Metadata-Version: 2.4
Name: orbit-eval
Version: 0.8.2
Summary: Honest statistics for robot-policy evaluation: CRN-paired checkpoint comparisons vs retrain-level method comparisons, shipping the measured ORBIT noise atlas.
Author: ORBIT Research
License:                     GNU AFFERO GENERAL PUBLIC LICENSE
                               Version 3, 19 November 2007
        
         Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
         Everyone is permitted to copy and distribute verbatim copies
         of this license document, but changing it is not allowed.
        
                                    Preamble
        
          The GNU Affero General Public License is a free, copyleft license for
        software and other kinds of works, specifically designed to ensure
        cooperation with the community in the case of network server software.
        
          The licenses for most software and other practical works are designed
        to take away your freedom to share and change the works.  By contrast,
        our General Public Licenses are intended to guarantee your freedom to
        share and change all versions of a program--to make sure it remains free
        software for all its users.
        
          When we speak of free software, we are referring to freedom, not
        price.  Our General Public Licenses are designed to make sure that you
        have the freedom to distribute copies of free software (and charge for
        them if you wish), that you receive source code or can get it if you
        want it, that you can change the software or use pieces of it in new
        free programs, and that you know you can do these things.
        
          Developers that use our General Public Licenses protect your rights
        with two steps: (1) assert copyright on the software, and (2) offer
        you this License which gives you legal permission to copy, distribute
        and/or modify the software.
        
          A secondary benefit of defending all users' freedom is that
        improvements made in alternate versions of the program, if they
        receive widespread use, become available for other developers to
        incorporate.  Many developers of free software are heartened and
        encouraged by the resulting cooperation.  However, in the case of
        software used on network servers, this result may fail to come about.
        The GNU General Public License permits making a modified version and
        letting the public access it on a server without ever releasing its
        source code to the public.
        
          The GNU Affero General Public License is designed specifically to
        ensure that, in such cases, the modified source code becomes available
        to the community.  It requires the operator of a network server to
        provide the source code of the modified version running there to the
        users of that server.  Therefore, public use of a modified version, on
        a publicly accessible server, gives the public access to the source
        code of the modified version.
        
          An older license, called the Affero General Public License and
        published by Affero, was designed to accomplish similar goals.  This is
        a different license, not a version of the Affero GPL, but Affero has
        released a new version of the Affero GPL which permits relicensing under
        this license.
        
          The precise terms and conditions for copying, distribution and
        modification follow.
        
                               TERMS AND CONDITIONS
        
          0. Definitions.
        
          "This License" refers to version 3 of the GNU Affero General Public License.
        
          "Copyright" also means copyright-like laws that apply to other kinds of
        works, such as semiconductor masks.
        
          "The Program" refers to any copyrightable work licensed under this
        License.  Each licensee is addressed as "you".  "Licensees" and
        "recipients" may be individuals or organizations.
        
          To "modify" a work means to copy from or adapt all or part of the work
        in a fashion requiring copyright permission, other than the making of an
        exact copy.  The resulting work is called a "modified version" of the
        earlier work or a work "based on" the earlier work.
        
          A "covered work" means either the unmodified Program or a work based
        on the Program.
        
          To "propagate" a work means to do anything with it that, without
        permission, would make you directly or secondarily liable for
        infringement under applicable copyright law, except executing it on a
        computer or modifying a private copy.  Propagation includes copying,
        distribution (with or without modification), making available to the
        public, and in some countries other activities as well.
        
          To "convey" a work means any kind of propagation that enables other
        parties to make or receive copies.  Mere interaction with a user through
        a computer network, with no transfer of a copy, is not conveying.
        
          An interactive user interface displays "Appropriate Legal Notices"
        to the extent that it includes a convenient and prominently visible
        feature that (1) displays an appropriate copyright notice, and (2)
        tells the user that there is no warranty for the work (except to the
        extent that warranties are provided), that licensees may convey the
        work under this License, and how to view a copy of this License.  If
        the interface presents a list of user commands or options, such as a
        menu, a prominent item in the list meets this criterion.
        
          1. Source Code.
        
          The "source code" for a work means the preferred form of the work
        for making modifications to it.  "Object code" means any non-source
        form of a work.
        
