd "$HERE"
+OUT=${1:?Usage: ./reproduce.sh /path/to/new-output-directory}
+if [ -e "$OUT" ]; then echo 'Refusing to overwrite an existing experiment' >&2; exit 1; fi
+mkdir -p "$OUT"
+OUT=$(CDPATH= cd -- "$OUT" && pwd)
+# Build artifacts stay outside the release payload.
+export CARGO_TARGET_DIR="$OUT/build"
+cargo build --release --locked --manifest-path source/tools/cortiq-decision/Cargo.toml
+"$CARGO_TARGET_DIR/release/cortiq-decision" eval cortiq-banking77.cmf evidence/banking77/data/test.jsonl -o "$OUT/local-test.jsonl" > "$OUT/eval-summary.json"
+python3 -m unittest discover -s source/tools/cortiq-decision/tests -p 'test_*.py'
+printf 'Saved independent offline run to %s\n' "$OUT"

source/tools/cortiq-decision/decision_server.py

@@ -77,7 +77,7 @@ def execute(c, label, probabilities):
 def accepts(entry, prediction):
     """One gate shared by serving and regression acceptance (including abstention)."""
     p=prediction;threshold=entry['min_confidence']
-    return (threshold is not None and p['confidence']>=threshold
+    return (threshold is not None and p['confidence']>0. and p['confidence']>=threshold
             and p['margin']>=entry.get('min_margin',0.)
             and p['novelty_score']<=entry.get('max_novelty',1.)
             and (not p['is_novel'] or entry.get('ignore_legacy_novelty',False)))

source/tools/cortiq-decision/src/main.rs

@@ -374,6 +374,8 @@ fn bench(m: &DecisionModel, n: usize) -> Result<()> {
             },
             n,
         )?;
+        // Only the macOS branch appends a Metal measurement.
+        #[allow(unused_mut)]
         let mut row = json!({"batch":batch,"reference":reference_times,"packed_cpu":packed_times,"cpu_bit_exact":exact});
         #[cfg(target_os = "macos")]
         {

source/tools/cortiq-decision/src/vulkan.rs

@@ -289,3 +289,15 @@ pub fn evaluate(model: &Path, embeddings: &Path, output: &Path) -> Result<()> {
     );
     Ok(())
 }
+
+#[cfg(test)]
+mod tests {
+    #[test]
+    fn shader_parses_and_has_valid_uniform_barriers() {
+        let module = wgpu::naga::front::wgsl::parse_str(include_str!("affine.wgsl")).unwrap();
+        wgpu::naga::valid::Validator::new(
+            wgpu::naga::valid::ValidationFlags::all(),
+            wgpu::naga::valid::Capabilities::empty(),
+        ).validate(&module).unwrap();
+    }
+}

