Release Banking77 Intent Error Predictor v1.0.0

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+model.skops filter=lfs diff=lfs merge=lfs -text
+primary_baseline.skops filter=lfs diff=lfs merge=lfs -text

DATASET_LICENSE.txt

@@ -0,0 +1,395 @@
+Attribution 4.0 International
+
+=======================================================================
+
+Creative Commons Corporation ("Creative Commons") is not a law firm and
+does not provide legal services or legal advice. Distribution of
+Creative Commons public licenses does not create a lawyer-client or
+other relationship. Creative Commons makes its licenses and related
+information available on an "as-is" basis. Creative Commons gives no
+warranties regarding its licenses, any material licensed under their
+terms and conditions, or any related information. Creative Commons
+disclaims all liability for damages resulting from their use to the
+fullest extent possible.
+
+Using Creative Commons Public Licenses
+
+Creative Commons public licenses provide a standard set of terms and
+conditions that creators and other rights holders may use to share
+original works of authorship and other material subject to copyright
+and certain other rights specified in the public license below. The
+following considerations are for informational purposes only, are not
+exhaustive, and do not form part of our licenses.
+
+     Considerations for licensors: Our public licenses are
+     intended for use by those authorized to give the public
+     permission to use material in ways otherwise restricted by
+     copyright and certain other rights. Our licenses are
+     irrevocable. Licensors should read and understand the terms
+     and conditions of the license they choose before applying it.
+     Licensors should also secure all rights necessary before
+     applying our licenses so that the public can reuse the
+     material as expected. Licensors should clearly mark any
+     material not subject to the license. This includes other CC-
+     licensed material, or material used under an exception or
+     limitation to copyright. More considerations for licensors:
+	wiki.creativecommons.org/Considerations_for_licensors
+
+     Considerations for the public: By using one of our public
+     licenses, a licensor grants the public permission to use the
+     licensed material under specified terms and conditions. If
+     the licensor's permission is not necessary for any reason--for
+     example, because of any applicable exception or limitation to
+     copyright--then that use is not regulated by the license. Our
+     licenses grant only permissions under copyright and certain
+     other rights that a licensor has authority to grant. Use of
+     the licensed material may still be restricted for other
+     reasons, including because others have copyright or other
+     rights in the material. A licensor may make special requests,
+     such as asking that all changes be marked or described.
+     Although not required by our licenses, you are encouraged to
+     respect those requests where reasonable. More_considerations
+     for the public:
+	wiki.creativecommons.org/Considerations_for_licensees
+
+=======================================================================
+
+Creative Commons Attribution 4.0 International Public License
+
+By exercising the Licensed Rights (defined below), You accept and agree
+to be bound by the terms and conditions of this Creative Commons
+Attribution 4.0 International Public License ("Public License"). To the
+extent this Public License may be interpreted as a contract, You are
+granted the Licensed Rights in consideration of Your acceptance of
+these terms and conditions, and the Licensor grants You such rights in
+consideration of benefits the Licensor receives from making the
+Licensed Material available under these terms and conditions.
+
+
+Section 1 -- Definitions.
+
+  a. Adapted Material means material subject to Copyright and Similar
+     Rights that is derived from or based upon the Licensed Material
+     and in which the Licensed Material is translated, altered,
+     arranged, transformed, or otherwise modified in a manner requiring
+     permission under the Copyright and Similar Rights held by the
+     Licensor. For purposes of this Public License, where the Licensed
+     Material is a musical work, performance, or sound recording,
+     Adapted Material is always produced where the Licensed Material is
+     synched in timed relation with a moving image.
+
+  b. Adapter's License means the license You apply to Your Copyright
+     and Similar Rights in Your contributions to Adapted Material in
+     accordance with the terms and conditions of this Public License.
+
+  c. Copyright and Similar Rights means copyright and/or similar rights
+     closely related to copyright including, without limitation,
+     performance, broadcast, sound recording, and Sui Generis Database
+     Rights, without regard to how the rights are labeled or
+     categorized. For purposes of this Public License, the rights
+     specified in Section 2(b)(1)-(2) are not Copyright and Similar
+     Rights.
+
+  d. Effective Technological Measures means those measures that, in the
+     absence of proper authority, may not be circumvented under laws
+     fulfilling obligations under Article 11 of the WIPO Copyright
+     Treaty adopted on December 20, 1996, and/or similar international
+     agreements.
+
+  e. Exceptions and Limitations means fair use, fair dealing, and/or
+     any other exception or limitation to Copyright and Similar Rights
+     that applies to Your use of the Licensed Material.
+
+  f. Licensed Material means the artistic or literary work, database,
+     or other material to which the Licensor applied this Public
+     License.
+
+  g. Licensed Rights means the rights granted to You subject to the
+     terms and conditions of this Public License, which are limited to
+     all Copyright and Similar Rights that apply to Your use of the
+     Licensed Material and that the Licensor has authority to license.
+
+  h. Licensor means the individual(s) or entity(ies) granting rights
+     under this Public License.
+
+  i. Share means to provide material to the public by any means or
+     process that requires permission under the Licensed Rights, such
+     as reproduction, public display, public performance, distribution,
+     dissemination, communication, or importation, and to make material
+     available to the public including in ways that members of the
+     public may access the material from a place and at a time
+     individually chosen by them.
+
+  j. Sui Generis Database Rights means rights other than copyright
+     resulting from Directive 96/9/EC of the European Parliament and of
+     the Council of 11 March 1996 on the legal protection of databases,
+     as amended and/or succeeded, as well as other essentially
+     equivalent rights anywhere in the world.
+
+  k. You means the individual or entity exercising the Licensed Rights
+     under this Public License. Your has a corresponding meaning.
+
+
+Section 2 -- Scope.
+
+  a. License grant.
+
+       1. Subject to the terms and conditions of this Public License,
+          the Licensor hereby grants You a worldwide, royalty-free,
+          non-sublicensable, non-exclusive, irrevocable license to
+          exercise the Licensed Rights in the Licensed Material to:
+
+            a. reproduce and Share the Licensed Material, in whole or
+               in part; and
+
+            b. produce, reproduce, and Share Adapted Material.
+
+       2. Exceptions and Limitations. For the avoidance of doubt, where
+          Exceptions and Limitations apply to Your use, this Public
+          License does not apply, and You do not need to comply with
+          its terms and conditions.
+
+       3. Term. The term of this Public License is specified in Section
+          6(a).
+
+       4. Media and formats; technical modifications allowed. The
+          Licensor authorizes You to exercise the Licensed Rights in
+          all media and formats whether now known or hereafter created,
+          and to make technical modifications necessary to do so. The
+          Licensor waives and/or agrees not to assert any right or
+          authority to forbid You from making technical modifications
+          necessary to exercise the Licensed Rights, including
+          technical modifications necessary to circumvent Effective
+          Technological Measures. For purposes of this Public License,
+          simply making modifications authorized by this Section 2(a)
+          (4) never produces Adapted Material.
+
+       5. Downstream recipients.
+
+            a. Offer from the Licensor -- Licensed Material. Every
+               recipient of the Licensed Material automatically
+               receives an offer from the Licensor to exercise the
+               Licensed Rights under the terms and conditions of this
+               Public License.
+
+            b. No downstream restrictions. You may not offer or impose
+               any additional or different terms or conditions on, or
+               apply any Effective Technological Measures to, the
+               Licensed Material if doing so restricts exercise of the
+               Licensed Rights by any recipient of the Licensed
+               Material.
+
+       6. No endorsement. Nothing in this Public License constitutes or
+          may be construed as permission to assert or imply that You
+          are, or that Your use of the Licensed Material is, connected
+          with, or sponsored, endorsed, or granted official status by,
+          the Licensor or others designated to receive attribution as
+          provided in Section 3(a)(1)(A)(i).
+
+  b. Other rights.
+
+       1. Moral rights, such as the right of integrity, are not
+          licensed under this Public License, nor are publicity,
+          privacy, and/or other similar personality rights; however, to
+          the extent possible, the Licensor waives and/or agrees not to
+          assert any such rights held by the Licensor to the limited
+          extent necessary to allow You to exercise the Licensed
+          Rights, but not otherwise.
+
+       2. Patent and trademark rights are not licensed under this
+          Public License.
+
+       3. To the extent possible, the Licensor waives any right to
+          collect royalties from You for the exercise of the Licensed
+          Rights, whether directly or through a collecting society
+          under any voluntary or waivable statutory or compulsory
+          licensing scheme. In all other cases the Licensor expressly
+          reserves any right to collect such royalties.
+
+
+Section 3 -- License Conditions.
+
+Your exercise of the Licensed Rights is expressly made subject to the
+following conditions.
+
+  a. Attribution.
+
+       1. If You Share the Licensed Material (including in modified
+          form), You must:
+
+            a. retain the following if it is supplied by the Licensor
+               with the Licensed Material:
+
+                 i. identification of the creator(s) of the Licensed
+                    Material and any others designated to receive
+                    attribution, in any reasonable manner requested by
+                    the Licensor (including by pseudonym if
+                    designated);
+
+                ii. a copyright notice;
+
+               iii. a notice that refers to this Public License;
+
+                iv. a notice that refers to the disclaimer of
+                    warranties;
+
+                 v. a URI or hyperlink to the Licensed Material to the
+                    extent reasonably practicable;
+
+            b. indicate if You modified the Licensed Material and
+               retain an indication of any previous modifications; and
+
+            c. indicate the Licensed Material is licensed under this
+               Public License, and include the text of, or the URI or
+               hyperlink to, this Public License.
+
+       2. You may satisfy the conditions in Section 3(a)(1) in any
+          reasonable manner based on the medium, means, and context in
+          which You Share the Licensed Material. For example, it may be
+          reasonable to satisfy the conditions by providing a URI or
+          hyperlink to a resource that includes the required
+          information.
+
+       3. If requested by the Licensor, You must remove any of the
+          information required by Section 3(a)(1)(A) to the extent
+          reasonably practicable.
+
+       4. If You Share Adapted Material You produce, the Adapter's
+          License You apply must not prevent recipients of the Adapted
+          Material from complying with this Public License.
+
+
+Section 4 -- Sui Generis Database Rights.
+
+Where the Licensed Rights include Sui Generis Database Rights that
+apply to Your use of the Licensed Material:
+
+  a. for the avoidance of doubt, Section 2(a)(1) grants You the right
+     to extract, reuse, reproduce, and Share all or a substantial
+     portion of the contents of the database;
+
+  b. if You include all or a substantial portion of the database
+     contents in a database in which You have Sui Generis Database
+     Rights, then the database in which You have Sui Generis Database
+     Rights (but not its individual contents) is Adapted Material; and
+
+  c. You must comply with the conditions in Section 3(a) if You Share
+     all or a substantial portion of the contents of the database.
+
+For the avoidance of doubt, this Section 4 supplements and does not
+replace Your obligations under this Public License where the Licensed
+Rights include other Copyright and Similar Rights.
+
+
+Section 5 -- Disclaimer of Warranties and Limitation of Liability.
+
+  a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE
+     EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS
+     AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF
+     ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS,
+     IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION,
+     WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR
+     PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS,
+     ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT
+     KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT
+     ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU.
+
+  b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE
+     TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION,
+     NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT,
+     INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES,
+     COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR
+     USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN
+     ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR
+     DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR
+     IN PART, THIS LIMITATION MAY NOT APPLY TO YOU.
+
+  c. The disclaimer of warranties and limitation of liability provided
+     above shall be interpreted in a manner that, to the extent
+     possible, most closely approximates an absolute disclaimer and
+     waiver of all liability.
+
+
+Section 6 -- Term and Termination.
+
+  a. This Public License applies for the term of the Copyright and
+     Similar Rights licensed here. However, if You fail to comply with
+     this Public License, then Your rights under this Public License
+     terminate automatically.
+
+  b. Where Your right to use the Licensed Material has terminated under
+     Section 6(a), it reinstates:
+
+       1. automatically as of the date the violation is cured, provided
+          it is cured within 30 days of Your discovery of the
+          violation; or
+
+       2. upon express reinstatement by the Licensor.
+
+     For the avoidance of doubt, this Section 6(b) does not affect any
+     right the Licensor may have to seek remedies for Your violations
+     of this Public License.
+
+  c. For the avoidance of doubt, the Licensor may also offer the
+     Licensed Material under separate terms or conditions or stop
+     distributing the Licensed Material at any time; however, doing so
+     will not terminate this Public License.
+
+  d. Sections 1, 5, 6, 7, and 8 survive termination of this Public
+     License.
+
+
+Section 7 -- Other Terms and Conditions.
+
+  a. The Licensor shall not be bound by any additional or different
+     terms or conditions communicated by You unless expressly agreed.
+
+  b. Any arrangements, understandings, or agreements regarding the
+     Licensed Material not stated herein are separate from and
+     independent of the terms and conditions of this Public License.
+
+
+Section 8 -- Interpretation.
+
+  a. For the avoidance of doubt, this Public License does not, and
+     shall not be interpreted to, reduce, limit, restrict, or impose
+     conditions on any use of the Licensed Material that could lawfully
+     be made without permission under this Public License.
+
+  b. To the extent possible, if any provision of this Public License is
+     deemed unenforceable, it shall be automatically reformed to the
+     minimum extent necessary to make it enforceable. If the provision
+     cannot be reformed, it shall be severed from this Public License
+     without affecting the enforceability of the remaining terms and
+     conditions.
+
+  c. No term or condition of this Public License will be waived and no
+     failure to comply consented to unless expressly agreed to by the
+     Licensor.
+
+  d. Nothing in this Public License constitutes or may be interpreted
+     as a limitation upon, or waiver of, any privileges and immunities
+     that apply to the Licensor or You, including from the legal
+     processes of any jurisdiction or authority.
+
+
+=======================================================================
+
+Creative Commons is not a party to its public
+licenses. Notwithstanding, Creative Commons may elect to apply one of
+its public licenses to material it publishes and in those instances
+will be considered the “Licensor.” The text of the Creative Commons
+public licenses is dedicated to the public domain under the CC0 Public
+Domain Dedication. Except for the limited purpose of indicating that
+material is shared under a Creative Commons public license or as
+otherwise permitted by the Creative Commons policies published at
+creativecommons.org/policies, Creative Commons does not authorize the
+use of the trademark "Creative Commons" or any other trademark or logo
+of Creative Commons without its prior written consent including,
+without limitation, in connection with any unauthorized modifications
+to any of its public licenses or any other arrangements,
+understandings, or agreements concerning use of licensed material. For
+the avoidance of doubt, this paragraph does not form part of the
+public licenses.
+
+Creative Commons may be contacted at creativecommons.org.