          A "Standard Interface" means an interface that either is an official
        standard defined by a recognized standards body, or, in the case of
        interfaces specified for a particular programming language, one that
        is widely used among developers working in that language.
        
          The "System Libraries" of an executable work include anything, other
        than the work as a whole, that (a) is included in the normal form of
        packaging a Major Component, but which is not part of that Major
        Component, and (b) serves only to enable use of the work with that
        Major Component, or to implement a Standard Interface for which an
        implementation is available to the public in source code form.  A
        "Major Component", in this context, means a major essential component
        (kernel, window system, and so on) of the specific operating system
        (if any) on which the executable work runs, or a compiler used to
        produce the work, or an object code interpreter used to run it.
        
          The "Corresponding Source" for a work in object code form means all
        the source code needed to generate, install, and (for an executable
        work) run the object code and to modify the work, including scripts to
        control those activities.  However, it does not include the work's
        System Libraries, or general-purpose tools or generally available free
        programs which are used unmodified in performing those activities but
        which are not part of the work.  For example, Corresponding Source
        includes interface definition files associated with source files for
        the work, and the source code for shared libraries and dynamically
        linked subprograms that the work is specifically designed to require,
        such as by intimate data communication or control flow between those
        subprograms and other parts of the work.
        
          The Corresponding Source need not include anything that users
        can regenerate automatically from other parts of the Corresponding
        Source.
        
          The Corresponding Source for a work in source code form is that
        same work.
        
          2. Basic Permissions.
        
          All rights granted under this License are granted for the term of
        copyright on the Program, and are irrevocable provided the stated
        conditions are met.  This License explicitly affirms your unlimited
        permission to run the unmodified Program.  The output from running a
        covered work is covered by this License only if the output, given its
        content, constitutes a covered work.  This License acknowledges your
        rights of fair use or other equivalent, as provided by copyright law.
        
          You may make, run and propagate covered works that you do not
        convey, without conditions so long as your license otherwise remains
        in force.  You may convey covered works to others for the sole purpose
        of having them make modifications exclusively for you, or provide you
        with facilities for running those works, provided that you comply with
        the terms of this License in conveying all material for which you do
        not control copyright.  Those thus making or running the covered works
        for you must do so exclusively on your behalf, under your direction
        and control, on terms that prohibit them from making any copies of
        your copyrighted material outside their relationship with you.
        
          Conveying under any other circumstances is permitted solely under
        the conditions stated below.  Sublicensing is not allowed; section 10
        makes it unnecessary.
        
          3. Protecting Users' Legal Rights From Anti-Circumvention Law.
        
          No covered work shall be deemed part of an effective technological
        measure under any applicable law fulfilling obligations under article
        11 of the WIPO copyright treaty adopted on 20 December 1996, or
        similar laws prohibiting or restricting circumvention of such
        measures.
        
          When you convey a covered work, you waive any legal power to forbid
        circumvention of technological measures to the extent such circumvention
        is effected by exercising rights under this License with respect to
        the covered work, and you disclaim any intention to limit operation or
        modification of the work as a means of enforcing, against the work's
        users, your or third parties' legal rights to forbid circumvention of
        technological measures.
        
          4. Conveying Verbatim Copies.
        
          You may convey verbatim copies of the Program's source code as you
        receive it, in any medium, provided that you conspicuously and
        appropriately publish on each copy an appropriate copyright notice;
        keep intact all notices stating that this License and any
        non-permissive terms added in accord with section 7 apply to the code;
        keep intact all notices of the absence of any warranty; and give all
        recipients a copy of this License along with the Program.
        
          You may charge any price or no price for each copy that you convey,
        and you may offer support or warranty protection for a fee.
        
          5. Conveying Modified Source Versions.
        
          You may convey a work based on the Program, or the modifications to
        produce it from the Program, in the form of source code under the
        terms of section 4, provided that you also meet all of these conditions:
        
            a) The work must carry prominent notices stating that you modified
            it, and giving a relevant date.
        
            b) The work must carry prominent notices stating that it is
            released under this License and any conditions added under section
            7.  This requirement modifies the requirement in section 4 to
            "keep intact all notices".
        
            c) You must license the entire work, as a whole, under this
            License to anyone who comes into possession of a copy.  This
            License will therefore apply, along with any applicable section 7
            additional terms, to the whole of the work, and all its parts,
            regardless of how they are packaged.  This License gives no
            permission to license the work in any other way, but it does not
            invalidate such permission if you have separately received it.
        
            d) If the work has interactive user interfaces, each must display
            Appropriate Legal Notices; however, if the Program has interactive
            interfaces that do not display Appropriate Legal Notices, your
            work need not make them do so.
        