source/tools/cortiq-decision/tests/contracts.rs

@@ -0,0 +1,216 @@
+use cortiq_core::format::{check_required_features, features};
+use cortiq_decision::{DecisionModel, ENCODER_FILES, Seed, Task, packed::Packed};
+use std::sync::atomic::{AtomicU64, Ordering};
+static ID: AtomicU64 = AtomicU64::new(0);
+fn seed(d: usize) -> Seed {
+    let tasks = (0..7)
+        .map(|i| Task {
+            id: i * 3,
+            label: format!("t{i}"),
+            mean: (0..d).map(|j| (i + j) as f32 * 0.001).collect(),
+            // Deliberately NOT orthonormal, with variable ranks, to catch energy shortcuts.
+            basis: (0..1 + i % 3)
+                .map(|k| (0..d).map(|j| ((i + j + k) % 7) as f32 * 0.009).collect())
+                .collect(),
+            err_mean: 0.1,
+            err_std: 0.02,
+            n_train: 7,
+            frozen: i % 2 == 0,
+            active: i != 2,
+        })
+        .collect();
+    Seed {
+        taxonomy_id: "test".into(),
+        taxonomy_version: 1,
+        model_version: "fixture".into(),
+        input_dim: d,
+        k: 3,
+        temperature: 0.12,
+        novelty_theta: 0.7,
+        tasks,
+    }
+}
+struct Temp(std::path::PathBuf);
+impl Temp {
+    fn new() -> Self {
+        let p = std::env::temp_dir().join(format!(
+            "cmf-decision-test-{}-{}",
+            std::process::id(),
+            ID.fetch_add(1, Ordering::Relaxed)
+        ));
+        std::fs::create_dir(&p).unwrap();
+        Self(p)
+    }
+}
+impl Drop for Temp {
+    fn drop(&mut self) {
+        std::fs::remove_dir_all(&self.0).unwrap();
+    }
+}
+fn package() -> (Temp, DecisionModel) {
+    let t = Temp::new();
+    for name in ENCODER_FILES {
+        std::fs::write(t.0.join(name), format!("fixture-{name}")).unwrap();
+    }
+    std::fs::write(
+        t.0.join("seed.json"),
+        serde_json::to_vec(&seed(37)).unwrap(),
+    )
+    .unwrap();
+    cortiq_decision::export(
+        &t.0.join("seed.json"),
+        &t.0,
+        "fixture",
+        "mean",
+        &t.0.join("model.cmf"),
+    )
+    .unwrap();
+    let m = DecisionModel::open(&t.0.join("model.cmf")).unwrap();
+    (t, m)
+}
+#[test]
+fn policy_binds_model_and_abstains_without_gate() {
+    use cortiq_decision::policy::Policy;
+    let (_temp, m) = package();
+    let mut labels: Vec<_> = m.seed.active().map(|t| t.label.clone()).collect();
+    labels.sort();
+    let mut p: Policy = serde_json::from_value(serde_json::json!({
+        "schema":1,"model_version":m.seed.model_version,"seed_sha256":m.manifest.source_sha256,
+        "representation_id":m.manifest.representation_id,"taxonomy_id":m.seed.taxonomy_id,
+        "labels":labels,"min_confidence":null,"target_selective_accuracy":0.95,
+        "calibration_gate":null,"fallback":"abstain_or_explicit_oracle",
+        "oracle_model":"typesafe/jev-1.13","auto_train_from_oracle":false,"scope":"test"
+    }))
+    .unwrap();
+    p.validate(&m).unwrap();
+    let mut prediction = m.seed.predict(&[0.; 37]).unwrap();
+    prediction.is_novel = false;
+    assert_eq!(p.decide(&prediction).action, "abstain");
+    p.min_confidence = Some(0.);
+    assert_eq!(p.decide(&prediction).action, "local");
+    prediction.is_novel = true;
+    assert_eq!(p.decide(&prediction).action, "abstain");
+    p.representation_id = "different".into();
+    assert!(p.validate(&m).is_err());
+    p.representation_id = m.manifest.representation_id.clone();
+    p.auto_train_from_oracle = true;
+    assert!(p.validate(&m).is_err());
+}
+#[test]
+fn packed_exact_nonorthogonal_variable_rank_and_tail() {
+    for d in [3, 17, 37, 384] {
+        let s = seed(d);
+        let mut packed = Packed::new(&s).unwrap();
+        for n in 0..100 {
+            let x: Vec<_> = (0..d).map(|i| ((i + n) % 19) as f32 * 0.13 - 1.0).collect();