LICENSE

@@ -0,0 +1,22 @@
+MIT License
+
+Copyright (c) 2026 Ali Zakaee
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+

README.md

@@ -0,0 +1,79 @@
+---
+library_name: sklearn
+pipeline_tag: text-classification
+tags:
+- sklearn
+- scikit-learn
+- skops
+- intent-classification
+- selective-classification
+- error-prediction
+- banking
+- advisory
+- dataset:PolyAI/banking77
+license: mit
+model-index:
+- name: Banking77 Intent Error Predictor
+  results: []
+---
+
+# Banking77 Intent Error Predictor
+
+Maintainer: Ali Zakaee (ITheEqualizer)
+
+This is a compact advisory complement for the included 77-way banking-support intent router. It estimates when that primary router is likely to be wrong so a fixed review budget can be spent on the riskiest requests. It does not approve transactions, make financial decisions, or replace access control or human review.
+
+## Quickstart
+
+```bash
+python -m venv .venv
+. .venv/bin/activate
+python -m pip install -r requirements.txt
+python -m banking_intent_error_predictor.reference_consumer "I was charged twice for one transfer"
+```
+
+The consumer returns either `enqueue_human_review` or an advisory `route_to_<intent>_handler` action and never executes that action.
+
+Positive review example: a short or ambiguous request whose score geometry resembles a known primary-router error. Negative review example: a clear request such as “How do I activate my new card?” when the primary prediction is confident and the learned error risk stays below the reviewed threshold.
+
+## Task and data
+
+The input is one non-empty English query of at most 512 characters. The bundled primary model produces probabilities in the exact 77-label BANKING77 order. The complement transforms those probabilities into class scores, predicted-intent one-hot features, top-score margin, normalized entropy, length, word count, digit count, question-mark presence, and a small negation indicator. It returns an error probability and recommends review at `0.27443790545050933` or above.
+
+BANKING77 contains 13,083 author-released online-banking queries under CC BY 4.0. Data is pinned to source commit `57ec275d8078af65b7731c2a98be812d844a6d6b`; exact file hashes are in `dataset_manifest.json`. Normalized-text hashes create 60/20/20 primary-train, complement-train, and validation partitions. Seven official-test rows overlapping development text are removed; the other 3,073 official-test rows form the untouched lockbox. Duplicate normalized text cannot cross development partitions.
+
+## Architecture and measured results
+
+The primary baseline is word/character TF-IDF plus multinomial logistic regression. The error predictor is histogram gradient boosting with 140 iterations, 15 leaves, learning rate 0.06, minimum leaf size 25, balanced binary loss, and L2 regularization 1.0. This was chosen over publishing another saturated primary intent classifier because the measured downstream decision is which requests deserve review.
+
+At about 20% review on validation, the learned model caught 168 of 210 primary errors (80.0%) versus 158 of 210 (75.2%) for the top-two-margin rule. Routed accuracy was 97.38% at 79.97% coverage versus 96.75% for margin at the same coverage. On the untouched lockbox, it caught 263 of 335 errors (78.51%); 72 errors were missed and 326 correct predictions were unnecessarily reviewed. Routed accuracy was 97.10% at 80.83% coverage.
+
+The contradictory result matters: ranking PR-AUC was 0.485 on validation and 0.520 on lockbox, below margin-only PR-AUC of 0.516 and 0.586. The release claim is therefore limited to the predeclared fixed review-budget operating point, not better global ranking or universal calibration. The lockbox error-recall Wilson 95% interval is approximately 73.8%–82.6%.
+
+## Prediction trace
+
+For “I was charged twice for one transfer,” text is normalized with NFKC, case folding, and whitespace collapse. The primary TF-IDF model produces 77 probabilities and an intent. Those probabilities plus bounded shape features enter the gradient-boosted error model. Its score is compared with the fixed threshold; above threshold the reference consumer requests review, otherwise it exposes the primary intent route.
+
+Model and threshold selection used validation only. The official-test lockbox was evaluated once for the selected specification. A second clean fit produced exactly identical validation and lockbox scores and decisions; serialization round-trip was exact.
+
+## Limitations and integration
+
+- English only; label coverage reflects BANKING77 and is not a bank’s live taxonomy.
+- The complement is specific to the bundled primary model, label order, preprocessing, and versioned checksums.
+- BANKING77 is small and older than current products, policies, fraud patterns, and user language.
+- Error-ranking PR-AUC is worse than the margin baseline even though the fixed-budget decision is better.
+- Review recommendation is advisory. Never use it to approve, deny, block, refund, authenticate, or move money.
+- Monitor review rate, per-intent misses, drift, and calibration on local labeled traffic. Roll back by disabling the complement and using the included margin-only baseline or sending all requests to the existing safe path.
+
+## Reproduction
+
+```bash
+python -m pip install -r requirements-train.txt
+python -m banking_intent_error_predictor.release_train --cache-dir cache --output-dir reproduced --reference reproduction_reference.npz
+```
+
+The command downloads only immutable, checksummed public source files, fits locally, permits at most eight numeric threads, and requires exact decision-score reproduction.
+
+## Licensing
+
+Code and newly fitted model weights are MIT licensed. BANKING77 is CC BY 4.0; attribution and source hashes are in `THIRD_PARTY_NOTICES.md` and `dataset_manifest.json`.