          A compilation of a covered work with other separate and independent
        works, which are not by their nature extensions of the covered work,
        and which are not combined with it such as to form a larger program,
        in or on a volume of a storage or distribution medium, is called an
        "aggregate" if the compilation and its resulting copyright are not
        used to limit the access or legal rights of the compilation's users
        beyond what the individual works permit.  Inclusion of a covered work
        in an aggregate does not cause this License to apply to the other
        parts of the aggregate.
        
          6. Conveying Non-Source Forms.
        
          You may convey a covered work in object code form under the terms
        of sections 4 and 5, provided that you also convey the
        machine-readable Corresponding Source under the terms of this License,
        in one of these ways:
        
            a) Convey the object code in, or embodied in, a physical product
            (including a physical distribution medium), accompanied by the
            Corresponding Source fixed on a durable physical medium
            customarily used for software interchange.
        
            b) Convey the object code in, or embodied in, a physical product
            (including a physical distribution medium), accompanied by a
            written offer, valid for at least three years and valid for as
            long as you offer spare parts or customer support for that product
            model, to give anyone who possesses the object code either (1) a
            copy of the Corresponding Source for all the software in the
            product that is covered by this License, on a durable physical
            medium customarily used for software interchange, for a price no
            more than your reasonable cost of physically performing this
            conveying of source, or (2) access to copy the
            Corresponding Source from a network server at no charge.
        
            c) Convey individual copies of the object code with a copy of the
            written offer to provide the Corresponding Source.  This
            alternative is allowed only occasionally and noncommercially, and
            only if you received the object code with such an offer, in accord
            with subsection 6b.
        
            d) Convey the object code by offering access from a designated
            place (gratis or for a charge), and offer equivalent access to the
            Corresponding Source in the same way through the same place at no
            further charge.  You need not require recipients to copy the
            Corresponding Source along with the object code.  If the place to
            copy the object code is a network server, the Corresponding Source
            may be on a different server (operated by you or a third party)
            that supports equivalent copying facilities, provided you maintain
            clear directions next to the object code saying where to find the
            Corresponding Source.  Regardless of what server hosts the
            Corresponding Source, you remain obligated to ensure that it is
            available for as long as needed to satisfy these requirements.
        
            e) Convey the object code using peer-to-peer transmission, provided
            you inform other peers where the object code and Corresponding
            Source of the work are being offered to the general public at no
            charge under subsection 6d.
        
          A separable portion of the object code, whose source code is excluded
        from the Corresponding Source as a System Library, need not be
        included in conveying the object code work.
        
          A "User Product" is either (1) a "consumer product", which means any
        tangible personal property which is normally used for personal, family,
        or household purposes, or (2) anything designed or sold for incorporation
        into a dwelling.  In determining whether a product is a consumer product,
        doubtful cases shall be resolved in favor of coverage.  For a particular
        product received by a particular user, "normally used" refers to a
        typical or common use of that class of product, regardless of the status
        of the particular user or of the way in which the particular user
        actually uses, or expects or is expected to use, the product.  A product
        is a consumer product regardless of whether the product has substantial
        commercial, industrial or non-consumer uses, unless such uses represent
        the only significant mode of use of the product.
        
          "Installation Information" for a User Product means any methods,
        procedures, authorization keys, or other information required to install
        and execute modified versions of a covered work in that User Product from
        a modified version of its Corresponding Source.  The information must
        suffice to ensure that the continued functioning of the modified object
        code is in no case prevented or interfered with solely because
        modification has been made.
        
          If you convey an object code work under this section in, or with, or
        specifically for use in, a User Product, and the conveying occurs as
        part of a transaction in which the right of possession and use of the
        User Product is transferred to the recipient in perpetuity or for a
        fixed term (regardless of how the transaction is characterized), the
        Corresponding Source conveyed under this section must be accompanied
        by the Installation Information.  But this requirement does not apply
        if neither you nor any third party retains the ability to install
        modified object code on the User Product (for example, the work has
        been installed in ROM).
        