+            let expected = s.errors(&x).unwrap();
+            let mut out = vec![0.; expected.len()];
+            packed.errors_into(&x, &mut out).unwrap();
+            assert_eq!(out, expected);
+        }
+    }
+}
+#[test]
+fn rejects_invalid_shapes_numbers_and_calibration() {
+    let mut s = seed(17);
+    s.validate().unwrap();
+    assert!(s.predict(&[1.]).is_err());
+    assert!(s.predict(&[f32::NAN; 17]).is_err());
+    s.tasks[0].err_std = 0.;
+    assert!(s.validate().is_err());
+    s.tasks[0].err_std = 1.;
+    s.tasks[0].basis[0].pop();
+    assert!(s.validate().is_err());
+    let mut s = seed(17);
+    s.tasks[1].id = s.tasks[0].id;
+    assert!(s.validate().is_err());
+    let mut s = seed(17);
+    s.temperature = f32::INFINITY;
+    assert!(s.validate().is_err());
+}
+#[test]
+fn empty_active_set_abstains_and_stable_ties() {
+    let mut s = seed(17);
+    for t in &mut s.tasks {
+        t.active = false;
+    }
+    assert!(s.predict(&[0.; 17]).unwrap().is_novel);
+    s.tasks[0].active = true;
+    s.tasks[1] = s.tasks[0].clone();
+    s.tasks[1].id = 99;
+    s.tasks[1].label = "second".into();
+    assert_eq!(s.predict(&[0.; 17]).unwrap().task_id, 0);
+}
+#[test]
+fn roundtrip_f32_flags_identity_and_legacy_reader() {
+    let (_t, m) = package();
+    let s = seed(37);
+    assert_eq!(
+        serde_json::to_vec(&s).unwrap(),
+        serde_json::to_vec(&m.seed).unwrap()
+    );
+    assert_ne!(m.container.required_features & features::DECISION, 0);
+    assert!(
+        check_required_features(
+            m.container.required_features,
+            features::SUPPORTED & !features::DECISION
+        )
+        .is_err()
+    );
+    assert!(m.predict_vector(&[0.; 37], "wrong").is_err());
+    assert!(
+        m.predict_vector(&[0.; 37], &m.manifest.representation_id)
+            .is_ok()
+    );
+}
+#[test]
+fn corrupted_payload_and_missing_feature_refused() {
+    let (t, m) = package();
+    let original = std::fs::read(t.0.join("model.cmf")).unwrap();
+    let e = m.container.tensor("decision.task.0.mean").unwrap();
+    let off = m.container.entry_abs_offset(e).unwrap();
+    let mut corrupt = original.clone();
+    corrupt[off] ^= 1;
+    std::fs::write(t.0.join("corrupt.cmf"), corrupt).unwrap();
+    assert!(DecisionModel::open(&t.0.join("corrupt.cmf")).is_err());
+    let mut corrupt = original;
+    let f = u32::from_le_bytes(corrupt[12..16].try_into().unwrap()) & !features::DECISION;
+    corrupt[12..16].copy_from_slice(&f.to_le_bytes());
+    std::fs::write(t.0.join("missing.cmf"), corrupt).unwrap();
+    assert!(DecisionModel::open(&t.0.join("missing.cmf")).is_err());
+}
+#[test]
+fn exporter_does_not_overwrite() {
+    let (t, _) = package();
+    assert!(
+        cortiq_decision::export(
+            &t.0.join("seed.json"),
+            &t.0,
+            "fixture",
+            "mean",
+            &t.0.join("model.cmf")
+        )
+        .is_err()
+    );
+}
+#[cfg(target_os = "macos")]
+#[test]
+fn metal_handles_dimension_tail_variable_rank_and_batch() {
+    let s = seed(37);
+    let mut gpu = cortiq_decision::metal::MetalScorer::new(&s, 8).unwrap();
+    let xs: Vec<_> = (0..37 * 8).map(|i| (i % 17) as f32 * 0.02 - 0.12).collect();
+    let nt = s.active().count();
+    let mut out = vec![0.; 8 * nt];
+    gpu.errors_into(&xs, &mut out).unwrap();
+    for (x, got) in xs.chunks_exact(37).zip(out.chunks_exact(nt)) {
+        let expected = s.errors(x).unwrap();
+        for (a, b) in got.iter().zip(expected) {
+            assert!((a - b).abs() < 2e-5, "{a} vs {b}");
+        }
+    }
+    assert!(gpu.errors_into(&xs[..2], &mut out).is_err());
+    assert!(
+        gpu.errors_into(&vec![0.; 37 * 9], &mut vec![0.; 9 * nt])
+            .is_err()
+    );
+}