THIRD_PARTY_NOTICES.md

@@ -0,0 +1,11 @@
+# Third-Party Notices
+
+## BANKING77
+
+The training data is BANKING77 from PolyAI, pinned to source commit `57ec275d8078af65b7731c2a98be812d844a6d6b` and licensed under Creative Commons Attribution 4.0 International.
+
+Citation: Iñigo Casanueva, Tadas Temčinas, Daniela Gerz, Matthew Henderson, and Ivan Vulić. “Efficient Intent Detection with Dual Sentence Encoders.” NLP for ConvAI, 2020.
+
+Source: https://github.com/PolyAI-LDN/task-specific-datasets/tree/57ec275d8078af65b7731c2a98be812d844a6d6b/banking_data
+
+License: https://github.com/PolyAI-LDN/task-specific-datasets/blob/57ec275d8078af65b7731c2a98be812d844a6d6b/LICENSE

banking_intent_error_predictor/__init__.py

@@ -0,0 +1 @@
+"""Compact error prediction for a BANKING77 intent router."""

banking_intent_error_predictor/predict.py

@@ -0,0 +1,134 @@
+from __future__ import annotations
+
+import hashlib
+import json
+import math
+import unicodedata
+from pathlib import Path
+from typing import Any, cast
+
+import numpy as np
+import skops.io as sio
+
+
+def sha256_file(path: Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as stream:
+        for chunk in iter(lambda: stream.read(1 << 20), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def normalize_text(value: str, maximum_characters: int) -> str:
+    if not isinstance(value, str):
+        raise TypeError("input must be a string")
+    if len(value) > maximum_characters:
+        raise ValueError(f"input must not exceed {maximum_characters} characters")
+    normalized = " ".join(unicodedata.normalize("NFKC", value).casefold().split())
+    if not normalized:
+        raise ValueError("input must not be empty")
+    return normalized
+
+
+def risk_features(probabilities: np.ndarray, texts: list[str]) -> np.ndarray:
+    clipped = np.clip(probabilities, 1e-12, 1.0)
+    ordered = np.sort(clipped, axis=1)
+    top1 = ordered[:, -1]
+    top2 = ordered[:, -2]
+    entropy = -(clipped * np.log(clipped)).sum(axis=1) / math.log(clipped.shape[1])
+    shapes = np.asarray(
+        [
+            [
+                min(len(text), 512) / 512,
+                min(len(text.split()), 100) / 100,
+                min(sum(ch.isdigit() for ch in text), 20) / 20,
+                float("?" in text),
+                float(
+                    any(
+                        token in text.split()
+                        for token in ("not", "no", "never", "wrong")
+                    )
+                ),
+            ]
+            for text in texts
+        ],
+        dtype=np.float64,
+    )
+    predicted_one_hot = np.zeros_like(clipped)
+    predicted_one_hot[np.arange(len(clipped)), np.argmax(clipped, axis=1)] = 1.0
+    return np.column_stack(
+        [clipped, predicted_one_hot, top1, top2, top1 - top2, entropy, shapes]
+    )
+
+
+def _load_checked(path: Path, expected_hash: str) -> dict[str, Any]:
+    if sha256_file(path) != expected_hash:
+        raise RuntimeError(f"artifact checksum mismatch: {path.name}")
+    untrusted = sio.get_untrusted_types(file=path)
+    if untrusted:
+        raise RuntimeError(f"artifact requires untrusted skops types: {untrusted}")
+    return cast(dict[str, Any], sio.load(path, trusted=[]))
+
+
+def predict_text(
+    text: str,
+    *,
+    primary_path: Path = Path("primary_baseline.skops"),
+    risk_path: Path = Path("model.skops"),
+    config_path: Path = Path("config.json"),
+) -> dict[str, Any]:
+    config = cast(dict[str, Any], json.loads(config_path.read_text(encoding="utf-8")))
+    if config.get("schema_version") != "1.0":
+        raise RuntimeError("unsupported configuration schema")
+    normalized = normalize_text(text, int(config["input"]["maximum_characters"]))
+    primary = _load_checked(primary_path, str(config["primary_artifact_sha256"]))
+    candidate = _load_checked(risk_path, str(config["artifact_sha256"]))
+    labels = cast(list[str], config["labels"])
+    if labels != primary["labels"] or labels != candidate["labels"]:
+        raise RuntimeError("label mapping mismatch")
+    probabilities = np.asarray(
+        primary["classifier"].predict_proba(
+            primary["features"].transform([normalized])
+        ),
+        dtype=np.float64,
+    )
+    if probabilities.shape != (1, len(labels)) or not np.isfinite(probabilities).all():
+        raise RuntimeError("primary model returned invalid probabilities")
+    predicted = int(probabilities[0].argmax())
+    error_risk = float(
+        candidate["classifier"].predict_proba(
+            risk_features(probabilities, [normalized])
+        )[0, 1]
+    )
+    threshold = float(config["review_at_or_above_error_risk"])
+    return {
+        "intent": labels[predicted],
+        "intent_score": float(probabilities[0, predicted]),
+        "error_risk": error_risk,
+        "review": error_risk >= threshold,
+        "review_threshold": threshold,
+        "advisory_only": True,
+        "score_semantics": config["score_semantics"],
+    }
+
+
+def main() -> None:
+    import argparse
+
+    parser = argparse.ArgumentParser(description="Route one banking-support query")
+    parser.add_argument("text")
+    parser.add_argument("--primary", type=Path, default=Path("primary_baseline.skops"))
+    parser.add_argument("--risk-model", type=Path, default=Path("model.skops"))
+    parser.add_argument("--config", type=Path, default=Path("config.json"))
+    args = parser.parse_args()
+    result = predict_text(
+        args.text,
+        primary_path=args.primary,
+        risk_path=args.risk_model,
+        config_path=args.config,
+    )
+    print(json.dumps(result, indent=2, sort_keys=True))
+
+
+if __name__ == "__main__":
+    main()

banking_intent_error_predictor/reference_consumer.py

@@ -0,0 +1,51 @@
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+from typing import Any
+
+from banking_intent_error_predictor.predict import predict_text
+
+
+def route_query(
+    text: str,
+    *,
+    primary_path: Path,
+    risk_path: Path,
+    config_path: Path,
+) -> dict[str, Any]:
+    prediction = predict_text(
+        text,
+        primary_path=primary_path,
+        risk_path=risk_path,
+        config_path=config_path,
+    )
+    action = (
+        "enqueue_human_review"
+        if prediction["review"]
+        else f"route_to_{prediction['intent']}_handler"
+    )
+    return {"action": action, "prediction": prediction, "executes_action": False}
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(
+        description="Demonstrate advisory routing without executing an action"
+    )
+    parser.add_argument("text")
+    parser.add_argument("--primary", type=Path, default=Path("primary_baseline.skops"))
+    parser.add_argument("--risk-model", type=Path, default=Path("model.skops"))
+    parser.add_argument("--config", type=Path, default=Path("config.json"))
+    args = parser.parse_args()
+    result = route_query(
+        args.text,
+        primary_path=args.primary,
+        risk_path=args.risk_model,
+        config_path=args.config,
+    )
+    print(json.dumps(result, indent=2, sort_keys=True))
+
+
+if __name__ == "__main__":
+    main()