          The requirement to provide Installation Information does not include a
        requirement to continue to provide support service, warranty, or updates
        for a work that has been modified or installed by the recipient, or for
        the User Product in which it has been modified or installed.  Access to a
        network may be denied when the modification itself materially and
        adversely affects the operation of the network or violates the rules and
        protocols for communication across the network.
        
          Corresponding Source conveyed, and Installation Information provided,
        in accord with this section must be in a format that is publicly
        documented (and with an implementation available to the public in
        source code form), and must require no special password or key for
        unpacking, reading or copying.
        
          7. Additional Terms.
        
          "Additional permissions" are terms that supplement the terms of this
        License by making exceptions from one or more of its conditions.
        Additional permissions that are applicable to the entire Program shall
        be treated as though they were included in this License, to the extent
        that they are valid under applicable law.  If additional permissions
        apply only to part of the Program, that part may be used separately
        under those permissions, but the entire Program remains governed by
        this License without regard to the additional permissions.
        
          When you convey a copy of a covered work, you may at your option
        remove any additional permissions from that copy, or from any part of
        it.  (Additional permissions may be written to require their own
        removal in certain cases when you modify the work.)  You may place
        additional permissions on material, added by you to a covered work,
        for which you have or can give appropriate copyright permission.
        
          Notwithstanding any other provision of this License, for material you
        add to a covered work, you may (if authorized by the copyright holders of
        that material) supplement the terms of this License with terms:
        
            a) Disclaiming warranty or limiting liability differently from the
            terms of sections 15 and 16 of this License; or
        
            b) Requiring preservation of specified reasonable legal notices or
            author attributions in that material or in the Appropriate Legal
            Notices displayed by works containing it; or
        
            c) Prohibiting misrepresentation of the origin of that material, or
            requiring that modified versions of such material be marked in
            reasonable ways as different from the original version; or
        
            d) Limiting the use for publicity purposes of names of licensors or
            authors of the material; or
        
            e) Declining to grant rights under trademark law for use of some
            trade names, trademarks, or service marks; or
        
            f) Requiring indemnification of licensors and authors of that
            material by anyone who conveys the material (or modified versions of
            it) with contractual assumptions of liability to the recipient, for
            any liability that these contractual assumptions directly impose on
            those licensors and authors.
        
          All other non-permissive additional terms are considered "further
        restrictions" within the meaning of section 10.  If the Program as you
        received it, or any part of it, contains a notice stating that it is
        governed by this License along with a term that is a further
        restriction, you may remove that term.  If a license document contains
        a further restriction but permits relicensing or conveying under this
        License, you may add to a covered work material governed by the terms
        of that license document, provided that the further restriction does
        not survive such relicensing or conveying.
        
          If you add terms to a covered work in accord with this section, you
        must place, in the relevant source files, a statement of the
        additional terms that apply to those files, or a notice indicating
        where to find the applicable terms.
        
          Additional terms, permissive or non-permissive, may be stated in the
        form of a separately written license, or stated as exceptions;
        the above requirements apply either way.
        
          8. Termination.
        
          You may not propagate or modify a covered work except as expressly
        provided under this License.  Any attempt otherwise to propagate or
        modify it is void, and will automatically terminate your rights under
        this License (including any patent licenses granted under the third
        paragraph of section 11).
        
          However, if you cease all violation of this License, then your
        license from a particular copyright holder is reinstated (a)
        provisionally, unless and until the copyright holder explicitly and
        finally terminates your license, and (b) permanently, if the copyright
        holder fails to notify you of the violation by some reasonable means
        prior to 60 days after the cessation.
        
          Moreover, your license from a particular copyright holder is
        reinstated permanently if the copyright holder notifies you of the
        violation by some reasonable means, this is the first time you have
        received notice of violation of this License (for any work) from that
        copyright holder, and you cure the violation prior to 30 days after
        your receipt of the notice.
        
          Termination of your rights under this section does not terminate the
        licenses of parties who have received copies or rights from you under
        this License.  If your rights have been terminated and not permanently
        reinstated, you do not qualify to receive new licenses for the same
        material under section 10.
        