source/tools/cortiq-decision/tests/test_decisions_protocol.py

@@ -0,0 +1,113 @@
+import copy
+import json
+from pathlib import Path
+import sys
+import tempfile
+import unittest
+from unittest.mock import patch
+
+sys.path.insert(0,str(Path(__file__).resolve().parents[1]))
+from decisions_protocol import (MODEL, ProtocolError, compile_question, concentration,
+                                evaluate, local_answer, validate_request, validate_oracle_response)
+from decision_server import Engine, contract_id
+
+CHOICE={'type':'choice','instructions':'Pick the kind','criteria':{'a':'Alpha','b':'Beta'}}
+NOUL={'type':'noul','instructions':'Is the proposition true?'}
+SCORE={'type':'score','instructions':'How severe?','criteria':['low','medium','high']}
+
+
+def prediction(p):
+    return {'task_label':max(p,key=p.get),'confidence':max(p.values()),'margin':.2,
+            'is_novel':False,'novelty_score':.1,
+            'scores':[{'task_label':k,'probability':v} for k,v in p.items()]}
+
+
+class WorkerStub:
+    rep='fixture'
+    def __init__(self,*args):self.embeds=0;self.predictions=0
+    def load(self,entry):return 'fixture-slot'
+    def call(self,**r):
+        if r['op']=='inspect':return {'metadata':{'model_version':'fixture'}}
+        if r['op']=='embed':self.embeds+=1;return {'vectors':[[1.,0.]]}
+        self.predictions+=1;return {'prediction':prediction({'a':.97,'b':.03}),'model_version':'fixture'}
+
+
+class NativeProtocol(unittest.TestCase):
+    def test_noul_without_criteria(self):
+        c=compile_question(NOUL,'text')
+        self.assertEqual(c['type'],'binary')
+        self.assertEqual(local_answer(NOUL,c,prediction({'true':.7,'false':.3})),{'type':'noul','noul':.7})
+
+    def test_fractional_score_and_legend(self):
+        c=compile_question(SCORE,'text');a=local_answer(SCORE,c,prediction({'0':0.,'1':.85,'2':.15}))
+        self.assertAlmostEqual(a['score'],1.15)
+        self.assertEqual(a['legend'],{'0':'low','1':'medium','2':'high'})
+        self.assertNotEqual(a['confidence'],.85)
+
+    def test_concentration_not_probability(self):
+        self.assertEqual(concentration({'a':.5,'b':.5}),0.)
+        self.assertEqual(concentration({'a':1.,'b':0.}),1.)
+
+    def test_structured_state_and_rubrics(self):
+        q={'type':'choice','instructions':{'question':'Pick','policy':['one','two']},'criteria':{'a':None,'b':{'meaning':'B'}}}
+        c=compile_question(q,{'message':'text'})
+        self.assertEqual(c['input_template'],'json-state-v1');self.assertEqual(c['criteria']['a'],'a')
+        self.assertIn('policy',c['instructions'])
+
+    def test_strict_limits_and_model_identity(self):
+        for state in [None,2,False,float('nan')]:
+            with self.assertRaises(ProtocolError):validate_request({'model':MODEL,'state':state,'questions':{'q':NOUL}})
+        with self.assertRaises(ProtocolError):compile_question({**SCORE,'criteria':['x']*11},'text')
+        with self.assertRaises(ProtocolError):validate_request({'model':'typesafe/jev-1.13','state':'text','questions':{'q':NOUL}})
+
+    def test_ids_not_part_of_inference_shared_encoder(self):
+        with tempfile.TemporaryDirectory() as d,patch('decision_server.Worker',WorkerStub):
+            c=compile_question(CHOICE,'text');cid=contract_id(c)
+            cfg={'state_dir':d,'binary':'unused','seed_model':'unused','families':{cid:{'name':'fixture','model':'unused','artifact_sha256':'fixture','contract':c,'legacy_adapter':True,'min_confidence':.8}},'oracle':{'enabled':False,'budget_usd':1,'max_calls':10}}
+            e=Engine(cfg)
+            result=evaluate(e,{'model':MODEL,'state':'text','questions':{'untrusted_name':CHOICE,'different_id':CHOICE}})
+            self.assertEqual(result['answers']['untrusted_name'],result['answers']['different_id'])
+            self.assertEqual(e.worker.embeds,1);self.assertEqual(e.worker.predictions,1)