banking_intent_error_predictor/release_train.py

@@ -0,0 +1,81 @@
+from __future__ import annotations
+
+import argparse
+import json
+import resource
+import time
+from pathlib import Path
+
+import numpy as np
+import skops.io as sio
+from threadpoolctl import threadpool_limits
+
+from banking_intent_error_predictor.training import (
+    SOURCE_REVISION,
+    download,
+    fit_once,
+    load_rows,
+    sha256_file,
+)
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Reproduce the reviewed release")
+    parser.add_argument("--cache-dir", type=Path, required=True)
+    parser.add_argument("--output-dir", type=Path, required=True)
+    parser.add_argument("--reference", type=Path, required=True)
+    args = parser.parse_args()
+    args.output_dir.mkdir(parents=True, exist_ok=False)
+    train_rows = load_rows(download(args.cache_dir, "train.csv"))
+    test_rows = load_rows(download(args.cache_dir, "test.csv"))
+    download(args.cache_dir, "categories.json")
+    wall_start = time.perf_counter()
+    cpu_start = time.process_time()
+    with threadpool_limits(limits=8):
+        result = fit_once(train_rows, test_rows)
+    reference = result["reproduction_reference"]
+    with np.load(args.reference, allow_pickle=False) as expected:
+        exact = all(
+            np.array_equal(reference[key], expected[key]) for key in expected.files
+        )
+    if not exact:
+        raise RuntimeError("reproduced decision scores differ from v1.0.0")
+    primary_path = args.output_dir / "primary_baseline.skops"
+    candidate_path = args.output_dir / "model.skops"
+    sio.dump(
+        {
+            "features": result["models"]["feature_extractor"],
+            "classifier": result["models"]["primary"],
+            "labels": result["labels"],
+        },
+        primary_path,
+    )
+    sio.dump(
+        {
+            "classifier": result["models"]["candidate"],
+            "labels": result["labels"],
+            "review_rate": 0.20,
+        },
+        candidate_path,
+    )
+    for path in (primary_path, candidate_path):
+        if sio.get_untrusted_types(file=path):
+            raise RuntimeError(f"reproduced artifact requires untrusted types: {path}")
+    report = {
+        "dataset_revision": SOURCE_REVISION,
+        "reference_outputs_exact": exact,
+        "primary_artifact_sha256": sha256_file(primary_path),
+        "candidate_artifact_sha256": sha256_file(candidate_path),
+        "wall_seconds": time.perf_counter() - wall_start,
+        "cpu_seconds": time.process_time() - cpu_start,
+        "thread_limit": 8,
+        "peak_rss_bytes": int(resource.getrusage(resource.RUSAGE_SELF).ru_maxrss),
+    }
+    (args.output_dir / "reproduction.json").write_text(
+        json.dumps(report, indent=2, sort_keys=True) + "\n", encoding="utf-8"
+    )
+    print(json.dumps(report, indent=2, sort_keys=True))
+
+
+if __name__ == "__main__":
+    main()