          9. Acceptance Not Required for Having Copies.
        
          You are not required to accept this License in order to receive or
        run a copy of the Program.  Ancillary propagation of a covered work
        occurring solely as a consequence of using peer-to-peer transmission
        to receive a copy likewise does not require acceptance.  However,
        nothing other than this License grants you permission to propagate or
        modify any covered work.  These actions infringe copyright if you do
        not accept this License.  Therefore, by modifying or propagating a
        covered work, you indicate your acceptance of this License to do so.
        
          10. Automatic Licensing of Downstream Recipients.
        
          Each time you convey a covered work, the recipient automatically
        receives a license from the original licensors, to run, modify and
        propagate that work, subject to this License.  You are not responsible
        for enforcing compliance by third parties with this License.
        
          An "entity transaction" is a transaction transferring control of an
        organization, or substantially all assets of one, or subdividing an
        organization, or merging organizations.  If propagation of a covered
        work results from an entity transaction, each party to that
        transaction who receives a copy of the work also receives whatever
        licenses to the work the party's predecessor in interest had or could
        give under the previous paragraph, plus a right to possession of the
        Corresponding Source of the work from the predecessor in interest, if
        the predecessor has it or can get it with reasonable efforts.
        
          You may not impose any further restrictions on the exercise of the
        rights granted or affirmed under this License.  For example, you may
        not impose a license fee, royalty, or other charge for exercise of
        rights granted under this License, and you may not initiate litigation
        (including a cross-claim or counterclaim in a lawsuit) alleging that
        any patent claim is infringed by making, using, selling, offering for
        sale, or importing the Program or any portion of it.
        
          11. Patents.
        
          A "contributor" is a copyright holder who authorizes use under this
        License of the Program or a work on which the Program is based.  The
        work thus licensed is called the contributor's "contributor version".
        
          A contributor's "essential patent claims" are all patent claims
        owned or controlled by the contributor, whether already acquired or
        hereafter acquired, that would be infringed by some manner, permitted
        by this License, of making, using, or selling its contributor version,
        but do not include claims that would be infringed only as a
        consequence of further modification of the contributor version.  For
        purposes of this definition, "control" includes the right to grant
        patent sublicenses in a manner consistent with the requirements of
        this License.
        
          Each contributor grants you a non-exclusive, worldwide, royalty-free
        patent license under the contributor's essential patent claims, to
        make, use, sell, offer for sale, import and otherwise run, modify and
        propagate the contents of its contributor version.
        
          In the following three paragraphs, a "patent license" is any express
        agreement or commitment, however denominated, not to enforce a patent
        (such as an express permission to practice a patent or covenant not to
        sue for patent infringement).  To "grant" such a patent license to a
        party means to make such an agreement or commitment not to enforce a
        patent against the party.
        
          If you convey a covered work, knowingly relying on a patent license,
        and the Corresponding Source of the work is not available for anyone
        to copy, free of charge and under the terms of this License, through a
        publicly available network server or other readily accessible means,
        then you must either (1) cause the Corresponding Source to be so
        available, or (2) arrange to deprive yourself of the benefit of the
        patent license for this particular work, or (3) arrange, in a manner
        consistent with the requirements of this License, to extend the patent
        license to downstream recipients.  "Knowingly relying" means you have
        actual knowledge that, but for the patent license, your conveying the
        covered work in a country, or your recipient's use of the covered work
        in a country, would infringe one or more identifiable patents in that
        country that you have reason to believe are valid.
        
          If, pursuant to or in connection with a single transaction or
        arrangement, you convey, or propagate by procuring conveyance of, a
        covered work, and grant a patent license to some of the parties
        receiving the covered work authorizing them to use, propagate, modify
        or convey a specific copy of the covered work, then the patent license
        you grant is automatically extended to all recipients of the covered
        work and works based on it.
        
          A patent license is "discriminatory" if it does not include within
        the scope of its coverage, prohibits the exercise of, or is
        conditioned on the non-exercise of one or more of the rights that are
        specifically granted under this License.  You may not convey a covered
        work if you are a party to an arrangement with a third party that is
        in the business of distributing software, under which you make payment
        to the third party based on the extent of your activity of conveying
        the work, and under which the third party grants, to any of the
        parties who would receive the covered work from you, a discriminatory
        patent license (a) in connection with copies of the covered work
        conveyed by you (or copies made from those copies), or (b) primarily
        for and in connection with specific products or compilations that
        contain the covered work, unless you entered into that arrangement,
        or that patent license was granted, prior to 28 March 2007.
        