+            self.assertEqual(result['usage']['cost'],0.);self.assertEqual(result['provider'],'Cortiq');e.db.close()
+
+    def test_unknown_not_fabricated_and_no_implicit_oracle(self):
+        with tempfile.TemporaryDirectory() as d,patch('decision_server.Worker',WorkerStub):
+            e=Engine({'state_dir':d,'binary':'unused','seed_model':'unused','families':{},'oracle':{'enabled':True,'budget_usd':1,'max_calls':10}},key='fixture')
+            with patch('decision_server.JevNativeClient') as client:
+                with self.assertRaises(ProtocolError) as ctx:evaluate(e,{'model':MODEL,'state':'text','questions':{'q':NOUL}})
+                self.assertEqual(ctx.exception.status,409);self.assertEqual(ctx.exception.body['error']['code'],'CMF_ABSTAIN');client.assert_not_called()
+            self.assertEqual(e.status()['oracle']['attempts'],0);e.db.close()
+
+    def test_batch_oracle_preserves_three_types_and_budget(self):
+        questions={'choice':CHOICE,'score':SCORE,'yes':NOUL}
+        response={'model':'typesafe/jev-1.13-20260917','answers':{
+            'choice':{'type':'choice','choice':'a','probabilities':{'a':.9,'b':.1},'confidence':.74},
+            'score':{'type':'score','score':1.15,'legend':{'0':'low','1':'medium','2':'high'},'probabilities':{'0':0.,'1':.85,'2':.15},'confidence':.77},
+            'yes':{'type':'noul','noul':.91}},'usage':{'input_tokens':300,'output_tokens':25,'cost':.0000126}}
+        self.assertIs(validate_oracle_response(response,questions),response)
+        with tempfile.TemporaryDirectory() as d,patch('decision_server.Worker',WorkerStub):
+            cfg={'state_dir':d,'binary':'unused','seed_model':'unused','families':{},'oracle':{'enabled':True,'budget_usd':1,'max_calls':1}}
+            e=Engine(cfg,key='fixture')
+            request={'model':MODEL,'state':{'message':'public fixture'},'questions':questions,'cmf':{'allow_oracle':True,'approved_for_external':True,'consent_store':True}}
+            with patch('decision_server.JevNativeClient.call',return_value={'response':response,'usage':response['usage']}) as call:
+                result=evaluate(e,request);self.assertEqual(call.call_count,1)
+                wire=json.loads(call.call_args.args[0]);self.assertEqual(wire['questions'],questions);self.assertEqual(wire['state'],request['state'])
+                self.assertEqual(result['answers'],response['answers'])
+                self.assertEqual(e.status()['review_queue'],3)
+                with self.assertRaises(ProtocolError) as ctx:evaluate(e,request)
+                self.assertEqual(ctx.exception.body['error']['details']['oracle_reason'],'oracle_budget_exhausted');self.assertEqual(call.call_count,1)
+            self.assertAlmostEqual(e.status()['oracle']['spent_usd'],.0000126);e.db.close()
+
+    def test_invalid_oracle_replies(self):
+        good={'model':'typesafe/jev-1.13','answers':{'q':{'type':'noul','noul':.5}},'usage':{'cost':0.,'input_tokens':1,'output_tokens':1}}
+        for bad in [{**good,'model':'different/model'}, {**good,'answers':{'q':{'type':'noul','noul':float('nan')}}}, {**good,'usage':{'cost':-1}}, {**good,'answers':{}}]:
+            with self.assertRaises(ValueError):validate_oracle_response(bad,{'q':NOUL})
+
+    def test_invalid_later_question_precedes_side_effects(self):
+        with tempfile.TemporaryDirectory() as d,patch('decision_server.Worker',WorkerStub):
+            e=Engine({'state_dir':d,'binary':'unused','seed_model':'unused','families':{},'oracle':{'enabled':False,'budget_usd':1,'max_calls':1}})
+            with self.assertRaises(ProtocolError):evaluate(e,{'model':MODEL,'state':'text','questions':{'one':NOUL,'two':{}},'cmf':{'consent_store':True}})
+            self.assertEqual(e.status()['review_queue'],0);self.assertEqual(e.worker.embeds,0);e.db.close()
+
+
+if __name__=='__main__':unittest.main()