banking_intent_error_predictor/training.py

@@ -0,0 +1,617 @@
+from __future__ import annotations
+
+import csv
+import hashlib
+import json
+import math
+import os
+import resource
+import time
+import unicodedata
+import urllib.request
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import skops.io as sio
+from sklearn.ensemble import HistGradientBoostingClassifier
+from sklearn.feature_extraction.text import TfidfVectorizer
+from sklearn.linear_model import LogisticRegression
+from sklearn.metrics import (
+    accuracy_score,
+    average_precision_score,
+    confusion_matrix,
+    f1_score,
+    precision_recall_fscore_support,
+    roc_auc_score,
+)
+from sklearn.pipeline import FeatureUnion
+
+CAMPAIGN_ID = "banking77-intent-error-predictor-v1"
+SOURCE_REVISION = "57ec275d8078af65b7731c2a98be812d844a6d6b"
+HUB_REVISION = "90d4e2ee5521c04fc1488f065b8b083658768c57"
+SOURCE_ROOT = (
+    "https://raw.githubusercontent.com/PolyAI-LDN/task-specific-datasets/"
+    f"{SOURCE_REVISION}/banking_data"
+)
+EXPECTED_SHA256 = {
+    "train.csv": "b06e26ac675513959a63135f11b94ea7786ed02da65db93a5650d8838cbc664b",
+    "test.csv": "d12d6e3bc4c3103966ae786dc435913c0c563dfa328f5a3646d0e62cfeeb474d",
+    "categories.json": (
+        "53261da888122daf2d120d925458631d9619e15d82e56052e7a42e535ce32b63"
+    ),
+}
+SEED = 20260811
+THREAD_LIMIT = 8
+REVIEW_RATE = 0.20
+
+
+@dataclass(frozen=True)
+class Row:
+    text: str
+    label: str
+
+
+def sha256_file(path: Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as handle:
+        for block in iter(lambda: handle.read(1 << 20), b""):
+            digest.update(block)
+    return digest.hexdigest()
+
+
+def normalize(text: str) -> str:
+    return " ".join(unicodedata.normalize("NFKC", text).casefold().split())
+
+
+def group_bucket(text: str) -> int:
+    return int(hashlib.sha256(normalize(text).encode()).hexdigest()[:8], 16) % 100
+
+
+def download(cache_dir: Path, filename: str) -> Path:
+    path = cache_dir / "banking77" / SOURCE_REVISION / filename
+    if not path.exists():
+        path.parent.mkdir(parents=True, exist_ok=True)
+        request = urllib.request.Request(
+            f"{SOURCE_ROOT}/{filename}",
+            headers={"User-Agent": "ITheEqualizer-banking77-audit/0.1"},
+        )
+        temporary = path.with_suffix(path.suffix + ".partial")
+        with urllib.request.urlopen(request, timeout=60) as response:
+            temporary.write_bytes(response.read())
+        temporary.replace(path)
+    actual = sha256_file(path)
+    expected = EXPECTED_SHA256[filename]
+    if actual != expected:
+        raise RuntimeError(f"checksum mismatch for {filename}: {actual}")
+    return path
+
+
+def load_rows(path: Path) -> list[Row]:
+    with path.open(newline="", encoding="utf-8") as handle:
+        rows = [
+            Row(text=row["text"], label=row["category"])
+            for row in csv.DictReader(handle)
+        ]
+    if not rows or any(not row.text.strip() or not row.label.strip() for row in rows):
+        raise RuntimeError(f"invalid or empty rows in {path.name}")
+    return rows
+
+
+def build_primary() -> tuple[FeatureUnion, LogisticRegression]:
+    features = FeatureUnion(
+        [
+            (
+                "word",
+                TfidfVectorizer(
+                    ngram_range=(1, 2),
+                    min_df=2,
+                    max_features=24000,
+                    sublinear_tf=True,
+                    strip_accents="unicode",
+                ),
+            ),
+            (
+                "char",
+                TfidfVectorizer(
+                    analyzer="char_wb",
+                    ngram_range=(3, 5),
+                    min_df=2,
+                    max_features=36000,
+                    sublinear_tf=True,
+                    strip_accents="unicode",
+                ),
+            ),
+        ]
+    )
+    model = LogisticRegression(
+        C=4.0,
+        max_iter=600,
+        solver="lbfgs",
+        random_state=SEED,
+    )
+    return features, model
+
+
+def risk_features(probabilities: np.ndarray, texts: list[str]) -> np.ndarray:
+    clipped = np.clip(probabilities, 1e-12, 1.0)
+    ordered = np.sort(clipped, axis=1)
+    top1 = ordered[:, -1]
+    top2 = ordered[:, -2]
+    entropy = -(clipped * np.log(clipped)).sum(axis=1) / math.log(clipped.shape[1])
+    shapes = np.asarray(
+        [
+            [
+                min(len(text), 512) / 512,
+                min(len(text.split()), 100) / 100,
+                min(sum(ch.isdigit() for ch in text), 20) / 20,
+                float("?" in text),
+                float(
+                    any(
+                        token in normalize(text).split()
+                        for token in ("not", "no", "never", "wrong")
+                    )
+                ),
+            ]
+            for text in texts
+        ],
+        dtype=np.float64,
+    )
+    predicted_one_hot = np.zeros_like(clipped)
+    predicted_one_hot[np.arange(len(clipped)), np.argmax(clipped, axis=1)] = 1.0
+    return np.column_stack(
+        [clipped, predicted_one_hot, top1, top2, top1 - top2, entropy, shapes]
+    )
+
+
+def error_metrics(
+    errors: np.ndarray,
+    risks: np.ndarray,
+    *,
+    threshold: float,
+) -> dict[str, Any]:
+    reviewed = risks >= threshold
+    tn, fp, fn, tp = confusion_matrix(errors, reviewed, labels=[0, 1]).ravel()
+    precision, recall, f1, _ = precision_recall_fscore_support(
+        errors, reviewed, average="binary", zero_division=0
+    )
+    routed = ~reviewed
+    return {
+        "error_prevalence": float(errors.mean()),
+        "roc_auc": float(roc_auc_score(errors, risks)),
+        "pr_auc": float(average_precision_score(errors, risks)),
+        "threshold": float(threshold),
+        "review_rate": float(reviewed.mean()),
+        "error_precision": float(precision),
+        "error_recall": float(recall),
+        "error_f1": float(f1),
+        "confusion": {"tn": int(tn), "fp": int(fp), "fn": int(fn), "tp": int(tp)},
+        "routed_accuracy": float(1.0 - errors[routed].mean()) if routed.any() else 0.0,
+        "coverage": float(routed.mean()),
+    }
+
+
+def review_threshold(risks: np.ndarray) -> float:
+    return float(np.quantile(risks, 1.0 - REVIEW_RATE, method="higher"))
+
+
+def dump_checked(model: Any, path: Path) -> dict[str, Any]:
+    sio.dump(model, path)
+    untrusted = sio.get_untrusted_types(file=path)
+    if untrusted:
+        raise RuntimeError(f"non-empty skops untrusted type set: {untrusted}")
+    return {
+        "path": path.name,
+        "bytes": path.stat().st_size,
+        "sha256": sha256_file(path),
+        "skops_untrusted_types": untrusted,
+    }
+
+
+def fit_once(train_rows: list[Row], test_rows: list[Row]) -> dict[str, Any]:
+    primary_train = [row for row in train_rows if group_bucket(row.text) < 60]
+    complement_train = [row for row in train_rows if 60 <= group_bucket(row.text) < 80]
+    validation = [row for row in train_rows if group_bucket(row.text) >= 80]
+    primary_groups = {normalize(row.text) for row in primary_train}
+    complement_groups = {normalize(row.text) for row in complement_train}
+    validation_groups = {normalize(row.text) for row in validation}
+    if (
+        primary_groups & complement_groups
+        or primary_groups & validation_groups
+        or complement_groups & validation_groups
+    ):
+        raise RuntimeError("group leakage across development partitions")
+    lockbox = [
+        row
+        for row in test_rows
+        if normalize(row.text)
+        not in primary_groups | complement_groups | validation_groups
+    ]
+    labels = sorted({row.label for row in train_rows})
+    if len(labels) != 77 or any(
+        {row.label for row in part} != set(labels)
+        for part in (primary_train, complement_train, validation, lockbox)
+    ):
+        raise RuntimeError("all four partitions must contain all 77 labels")
+
+    features, primary = build_primary()
+    x_primary = features.fit_transform([row.text for row in primary_train])
+    primary.fit(x_primary, [row.label for row in primary_train])
+
+    def primary_outputs(rows: list[Row]) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+        probs = primary.predict_proba(features.transform([row.text for row in rows]))
+        predictions = primary.classes_[np.argmax(probs, axis=1)]
+        truth = np.asarray([row.label for row in rows])
+        return probs, predictions, (predictions != truth).astype(np.int64)
+
+    complement_probs, _, complement_errors = primary_outputs(complement_train)
+    validation_probs, validation_predictions, validation_errors = primary_outputs(
+        validation
+    )
+    lockbox_probs, lockbox_predictions, lockbox_errors = primary_outputs(lockbox)
+
+    candidate = HistGradientBoostingClassifier(
+        learning_rate=0.06,
+        max_iter=140,
+        max_leaf_nodes=15,
+        min_samples_leaf=25,
+        l2_regularization=1.0,
+        class_weight="balanced",
+        random_state=SEED,
+    )
+    candidate.fit(
+        risk_features(complement_probs, [row.text for row in complement_train]),
+        complement_errors,
+    )
+    validation_risk = candidate.predict_proba(
+        risk_features(validation_probs, [row.text for row in validation])
+    )[:, 1]
+    lockbox_risk = candidate.predict_proba(
+        risk_features(lockbox_probs, [row.text for row in lockbox])
+    )[:, 1]
+    validation_margin_risk = (
+        1.0
+        - np.partition(validation_probs, -2, axis=1)[:, -1]
+        + np.partition(validation_probs, -2, axis=1)[:, -2]
+    )
+    lockbox_margin_risk = (
+        1.0
+        - np.partition(lockbox_probs, -2, axis=1)[:, -1]
+        + np.partition(lockbox_probs, -2, axis=1)[:, -2]
+    )
+    candidate_threshold = review_threshold(validation_risk)
+    margin_threshold = review_threshold(validation_margin_risk)
+
+    return {
+        "models": {
+            "feature_extractor": features,
+            "primary": primary,
+            "candidate": candidate,
+        },
+        "partitions": {
+            "primary_train": len(primary_train),
+            "complement_train": len(complement_train),
+            "validation": len(validation),
+            "lockbox": len(lockbox),
+            "lockbox_overlap_removed": len(test_rows) - len(lockbox),
+        },
+        "labels": labels,
+        "primary": {
+            "validation_accuracy": float(
+                accuracy_score(
+                    [row.label for row in validation], validation_predictions
+                )
+            ),
+            "validation_macro_f1": float(
+                f1_score(
+                    [row.label for row in validation],
+                    validation_predictions,
+                    average="macro",
+                )
+            ),
+            "lockbox_accuracy": float(
+                accuracy_score([row.label for row in lockbox], lockbox_predictions)
+            ),
+            "lockbox_macro_f1": float(
+                f1_score(
+                    [row.label for row in lockbox], lockbox_predictions, average="macro"
+                )
+            ),
+        },
+        "validation": {
+            "candidate": error_metrics(
+                validation_errors, validation_risk, threshold=candidate_threshold
+            ),
+            "margin_baseline": error_metrics(
+                validation_errors, validation_margin_risk, threshold=margin_threshold
+            ),
+        },
+        "lockbox": {
+            "candidate": error_metrics(
+                lockbox_errors, lockbox_risk, threshold=candidate_threshold
+            ),
+            "margin_baseline": error_metrics(
+                lockbox_errors, lockbox_margin_risk, threshold=margin_threshold
+            ),
+        },
+        "reproduction_reference": {
+            "validation_candidate_risk": validation_risk,
+            "validation_candidate_predictions": validation_risk >= candidate_threshold,
+            "lockbox_candidate_risk": lockbox_risk,
+            "lockbox_candidate_predictions": lockbox_risk >= candidate_threshold,
+        },
+    }
+
+
+def main() -> None:
+    import argparse
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--output", type=Path, required=True)
+    parser.add_argument("--cache-dir", type=Path, required=True)
+    parser.add_argument("--ledger", type=Path, required=True)
+    parser.add_argument("--state", type=Path, required=True)
+    args = parser.parse_args()
+    args.output.mkdir(parents=True, exist_ok=False)
+    cache_before = (
+        sum(path.stat().st_size for path in args.cache_dir.rglob("*") if path.is_file())
+        if args.cache_dir.exists()
+        else 0
+    )
+    for name in (
+        "OMP_NUM_THREADS",
+        "OPENBLAS_NUM_THREADS",
+        "MKL_NUM_THREADS",
+        "VECLIB_MAXIMUM_THREADS",
+        "NUMEXPR_NUM_THREADS",
+    ):
+        os.environ[name] = str(THREAD_LIMIT)
+    train_path = download(args.cache_dir, "train.csv")
+    test_path = download(args.cache_dir, "test.csv")
+    categories_path = download(args.cache_dir, "categories.json")
+    categories = json.loads(categories_path.read_text(encoding="utf-8"))
+    train_rows = load_rows(train_path)
+    test_rows = load_rows(test_path)
+    if sorted(categories) != sorted({row.label for row in train_rows}):
+        raise RuntimeError("category manifest mismatch")
+
+    code_hash = sha256_file(Path(__file__))
+    specification = {
+        "campaign_id": CAMPAIGN_ID,
+        "candidate": "histogram_gradient_boosting_score_error_predictor",
+        "consumer": "banking-support intent router with a human review queue",
+        "task": "predict whether a fixed BANKING77 primary router is wrong",
+        "input_contract": (
+            "77 probabilities in sorted Banking77 label order plus bounded "
+            "text-shape features"
+        ),
+        "output_contract": "error probability and advisory review decision",
+        "dataset_hub_revision": HUB_REVISION,
+        "dataset_source_revision": SOURCE_REVISION,
+        "dataset_hashes": EXPECTED_SHA256,
+        "preprocessing": (
+            "NFKC casefold whitespace normalization for group hashing; TF-IDF "
+            "primary; score and text-shape candidate features"
+        ),
+        "split": (
+            "normalized-text SHA-256 grouped 60/20/20 development partitions; "
+            "official test untouched lockbox with overlap removal"
+        ),
+        "primary": (
+            "24k word plus 36k character TF-IDF with multinomial logistic "
+            "regression C=4"
+        ),
+        "architecture": (
+            "histogram gradient boosting over probabilities, predicted-intent "
+            "one-hot, confidence geometry, and five text-shape features"
+        ),
+        "objective": "balanced binary log loss for primary-router error prediction",
+        "hyperparameters": {
+            "learning_rate": 0.06,
+            "max_iter": 140,
+            "max_leaf_nodes": 15,
+            "min_samples_leaf": 25,
+            "l2_regularization": 1.0,
+        },
+        "seed": SEED,
+        "code_sha256": code_hash,
+    }
+    spec_hash = hashlib.sha256(
+        json.dumps(specification, sort_keys=True, separators=(",", ":")).encode()
+    ).hexdigest()
+    if any(
+        json.loads(line).get("experiment_spec_hash") == spec_hash
+        for line in args.ledger.read_text(encoding="utf-8").splitlines()
+        if line.strip()
+    ):
+        raise RuntimeError("experiment specification already exists in ledger")
+
+    wall_start = time.perf_counter()
+    cpu_start = time.process_time()
+    usage_start = resource.getrusage(resource.RUSAGE_SELF)
+    result = fit_once(train_rows, test_rows)
+    telemetry = {
+        "wall_seconds": time.perf_counter() - wall_start,
+        "cpu_seconds": time.process_time() - cpu_start,
+        "peak_rss_bytes": resource.getrusage(resource.RUSAGE_SELF).ru_maxrss,
+        "minor_page_fault_delta": resource.getrusage(resource.RUSAGE_SELF).ru_minflt
+        - usage_start.ru_minflt,
+        "thread_limit": THREAD_LIMIT,
+        "single_training_process": True,
+    }
+    primary_bundle = {
+        "features": result["models"]["feature_extractor"],
+        "classifier": result["models"]["primary"],
+        "labels": result["labels"],
+    }
+    primary_artifact = dump_checked(
+        primary_bundle, args.output / "primary_baseline.skops"
+    )
+    candidate_bundle = {
+        "classifier": result["models"]["candidate"],
+        "labels": result["labels"],
+        "review_rate": REVIEW_RATE,
+    }
+    candidate_artifact = dump_checked(candidate_bundle, args.output / "model.skops")
+
+    rerun = fit_once(train_rows, test_rows)
+    exact_scores = all(
+        np.array_equal(
+            result["reproduction_reference"][key], rerun["reproduction_reference"][key]
+        )
+        for key in ("validation_candidate_risk", "lockbox_candidate_risk")
+    )
+    exact_predictions = all(
+        np.array_equal(
+            result["reproduction_reference"][key], rerun["reproduction_reference"][key]
+        )
+        for key in ("validation_candidate_predictions", "lockbox_candidate_predictions")
+    )
+    loaded = sio.load(args.output / "model.skops", trusted=[])
+    serialization_scores_exact = np.array_equal(
+        loaded["classifier"].predict_proba(
+            risk_features(
+                rerun["models"]["primary"].predict_proba(
+                    rerun["models"]["feature_extractor"].transform(
+                        [
+                            row.text
+                            for row in [
+                                r for r in train_rows if group_bucket(r.text) >= 80
+                            ]
+                        ]
+                    )
+                ),
+                [
+                    row.text
+                    for row in [r for r in train_rows if group_bucket(r.text) >= 80]
+                ],
+            )
+        )[:, 1],
+        rerun["reproduction_reference"]["validation_candidate_risk"],
+    )
+    reproduction = {
+        "clean_refit_scores_exact": exact_scores,
+        "clean_refit_predictions_exact": exact_predictions,
+        "serialization_scores_exact": serialization_scores_exact,
+    }
+    validation_candidate = result["validation"]["candidate"]
+    validation_margin = result["validation"]["margin_baseline"]
+    lockbox_candidate = result["lockbox"]["candidate"]
+    acceptance = {
+        "validation_error_recall_margin_delta_at_least_0.02": validation_candidate[
+            "error_recall"
+        ]
+        >= validation_margin["error_recall"] + 0.02,
+        "validation_routed_accuracy_gain_at_least_0.03": validation_candidate[
+            "routed_accuracy"
+        ]
+        >= result["primary"]["validation_accuracy"] + 0.03,
+        "lockbox_error_recall_at_least_0.50": lockbox_candidate["error_recall"] >= 0.50,
+        "artifact_at_most_2_mb": candidate_artifact["bytes"] <= 2_000_000,
+        "empty_skops_untrusted_type_set": not candidate_artifact[
+            "skops_untrusted_types"
+        ],
+        "exact_reproduction": all(reproduction.values()),
+    }
+    decision = (
+        "release_work_pending" if all(acceptance.values()) else "measured_gates_failed"
+    )
+    cache_after = sum(
+        path.stat().st_size for path in args.cache_dir.rglob("*") if path.is_file()
+    )
+    metrics = {
+        "record_type": "fit",
+        "campaign_id": CAMPAIGN_ID,
+        "candidate": specification["candidate"],
+        "hypothesis": (
+            "learned score-shape and intent features improve error capture over "
+            "margin-only triage at equal review rate"
+        ),
+        "experiment_spec_hash": spec_hash,
+        "code_sha256": code_hash,
+        "dataset_revision": SOURCE_REVISION,
+        "base_model_revision": None,
+        "configuration": specification,
+        "partitions": result["partitions"],
+        "primary": result["primary"],
+        "validation": result["validation"],
+        "lockbox": result["lockbox"],
+        "artifacts": {
+            "primary_baseline": primary_artifact,
+            "candidate": candidate_artifact,
+        },
+        "acceptance": acceptance,
+        "acceptance_decision": decision,
+        "reproduction": reproduction,
+        "telemetry": {**telemetry, "cache_growth_bytes": cache_after - cache_before},
+    }
+    (args.output / "metrics.json").write_text(
+        json.dumps(metrics, indent=2, sort_keys=True) + "\n", encoding="utf-8"
+    )
+    temporary_disk_growth = sum(
+        path.stat().st_size for path in args.output.rglob("*") if path.is_file()
+    )
+    metrics["telemetry"]["temporary_disk_growth_bytes"] = temporary_disk_growth
+    (args.output / "metrics.json").write_text(
+        json.dumps(metrics, indent=2, sort_keys=True) + "\n", encoding="utf-8"
+    )
+    ledger_record = {**metrics, "telemetry": metrics["telemetry"]}
+    with args.ledger.open("a", encoding="utf-8") as handle:
+        handle.write(
+            json.dumps(ledger_record, sort_keys=True, separators=(",", ":")) + "\n"
+        )
+        handle.flush()
+        os.fsync(handle.fileno())
+    state = {
+        "campaign_id": CAMPAIGN_ID,
+        "created_at": "2026-08-11T13:25:00Z",
+        "status": decision,
+        "consumer": specification["consumer"],
+        "task": specification["task"],
+        "differentiator": (
+            "sub-2 MB model-specific selective-routing complement rather than "
+            "another primary intent classifier"
+        ),
+        "trend_snapshot": "campaign/trends/20260811T1300Z-v2.json",
+        "preflight": "campaign/trends/20260811T1325Z-banking-error-preflight.json",
+        "current_experiment": {
+            "candidate": specification["candidate"],
+            "experiment_spec_hash": spec_hash,
+            "hypothesis": ledger_record["hypothesis"],
+        },
+        "last_run": metrics,
+        "release_blocker": None
+        if decision == "release_work_pending"
+        else "learned error risk did not clear every predeclared usefulness gate",
+        "next_action": (
+            "run one error-driven candidate using class-conditional calibration "
+            "and confusion-neighborhood features, selected on validation only; "
+            "abandon if it still fails to beat margin triage"
+        ),
+    }
+    temporary_state = args.state.with_suffix(".json.tmp")
+    temporary_state.write_text(
+        json.dumps(state, indent=2, sort_keys=True) + "\n", encoding="utf-8"
+    )
+    temporary_state.replace(args.state)
+    print(
+        json.dumps(
+            {
+                "campaign_id": CAMPAIGN_ID,
+                "decision": decision,
+                "spec_hash": spec_hash,
+                "candidate_artifact": candidate_artifact,
+                "validation": result["validation"],
+                "lockbox": result["lockbox"],
+                "reproduction": reproduction,
+                "telemetry": metrics["telemetry"],
+            },
+            indent=2,
+            sort_keys=True,
+        )
+    )
+
+
+if __name__ == "__main__":
+    main()