          Nothing in this License shall be construed as excluding or limiting
        any implied license or other defenses to infringement that may
        otherwise be available to you under applicable patent law.
        
          12. No Surrender of Others' Freedom.
        
          If conditions are imposed on you (whether by court order, agreement or
        otherwise) that contradict the conditions of this License, they do not
        excuse you from the conditions of this License.  If you cannot convey a
        covered work so as to satisfy simultaneously your obligations under this
        License and any other pertinent obligations, then as a consequence you may
        not convey it at all.  For example, if you agree to terms that obligate you
        to collect a royalty for further conveying from those to whom you convey
        the Program, the only way you could satisfy both those terms and this
        License would be to refrain entirely from conveying the Program.
        
          13. Remote Network Interaction; Use with the GNU General Public License.
        
          Notwithstanding any other provision of this License, if you modify the
        Program, your modified version must prominently offer all users
        interacting with it remotely through a computer network (if your version
        supports such interaction) an opportunity to receive the Corresponding
        Source of your version by providing access to the Corresponding Source
        from a network server at no charge, through some standard or customary
        means of facilitating copying of software.  This Corresponding Source
        shall include the Corresponding Source for any work covered by version 3
        of the GNU General Public License that is incorporated pursuant to the
        following paragraph.
        
          Notwithstanding any other provision of this License, you have
        permission to link or combine any covered work with a work licensed
        under version 3 of the GNU General Public License into a single
        combined work, and to convey the resulting work.  The terms of this
        License will continue to apply to the part which is the covered work,
        but the work with which it is combined will remain governed by version
        3 of the GNU General Public License.
        
          14. Revised Versions of this License.
        
          The Free Software Foundation may publish revised and/or new versions of
        the GNU Affero General Public License from time to time.  Such new versions
        will be similar in spirit to the present version, but may differ in detail to
        address new problems or concerns.
        
          Each version is given a distinguishing version number.  If the
        Program specifies that a certain numbered version of the GNU Affero General
        Public License "or any later version" applies to it, you have the
        option of following the terms and conditions either of that numbered
        version or of any later version published by the Free Software
        Foundation.  If the Program does not specify a version number of the
        GNU Affero General Public License, you may choose any version ever published
        by the Free Software Foundation.
        
          If the Program specifies that a proxy can decide which future
        versions of the GNU Affero General Public License can be used, that proxy's
        public statement of acceptance of a version permanently authorizes you
        to choose that version for the Program.
        
          Later license versions may give you additional or different
        permissions.  However, no additional obligations are imposed on any
        author or copyright holder as a result of your choosing to follow a
        later version.
        
          15. Disclaimer of Warranty.
        
          THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
        APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
        HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
        OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
        THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
        PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
        IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
        ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
        
          16. Limitation of Liability.
        
          IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
        WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
        THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
        GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
        USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
        DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
        PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
        EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
        SUCH DAMAGES.
        
          17. Interpretation of Sections 15 and 16.
        
          If the disclaimer of warranty and limitation of liability provided
        above cannot be given local legal effect according to their terms,
        reviewing courts shall apply local law that most closely approximates
        an absolute waiver of all civil liability in connection with the
        Program, unless a warranty or assumption of liability accompanies a
        copy of the Program in return for a fee.
        
                             END OF TERMS AND CONDITIONS
        
                    How to Apply These Terms to Your New Programs
        
          If you develop a new program, and you want it to be of the greatest
        possible use to the public, the best way to achieve this is to make it
        free software which everyone can redistribute and change under these terms.
        
          To do so, attach the following notices to the program.  It is safest
        to attach them to the start of each source file to most effectively
        state the exclusion of warranty; and each file should have at least
        the "copyright" line and a pointer to where the full notice is found.
        
            <one line to give the program's name and a brief idea of what it does.>
            Copyright (C) <year>  <name of author>
        
            This program is free software: you can redistribute it and/or modify
            it under the terms of the GNU Affero General Public License as published by
            the Free Software Foundation, either version 3 of the License, or
            (at your option) any later version.
        