source/tools/cortiq-decision/tests/test_harness.py

@@ -0,0 +1,27 @@
+import importlib.util, math, pathlib, unittest
+ROOT=pathlib.Path(__file__).resolve().parents[1]
+def module(name):
+    spec=importlib.util.spec_from_file_location(name,ROOT/(name+'.py'));m=importlib.util.module_from_spec(spec);spec.loader.exec_module(m);return m
+bench=module('openrouter_bench');summary=module('summarize')
+class HarnessTests(unittest.TestCase):
+    def test_only_pinned_jev_native_request(self):
+        r=bench.request_for('test')
+        self.assertEqual(r['model'],'typesafe/jev-1.13')
+        self.assertNotIn('models',r);self.assertNotIn('route',r)
+        self.assertEqual(set(r['questions']),{'task'})
+        self.assertEqual(r['questions']['task']['type'],'choice')
+        self.assertEqual(bench.ENDPOINT,'/api/alpha/decisions')
+    def test_metrics_known_example(self):
+        r=summary.metrics([{'truth':'a','choice':'a','probabilities':{'a':0.8,'b':0.2}}, {'truth':'b','choice':'a','probabilities':{'a':0.6,'b':0.4}}],['a','b'])
+        self.assertEqual(r['accuracy'],0.5)
+        self.assertAlmostEqual(r['macro_f1'],1/3)
+        self.assertAlmostEqual(r['brier_multiclass_sum'],0.4)
+        self.assertAlmostEqual(r['ece_10_bins_on_max_probability'],0.4)
+        self.assertLess(r['accuracy_wilson95'][0],0.5);self.assertGreater(r['accuracy_wilson95'][1],0.5)
+    def test_perfect_predictions(self):
+        r=summary.metrics([{'truth':'a','choice':'a','probabilities':{'a':1.}}],['a'])
+        self.assertEqual(r['accuracy'],1.);self.assertEqual(r['macro_f1'],1.)
+        self.assertEqual(r['brier_multiclass_sum'],0.);self.assertEqual(r['ece_10_bins_on_max_probability'],0.)
+    def test_percentiles(self):
+        self.assertEqual(summary.percentiles([1000.,3000.,2000.]),{'p50_ms':2.,'p95_ms':3.,'p99_ms':3.})
+if __name__=='__main__':unittest.main()

source/tools/cortiq-decision/tests/test_server.py

@@ -0,0 +1,66 @@
+import sys,unittest,tempfile,json
+from pathlib import Path
+from unittest.mock import patch
+sys.path.insert(0,str(Path(__file__).resolve().parents[1]))
+from decision_server import contract,contract_id,state_text,execute,Engine,accepts
+from lifecycle import classification_metrics,old_decision_regresses
+class Stub:
+    rep='fixture-rep'
+    def __init__(self,*args):pass
+    def call(self,**args):return {'metadata':{'model_version':'fixture'},'representation':{}}
+    def load(self,*args):return 'fixture'
+    def close(self):pass
+class ServerContracts(unittest.TestCase):
+    def test_choice_order_and_criteria_changes(self):
+        a={'type':'choice','instructions':'Choose','criteria':{'a':'Alpha','b':'Beta'}}
+        b={'type':'choice','instructions':'Choose','criteria':{'b':'Beta','a':'Alpha'}}
+        self.assertEqual(contract_id(contract(a)),contract_id(contract(b)))
+        b['criteria']['a']='NOT Alpha';self.assertNotEqual(contract_id(contract(a)),contract_id(contract(b)))
+    def test_score_order_is_semantic(self):
+        a=contract({'type':'score','instructions':'severity','criteria':['low','high']});b=contract({'type':'score','instructions':'severity','criteria':['high','low']})
+        self.assertNotEqual(contract_id(a),contract_id(b));self.assertEqual(contract(a),a)
+        self.assertEqual(execute(a,'1',{'0':.1,'1':.9})['score'],1)
+    def test_pair_cannot_silently_become_plain_text(self):
+        c=contract({'type':'binary','instructions':'relevant?','criteria':{'false':'no','true':'yes'},'input_template':'qa-pair-v1'})
+        with self.assertRaises(ValueError):state_text('plain text',c)
+        self.assertIn('[DOCUMENT]',state_text({'query':'q','document':'d'},c))
+        with self.assertRaises(ValueError):execute(c,'other',{'false':.1,'true':.9})