checksums.sha256

@@ -0,0 +1,19 @@
+7e7170e3cebf88a9f60c7b8421418323c09304da1af4d5e90f4da1dc1c8a2661  DATASET_LICENSE.txt
+467416515c08b60ef9ab26725c9b5274ef0c9a27d27df3c57809354585d52be9  LICENSE
+ee2b7da6c94b4ecc06d9dcee1c13bb8f1bafec0f1f5871e7250c9f79c28c1220  README.md
+d631b5f1c5bf3c0eed96fd8c3a54de9bca034b6b30c4158663a94e821d507f64  THIRD_PARTY_NOTICES.md
+90b33dfe9e877ffe98f36b1ef1f12b067ee80790e42cc80a048df7b55d996258  banking_intent_error_predictor/__init__.py
+4959067362dc69d8d64650eb778a9c96eea591f13a01497882d4edcfd5b83bb3  banking_intent_error_predictor/predict.py
+9a80f230ebc6ebf575d2c61cef3ceb0fac4cebdc517b7c0ba99426ac784dc5ed  banking_intent_error_predictor/reference_consumer.py
+5cf2da52441f87a132e792029153bed74788b0a08d4a99d92af936fb42649644  banking_intent_error_predictor/release_train.py
+852699c3ee6b8e545f9f0fd6c4cf8ccce0bccdea3d67ae4ceceefcdda1404bd7  banking_intent_error_predictor/training.py
+66f01a807155f28f117d18eaf4b28d75569d16e4e30020de1f40a0009537234b  config.json
+68577be2950e9c723d364651441d5e698791e5c00c3cc9bd77f37022a174a6a9  dataset_manifest.json
+047ac38d790bd894ca21cc141db044df79bd0fab42ad9658411eab4601c63c02  example.json
+d277fd9c63eada32626aba8d738613871f0a87c1cdc89cd10aae1bb5c4347d9c  metrics.json
+1d5a785c69b01135981c9052aa8d124cf3d423f7d15e558820502d37a18686b0  model.skops
+139c47e42e634b9e0e8b1a787e89c378a7175ebce761e5618a3deb74476b6bd2  model_manifest.json
+f735dfa1e498fef3c6a0d0a87664cd2e77acaadc892af7875c7ab30148a0e020  primary_baseline.skops
+5c267e1874f2568facf64893e1bd1acd56b922a61e592e5014080da6897d488f  reproduction_reference.npz
+f4103e1d55fd231fb632fb5b2bf232c719b9d881fcfbeaf5849ba687eb0a8c2e  requirements-train.txt
+6e4faa6fa81cd15034c467bf1a78aed2dcb31f91705d643314a420c7e6e69bd7  requirements.txt