            This program is distributed in the hope that it will be useful,
            but WITHOUT ANY WARRANTY; without even the implied warranty of
            MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
            GNU Affero General Public License for more details.
        
            You should have received a copy of the GNU Affero General Public License
            along with this program.  If not, see <https://www.gnu.org/licenses/>.
        
        Also add information on how to contact you by electronic and paper mail.
        
          If your software can interact with users remotely through a computer
        network, you should also make sure that it provides a way for users to
        get its source.  For example, if your program is a web application, its
        interface could display a "Source" link that leads users to an archive
        of the code.  There are many ways you could offer source, and different
        solutions will be better for different programs; see section 13 for the
        specific requirements.
        
          You should also get your employer (if you work as a programmer) or school,
        if any, to sign a "copyright disclaimer" for the program, if necessary.
        For more information on this, and how to apply and follow the GNU AGPL, see
        <https://www.gnu.org/licenses/>.
        
Project-URL: Homepage, https://orbiteval.com
Classifier: Development Status :: 4 - Beta
Classifier: Intended Audience :: Science/Research
Classifier: Intended Audience :: Developers
Classifier: License :: OSI Approved :: GNU Affero General Public License v3
Classifier: Operating System :: OS Independent
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.9
Classifier: Programming Language :: Python :: 3.10
Classifier: Programming Language :: Python :: 3.11
Classifier: Programming Language :: Python :: 3.12
Classifier: Programming Language :: Python :: 3.13
Classifier: Topic :: Scientific/Engineering
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Classifier: Topic :: Software Development :: Quality Assurance
Requires-Python: >=3.9
Description-Content-Type: text/markdown
License-File: LICENSE
Dynamic: license-file
Dynamic: requires-python

# orbit-eval

**Measure your sigma once — ~8 replicate retrains — and the gate prices every
ship decision after that.** Robot-policy success rates move by whole
percentage points between *identical-data retrains*; a gate that ignores that
lottery false-shipped 16% of ORBIT's seed-replicate null pairs. orbit-eval is
the statistical decision layer that prices it: calibrate your cell's
`sigma_run` one time, feed it to the gate (`--sigma-run` or an atlas regime),
and every promote/hold verdict after that is priced against measured
retraining noise instead of eval luck.

Honest statistics for robot-policy evaluation. A thin, stdlib-only
pip-installable port of the validated statistics from the ORBIT research
program, shipping the program's **measured noise atlas** as data.

## The one distinction that matters

Every command in this tool keeps two different questions separated, because
they differ in price by more than an order of magnitude:

| | question | noise that governs it | price |
|---|---|---|---|
| **(a)** | *"Did my checkpoint / dataset change move the metric?"* — fixed checkpoints, fixed data, compared under **common random numbers** (CRN: same eval seed → same initial states) | `sigma_0`, the harness floor (PushT/DP 1.91 pp; SmolVLA-ft 20K **0.00** — bit-deterministic) | **~2 eval seeds** |
| **(b)** | *"Is selection method A better than method B?"* — every arm is an **independent retrain**, usually on an independently drawn episode set | `sigma_run` (0.78–11.43 pp measured across the 16 atlas cells; the low extremes are floor-dominated LOW-SR cells, not stability results) and `sigma_set` — **k-dependent**: 19.01 pp at k=22 single-task, falling to ~2.4–2.8 pp (pure) at k≥88 on a multi-task pool | **~59 set draws per arm** for a 10 pp effect at k=22 single-task; **~3–4** at k≥88 multi-task (k=176 re-measured: K4B band B3, pooled σ_run 3.36 df=6; prices unchanged) |

**CRN pairing cannot rescue (b).** Pairing cancels eval-draw noise only;
`sigma_run` and `sigma_set` live in training, not in the eval draw. Any tool
that quotes you a McNemar p-value as evidence about a *selection method* is
lying to you. This one refuses to.

## Install

```bash
pip install orbit-eval
```

or, from a source checkout:

```bash
pip install -e .
```

No dependencies (pure stdlib). Python ≥ 3.9.