+    def test_review_transaction_and_durable_holdout(self):
+        with tempfile.TemporaryDirectory() as d,patch('decision_server.Worker',Stub):
+            cfg={'state_dir':d,'binary':'unused','seed_model':'unused','families':{},'oracle':{'enabled':False,'budget_usd':1,'max_calls':10}}
+            e=Engine(cfg);q={'type':'binary','instructions':'yes?','criteria':{'false':'no','true':'yes'}}
+            r=e.decide({'state':'test','questions':{'q':q},'options':{'consent_store':True}})['answers']['q']
+            with self.assertRaises(ValueError):e.review({'request_id':r['request_id'],'label':'true','split':'test','human_verified':True,'training_allowed':True,'label_source':'oracle_verified'})
+            e.review({'request_id':r['request_id'],'label':'true','split':'test','human_verified':True,'training_allowed':True,'label_source':'human_independent'})
+            self.assertEqual(e.status()['persisted_controls'],1);e.db.close()
+            e=Engine(cfg);self.assertEqual(e.status()['persisted_controls'],1)
+            r2=e.decide({'state':'TEST','questions':{'q':q},'options':{'consent_store':True}})['answers']['q']
+            with self.assertRaises(ValueError):e.review({'request_id':r2['request_id'],'label':'true','split':'train','human_verified':True,'training_allowed':True,'label_source':'human_independent'})
+            with self.assertRaises(ValueError):e.promote('unaccepted-job')
+            e.db.close()
+    def test_invalid_settings_never_partially_apply(self):
+        with tempfile.TemporaryDirectory() as d,patch('decision_server.Worker',Stub):
+            cfg={'state_dir':d,'binary':'unused','seed_model':'unused','families':{},'oracle':{'enabled':False,'budget_usd':1,'max_calls':10}}
+            e=Engine(cfg,key='original-test-value')
+            with self.assertRaises(ValueError):e.configure_oracle({'api_key':'changed-test-value','enabled':True,'budget_usd':20})
+            self.assertEqual(e.key,'original-test-value');self.assertFalse(e.config['oracle']['enabled'])
+            with self.assertRaises(ValueError):e.configure_oracle({'enabled':'true'})
+            e.db.close()
+    def test_invalid_second_question_cannot_store_or_bill_first(self):
+        with tempfile.TemporaryDirectory() as d,patch('decision_server.Worker',Stub):
+            cfg={'state_dir':d,'binary':'unused','seed_model':'unused','families':{},'oracle':{'enabled':False,'budget_usd':1,'max_calls':10}}
+            e=Engine(cfg)
+            with self.assertRaises(ValueError):e.decide({'state':'text','questions':{'valid':{'type':'binary','instructions':'yes?','criteria':{'false':'no','true':'yes'}},'invalid':{}},'options':{'consent_store':True}})
+            self.assertEqual(e.status()['review_queue'],0);e.db.close()
+    def test_old_correct_abstention_regression_is_rejected(self):
+        entry={'min_confidence':.8};before={'task_label':'true','confidence':.95,'margin':.1,'novelty_score':.1,'is_novel':False}
+        self.assertTrue(accepts(entry,before))
+        after={**before,'confidence':.7}
+        self.assertTrue(old_decision_regresses(before,after,'true',entry,entry))
+        self.assertTrue(old_decision_regresses(before,{**before,'task_label':'false'},'true',entry,entry))
+        self.assertFalse(old_decision_regresses(before,before,'true',entry,entry))
+    def test_imbalanced_accuracy_cannot_hide_missing_class(self):
+        rows=[{'truth':'major','prediction':'major'}]*99+[{'truth':'rare','prediction':'major'}]
+        result=classification_metrics(rows,{'major','rare'})
+        self.assertEqual(result['minimum_class_recall'],0.);self.assertLess(result['macro_f1'],.5)
+if __name__=='__main__':unittest.main()