config.json

@@ -0,0 +1,95 @@
+{
+  "advisory_only": true,
+  "artifact": "model.skops",
+  "artifact_sha256": "1d5a785c69b01135981c9052aa8d124cf3d423f7d15e558820502d37a18686b0",
+  "input": {
+    "language": "en",
+    "maximum_characters": 512,
+    "minimum_characters": 1,
+    "type": "string"
+  },
+  "labels": [
+    "Refund_not_showing_up",
+    "activate_my_card",
+    "age_limit",
+    "apple_pay_or_google_pay",
+    "atm_support",
+    "automatic_top_up",
+    "balance_not_updated_after_bank_transfer",
+    "balance_not_updated_after_cheque_or_cash_deposit",
+    "beneficiary_not_allowed",
+    "cancel_transfer",
+    "card_about_to_expire",
+    "card_acceptance",
+    "card_arrival",
+    "card_delivery_estimate",
+    "card_linking",
+    "card_not_working",
+    "card_payment_fee_charged",
+    "card_payment_not_recognised",
+    "card_payment_wrong_exchange_rate",
+    "card_swallowed",
+    "cash_withdrawal_charge",
+    "cash_withdrawal_not_recognised",
+    "change_pin",
+    "compromised_card",
+    "contactless_not_working",
+    "country_support",
+    "declined_card_payment",
+    "declined_cash_withdrawal",
+    "declined_transfer",
+    "direct_debit_payment_not_recognised",
+    "disposable_card_limits",
+    "edit_personal_details",
+    "exchange_charge",
+    "exchange_rate",
+    "exchange_via_app",
+    "extra_charge_on_statement",
+    "failed_transfer",
+    "fiat_currency_support",
+    "get_disposable_virtual_card",
+    "get_physical_card",
+    "getting_spare_card",
+    "getting_virtual_card",
+    "lost_or_stolen_card",
+    "lost_or_stolen_phone",
+    "order_physical_card",
+    "passcode_forgotten",
+    "pending_card_payment",
+    "pending_cash_withdrawal",
+    "pending_top_up",
+    "pending_transfer",
+    "pin_blocked",
+    "receiving_money",
+    "request_refund",
+    "reverted_card_payment?",
+    "supported_cards_and_currencies",
+    "terminate_account",
+    "top_up_by_bank_transfer_charge",
+    "top_up_by_card_charge",
+    "top_up_by_cash_or_cheque",
+    "top_up_failed",
+    "top_up_limits",
+    "top_up_reverted",
+    "topping_up_by_card",
+    "transaction_charged_twice",
+    "transfer_fee_charged",
+    "transfer_into_account",
+    "transfer_not_received_by_recipient",
+    "transfer_timing",
+    "unable_to_verify_identity",
+    "verify_my_identity",
+    "verify_source_of_funds",
+    "verify_top_up",
+    "virtual_card_not_working",
+    "visa_or_mastercard",
+    "why_verify_identity",
+    "wrong_amount_of_cash_received",
+    "wrong_exchange_rate_for_cash_withdrawal"
+  ],
+  "primary_artifact": "primary_baseline.skops",
+  "primary_artifact_sha256": "f735dfa1e498fef3c6a0d0a87664cd2e77acaadc892af7875c7ab30148a0e020",
+  "review_at_or_above_error_risk": 0.27443790545050933,
+  "schema_version": "1.0",
+  "score_semantics": "model-specific estimate that the bundled primary intent route is wrong"
+}

dataset_manifest.json

@@ -0,0 +1,25 @@
+{
+  "files": {
+    "categories.json": "53261da888122daf2d120d925458631d9619e15d82e56052e7a42e535ce32b63",
+    "test.csv": "d12d6e3bc4c3103966ae786dc435913c0c563dfa328f5a3646d0e62cfeeb474d",
+    "train.csv": "b06e26ac675513959a63135f11b94ea7786ed02da65db93a5650d8838cbc664b"
+  },
+  "hub_revision": "90d4e2ee5521c04fc1488f065b8b083658768c57",
+  "license": "CC BY 4.0",
+  "license_sha256": "7e7170e3cebf88a9f60c7b8421418323c09304da1af4d5e90f4da1dc1c8a2661",
+  "pii_review": "short customer-service queries; no secret patterns were introduced; real-world origin and demographic representativeness are undocumented",
+  "repository": "PolyAI/banking77",
+  "rows": {
+    "official_test": 3080,
+    "train": 10003
+  },
+  "source_repository": "PolyAI-LDN/task-specific-datasets",
+  "source_revision": "57ec275d8078af65b7731c2a98be812d844a6d6b",
+  "split": {
+    "complement_train": 1951,
+    "lockbox": 3073,
+    "lockbox_overlap_removed": 7,
+    "primary_train": 6050,
+    "validation": 2002
+  }
+}

example.json

@@ -0,0 +1,13 @@
+{
+  "expected": {
+    "advisory_only": true,
+    "error_risk": 0.011158242123685077,
+    "intent": "transaction_charged_twice",
+    "intent_score": 0.9511649716326432,
+    "review": false,
+    "review_threshold": 0.27443790545050933,
+    "score_semantics": "model-specific estimate that the bundled primary intent route is wrong"
+  },
+  "input": "I was charged twice for one transfer",
+  "source": "author-written"
+}

metrics.json

@@ -0,0 +1,172 @@
+{
+  "acceptance": {
+    "artifact_at_most_2_mb": true,
+    "empty_skops_untrusted_type_set": true,
+    "exact_reproduction": true,
+    "lockbox_error_recall_at_least_0.50": true,
+    "validation_error_recall_margin_delta_at_least_0.02": true,
+    "validation_routed_accuracy_gain_at_least_0.03": true
+  },
+  "acceptance_decision": "release_work_pending",
+  "artifacts": {
+    "candidate": {
+      "bytes": 1104573,
+      "path": "model.skops",
+      "sha256": "1d5a785c69b01135981c9052aa8d124cf3d423f7d15e558820502d37a18686b0",
+      "skops_untrusted_types": []
+    },
+    "primary_baseline": {
+      "bytes": 28966090,
+      "path": "primary_baseline.skops",
+      "sha256": "f735dfa1e498fef3c6a0d0a87664cd2e77acaadc892af7875c7ab30148a0e020",
+      "skops_untrusted_types": []
+    }
+  },
+  "base_model_revision": null,
+  "campaign_id": "banking77-intent-error-predictor-v1",
+  "candidate": "histogram_gradient_boosting_score_error_predictor",
+  "code_sha256": "852699c3ee6b8e545f9f0fd6c4cf8ccce0bccdea3d67ae4ceceefcdda1404bd7",
+  "configuration": {
+    "architecture": "histogram gradient boosting over probabilities, predicted-intent one-hot, confidence geometry, and five text-shape features",
+    "campaign_id": "banking77-intent-error-predictor-v1",
+    "candidate": "histogram_gradient_boosting_score_error_predictor",
+    "code_sha256": "852699c3ee6b8e545f9f0fd6c4cf8ccce0bccdea3d67ae4ceceefcdda1404bd7",
+    "consumer": "banking-support intent router with a human review queue",
+    "dataset_hashes": {
+      "categories.json": "53261da888122daf2d120d925458631d9619e15d82e56052e7a42e535ce32b63",
+      "test.csv": "d12d6e3bc4c3103966ae786dc435913c0c563dfa328f5a3646d0e62cfeeb474d",
+      "train.csv": "b06e26ac675513959a63135f11b94ea7786ed02da65db93a5650d8838cbc664b"
+    },
+    "dataset_hub_revision": "90d4e2ee5521c04fc1488f065b8b083658768c57",
+    "dataset_source_revision": "57ec275d8078af65b7731c2a98be812d844a6d6b",
+    "hyperparameters": {
+      "l2_regularization": 1.0,
+      "learning_rate": 0.06,
+      "max_iter": 140,
+      "max_leaf_nodes": 15,
+      "min_samples_leaf": 25
+    },
+    "input_contract": "77 probabilities in sorted Banking77 label order plus bounded text-shape features",
+    "objective": "balanced binary log loss for primary-router error prediction",
+    "output_contract": "error probability and advisory review decision",
+    "preprocessing": "NFKC casefold whitespace normalization for group hashing; TF-IDF primary; score and text-shape candidate features",
+    "primary": "24k word plus 36k character TF-IDF with multinomial logistic regression C=4",
+    "seed": 20260811,
+    "split": "normalized-text SHA-256 grouped 60/20/20 development partitions; official test untouched lockbox with overlap removal",
+    "task": "predict whether a fixed BANKING77 primary router is wrong"
+  },
+  "dataset_revision": "57ec275d8078af65b7731c2a98be812d844a6d6b",
+  "experiment_spec_hash": "8a4c7e16a4a0ef3615735c7214e429df5268dc1610612ce6e833aa9e3829ccdc",
+  "hypothesis": "learned score-shape and intent features improve error capture over margin-only triage at equal review rate",
+  "lockbox": {
+    "candidate": {
+      "confusion": {
+        "fn": 72,
+        "fp": 326,
+        "tn": 2412,
+        "tp": 263
+      },
+      "coverage": 0.8083306215424666,
+      "error_f1": 0.5692640692640693,
+      "error_precision": 0.4465195246179966,
+      "error_prevalence": 0.10901399284087211,
+      "error_recall": 0.7850746268656716,
+      "pr_auc": 0.5199652272918066,
+      "review_rate": 0.19166937845753335,
+      "roc_auc": 0.9035912475605902,
+      "routed_accuracy": 0.9710144927536232,
+      "threshold": 0.27443790545050933
+    },
+    "margin_baseline": {
+      "confusion": {
+        "fn": 78,
+        "fp": 366,
+        "tn": 2372,
+        "tp": 257
+      },
+      "coverage": 0.7972665148063781,
+      "error_f1": 0.5365344467640919,
+      "error_precision": 0.41252006420545745,
+      "error_prevalence": 0.10901399284087211,
+      "error_recall": 0.7671641791044777,
+      "pr_auc": 0.5863979952872228,
+      "review_rate": 0.20273348519362186,
+      "roc_auc": 0.9092310543702234,
+      "routed_accuracy": 0.9681632653061224,
+      "threshold": 0.6936486958535751
+    }
+  },
+  "partitions": {
+    "complement_train": 1951,
+    "lockbox": 3073,
+    "lockbox_overlap_removed": 7,
+    "primary_train": 6050,
+    "validation": 2002
+  },
+  "primary": {
+    "lockbox_accuracy": 0.8909860071591279,
+    "lockbox_macro_f1": 0.8910051474654282,
+    "validation_accuracy": 0.8951048951048951,
+    "validation_macro_f1": 0.8919223884704242
+  },
+  "record_type": "fit",
+  "release_validation": {
+    "bounded_inference_api": true,
+    "lockbox_scores_persisted": true,
+    "reference_consumer": true,
+    "representative_example_passed": true
+  },
+  "reproduction": {
+    "clean_refit_predictions_exact": true,
+    "clean_refit_scores_exact": true,
+    "serialization_scores_exact": true
+  },
+  "telemetry": {
+    "cache_growth_bytes": 1081070,
+    "cpu_seconds": 14.953067,
+    "minor_page_fault_delta": 37952,
+    "peak_rss_bytes": 743817216,
+    "single_training_process": true,
+    "temporary_disk_growth_bytes": 30076757,
+    "thread_limit": 8,
+    "wall_seconds": 6.962122584000099
+  },
+  "validation": {
+    "candidate": {
+      "confusion": {
+        "fn": 42,
+        "fp": 233,
+        "tn": 1559,
+        "tp": 168
+      },
+      "coverage": 0.7997002997002997,
+      "error_f1": 0.5499181669394435,
+      "error_precision": 0.41895261845386533,
+      "error_prevalence": 0.1048951048951049,
+      "error_recall": 0.8,
+      "pr_auc": 0.48498408831658685,
+      "review_rate": 0.2002997002997003,
+      "roc_auc": 0.8968856292517007,
+      "routed_accuracy": 0.9737663960024985,
+      "threshold": 0.27443790545050933
+    },
+    "margin_baseline": {
+      "confusion": {
+        "fn": 52,
+        "fp": 243,
+        "tn": 1549,
+        "tp": 158
+      },
+      "coverage": 0.7997002997002997,
+      "error_f1": 0.5171849427168577,
+      "error_precision": 0.3940149625935162,
+      "error_prevalence": 0.1048951048951049,
+      "error_recall": 0.7523809523809524,
+      "pr_auc": 0.5163906717576305,
+      "review_rate": 0.2002997002997003,
+      "roc_auc": 0.8986208545918366,
+      "routed_accuracy": 0.9675202998126171,
+      "threshold": 0.6936486958535751
+    }
+  }
+}

model.skops

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1d5a785c69b01135981c9052aa8d124cf3d423f7d15e558820502d37a18686b0
+size 1104573

model_manifest.json

@@ -0,0 +1,99 @@
+{
+  "artifact": "model.skops",
+  "artifact_bytes": 1104573,
+  "artifact_sha256": "1d5a785c69b01135981c9052aa8d124cf3d423f7d15e558820502d37a18686b0",
+  "base_model_revision": null,
+  "campaign_id": "banking77-intent-error-predictor-v1",
+  "dataset_revision": "57ec275d8078af65b7731c2a98be812d844a6d6b",
+  "experiment_spec_hash": "8a4c7e16a4a0ef3615735c7214e429df5268dc1610612ce6e833aa9e3829ccdc",
+  "labels": [
+    "Refund_not_showing_up",
+    "activate_my_card",
+    "age_limit",
+    "apple_pay_or_google_pay",
+    "atm_support",
+    "automatic_top_up",
+    "balance_not_updated_after_bank_transfer",
+    "balance_not_updated_after_cheque_or_cash_deposit",
+    "beneficiary_not_allowed",
+    "cancel_transfer",
+    "card_about_to_expire",
+    "card_acceptance",
+    "card_arrival",
+    "card_delivery_estimate",
+    "card_linking",
+    "card_not_working",
+    "card_payment_fee_charged",
+    "card_payment_not_recognised",
+    "card_payment_wrong_exchange_rate",
+    "card_swallowed",
+    "cash_withdrawal_charge",
+    "cash_withdrawal_not_recognised",
+    "change_pin",
+    "compromised_card",
+    "contactless_not_working",
+    "country_support",
+    "declined_card_payment",
+    "declined_cash_withdrawal",
+    "declined_transfer",
+    "direct_debit_payment_not_recognised",
+    "disposable_card_limits",
+    "edit_personal_details",
+    "exchange_charge",
+    "exchange_rate",
+    "exchange_via_app",
+    "extra_charge_on_statement",
+    "failed_transfer",
+    "fiat_currency_support",
+    "get_disposable_virtual_card",
+    "get_physical_card",
+    "getting_spare_card",
+    "getting_virtual_card",
+    "lost_or_stolen_card",
+    "lost_or_stolen_phone",
+    "order_physical_card",
+    "passcode_forgotten",
+    "pending_card_payment",
+    "pending_cash_withdrawal",
+    "pending_top_up",
+    "pending_transfer",
+    "pin_blocked",
+    "receiving_money",
+    "request_refund",
+    "reverted_card_payment?",
+    "supported_cards_and_currencies",
+    "terminate_account",
+    "top_up_by_bank_transfer_charge",
+    "top_up_by_card_charge",
+    "top_up_by_cash_or_cheque",
+    "top_up_failed",
+    "top_up_limits",
+    "top_up_reverted",
+    "topping_up_by_card",
+    "transaction_charged_twice",
+    "transfer_fee_charged",
+    "transfer_into_account",
+    "transfer_not_received_by_recipient",
+    "transfer_timing",
+    "unable_to_verify_identity",
+    "verify_my_identity",
+    "verify_source_of_funds",
+    "verify_top_up",
+    "virtual_card_not_working",
+    "visa_or_mastercard",
+    "why_verify_identity",
+    "wrong_amount_of_cash_received",
+    "wrong_exchange_rate_for_cash_withdrawal"
+  ],
+  "lockbox_error_recall": 0.7850746268656716,
+  "lockbox_routed_accuracy": 0.9710144927536232,
+  "maintainer": "Ali Zakaee (ITheEqualizer)",
+  "model_name": "Banking77 Intent Error Predictor",
+  "model_version": "1.0.0",
+  "primary_artifact_bytes": 28966090,
+  "primary_artifact_sha256": "f735dfa1e498fef3c6a0d0a87664cd2e77acaadc892af7875c7ab30148a0e020",
+  "release_summary": "New Banking77 primary-router error predictor with newly fitted weights, a fixed 20 percent review-budget contract, grouped development splits, an untouched official-test lockbox, and exact clean reproduction.",
+  "schema_version": "1.0",
+  "trusted_types": [],
+  "validation_error_recall": 0.8
+}

primary_baseline.skops

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f735dfa1e498fef3c6a0d0a87664cd2e77acaadc892af7875c7ab30148a0e020
+size 28966090

reproduction_reference.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c267e1874f2568facf64893e1bd1acd56b922a61e592e5014080da6897d488f
+size 40108

requirements-train.txt

@@ -0,0 +1,4 @@
+numpy==2.5.1
+scikit-learn==1.9.0
+skops==0.14.0
+threadpoolctl==3.6.0

requirements.txt

@@ -0,0 +1,3 @@
+numpy==2.5.1
+scikit-learn==1.9.0
+skops==0.14.0