source/tools/cortiq-decision/tests/test_support_gate.py

@@ -0,0 +1,11 @@
+import sys,unittest
+from pathlib import Path
+sys.path.insert(0,str(Path(__file__).resolve().parents[1]))
+from decision_server import accepts
+class SupportGate(unittest.TestCase):
+    def test_zero_support_cannot_be_forced_local_by_ignoring_legacy_novelty(self):
+        entry={'min_confidence':0.,'ignore_legacy_novelty':True}
+        p={'confidence':0.,'margin':0.,'novelty_score':.5,'is_novel':True}
+        self.assertFalse(accepts(entry,p))
+        p['confidence']=.9
+        self.assertTrue(accepts(entry,p))

source/tools/cortiq-decision/tests/test_training.py

@@ -0,0 +1,79 @@
+import importlib.util
+import json
+from pathlib import Path
+import tempfile
+import sys
+import unittest
+import numpy as np
+
+ROOT = Path(__file__).resolve().parents[1]
+def module(name):
+    spec = importlib.util.spec_from_file_location(name, ROOT / (name+'.py'))
+    m = importlib.util.module_from_spec(spec); spec.loader.exec_module(m)
+    return m
+train = module('train')
+bench = module('openrouter_bench')
+sys.path.insert(0, str(ROOT))
+oracle = module('oracle')
+
+class TrainingTests(unittest.TestCase):
+    def test_projection_nonorthogonal(self):
+        x = np.array([[2., 3.]], dtype=np.float32)
+        means = [np.array([1., 1.], dtype=np.float32)]
+        basis = [np.array([[.5, .5], [1., 0.]], dtype=np.float32)]
+        r = x[0]-means[0]
+        for b in basis[0]:
+            r = r-(r@b)*b
+        self.assertAlmostEqual(float(train.errors(x, means, basis, 2)[0, 0]), float(r@r))
+    def test_gate_all_correct_and_all_wrong(self):
+        p = np.array([[.999, .001]]*500)
+        best, _ = train.select_gate(p, np.zeros(500, dtype=int), .95, 100)
+        self.assertIsNotNone(best)
+        best, _ = train.select_gate(p, np.ones(500, dtype=int), .95, 100)
+        self.assertIsNone(best)
+    def test_softmax_temperature(self):
+        p = train.probabilities(np.array([[0., 1.]]), .2)
+        self.assertAlmostEqual(float(p.sum()), 1.)
+        self.assertGreater(p[0, 0], .99)
+    def test_reject_duplicate_embeddings(self):
+        with tempfile.TemporaryDirectory() as d:
+            p = Path(d)/'input.jsonl'
+            meta = {'record_type':'representation','rows':2,'representation':{'input_dim':1}}
+            row = {'text_sha256':'same','label':'a','vector':[1.]}
+            p.write_text('\n'.join(map(json.dumps,[meta,row,row])))
+            with self.assertRaises(ValueError): train.read_embeddings(p)
+    def test_custom_taxonomy_does_not_change_model(self):
+        q = {'instructions':'Test intent', 'criteria':{'a':'Class A', 'b':'Class B'}}
+        r = bench.request_for('state', q)
+        self.assertEqual(r['model'], 'typesafe/jev-1.13')
+        self.assertEqual(r['questions']['task']['criteria'], q['criteria'])
+    def test_refuse_substituted_models_and_invalid_cost(self):
+        r = {'model':'typesafe/jev-1.13-20260917','answers':{'task':{'type':'choice','choice':'a','probabilities':{'a':.9,'b':.1}}},'usage':{'cost':.001}}
+        bench.validate_result(r, {'a','b'})
+        r['model'] = 'some/other-model'
+        with self.assertRaises(ValueError): bench.validate_result(r, {'a','b'})
+        r['model'] = 'typesafe/jev-1.13'; r['usage']['cost'] = float('nan')
+        with self.assertRaises(ValueError): bench.validate_result(r, {'a','b'})
+    def test_reservation_and_request_cap(self):
+        self.assertGreater(bench.reservation(b'123', .042), 0)
+        with self.assertRaises(ValueError): bench.reservation(b'x'*32769, .042)
+    def test_oracle_requires_exact_question_binding(self):
+        q = json.dumps({'instructions':'Choose', 'criteria':{'a':'A','b':'B'}}).encode()
+        policy = {'question_sha256':bench.digest(q),'labels':['a','b'],'oracle_model':'typesafe/jev-1.13','auto_train_from_oracle':False}
+        oracle.validate_binding(policy, q)
+        with self.assertRaises(ValueError): oracle.validate_binding(policy, q+b' ')
+        policy['auto_train_from_oracle'] = True
+        with self.assertRaises(ValueError): oracle.validate_binding(policy, q)
+    def test_failures_do_not_get_fake_calibration(self):
+        summary = module('summarize')
+        m = summary.metrics([{'truth':'a','choice':None,'probabilities':None}], ['a','b'])
+        self.assertEqual(m['accuracy'], 0.)
+        self.assertIsNone(m['brier_multiclass_sum'])
+        self.assertIsNone(m['ece_10_bins_on_max_probability'])
+    def test_rounded_probability_sum(self):
+        r = {'model':'typesafe/jev-1.13','answers':{'task':{'type':'choice','choice':'a','probabilities':{'a':.55,'b':.24,'c':.10,'d':.10}}},'usage':{'cost':.001}}
+        bench.validate_result(r, {'a','b','c','d'})
+        r['answers']['task']['probabilities'] = dict.fromkeys(['a','b','c','d'],0.)
+        with self.assertRaises(ValueError): bench.validate_result(r, {'a','b','c','d'})
+
+if __name__ == '__main__': unittest.main()