## Quickstart (60 seconds)

Price a design before you run it — what can m retrains per arm actually
detect, and how many does your claimed effect need?

```console
$ orbit-eval power --regime smolvla-ft --effect 10
orbit-eval power — MDE = 2.80 * sigma * sqrt(2/m)  (alpha=.05 two-sided, power=.80)
regime: LIBERO / SmolVLA fine-tune k=22 (Tier 2)
  sigma_run = 3.31 pp  (90% chi-square CI [2.28, 6.34], df=6)
  sigma_0 (harness floor, CRN d=0) = 0.00 pp

arms: FIXED-SET — both arms retrain on the SAME fixed episode set,
so only sigma_run separates them (per-draw sigma = 3.31 pp)

  draws/arm    m=1    m=2    m=3    m=5    m=8    m=12   m=20   m=59
  MDE (pp)     13.1   9.3    7.6    5.9    4.6    3.8    2.9    1.7

  to detect a 10.0 pp effect: 2 draws per arm
  power at 2 draws/arm for this effect: 86%
[... provenance + honesty note trimmed ...]
```

Gate a candidate against an incumbent — two LeRobot `eval_info.json` files
(or training output dirs; sequential LIBERO eval declared so CRN pairing is
licensed):

```console
$ orbit-shipgate check incumbent_eval_info.json candidate_eval_info.json --sequential-eval
orbit-shipgate check  incumbent=incumbent_eval_info  candidate=candidate_eval_info
  engine 0.8.1  schema 2  gate-config sha256 62f4a20c1892

mode: CRN-PAIRED — one-sided exact McNemar + paired bootstrap
      (fixed-sample v0; anytime-valid sequential is v1)
  n episodes         500
  SR inc / cand      52.8 / 58.4 pp
  delta (cand-inc)   +5.6 pp
  95% CI             [+3.6, +7.6] pp
  discordant         inc-only b=0 / cand-only c=28 (rate 0.056)
  p_ship / p_worse   3.73e-09 / 1.0000  (one-sided, alpha 0.05)

VERDICT: SHIP -> exit 0
[... assumptions + caveats trimmed — the decision record carries them in full ...]
```

**Calibrate your own cell — this is the product loop.** The atlas prices only
the cells it measured; your stack is not one of them until you measure it.
Run ~8 replicate retrains of your training config (same data, same budget,
new training seeds), evaluate each on a fixed common-seed battery, and take
`sigma_run` from the pairwise gaps (`orbit_eval.stats.sigma_from_gaps`) —
a one-time cost that prices every ship decision after it. Feed it back as
`orbit-shipgate check --level retrain --sigma-run <pp>` (or a custom atlas regime
with an `sr_range`): the gate then prices the retraining lottery into every
verdict, and without it the gate still runs but records the no-prior
disclosure instead of inventing a number.

## Commands

```bash
orbit-eval audit <dir>            # scan eval outputs for silent defects
                                  #   (exit 0 clean / 1 flags or parse errors
                                  #    / 2 unusable path or no result files)
orbit-eval compare <runA> <runB>  # CRN-paired if possible, else unpaired
orbit-eval power --regime smolvla-ft --effect 10 [--arms method]
orbit-eval regress <base> <cand>  # CI gate: exit 0 ok / 1 regression
                                  #   / 2 invalid inputs / 3 underpowered
                                  #   (the design cannot detect the gate —
                                  #    a different remediation than 2)
orbit-eval selftest               # the tool verifies its own size/power/bias
orbit-eval release <log.csv>      # which skills did the new model break? (exit
                                  #   0 clean / 1 regressed / 3 cannot answer)
orbit-eval route build <cands>    # task-conditioned release over candidate
                                  #   checkpoints, with abstention (exit 0 /
                                  #   2 invalid inputs)
orbit-eval route plan --candidates 8 --tasks 10 --select-eps 50
```

`compare` and `regress` accept `--json` for machine-readable output (same
exit codes).

### `audit` — the defects that corrupted real waves

* **seed absent or seed == 1000** — LeRobot's *silent* CRN default: every
  default eval scores initial states 1000..1000+n−1. Great for pairing,
  poison for independence assumptions — either way you need to *know*.
* **truncated budgets** (`final_step != design_steps`) — a crashed run emits a
  complete-looking result with a plausible `sr`. Truncation is **directional**
  (undertrained scores LOW), not mean-zero; it corrupted 12/414 ORBIT runs and
  flipped the sign of one suite's headline result.
* **missing / tiny `n_eval`** — reported with the Wilson CI width it implies.
* **header/vector mismatch** — a file whose per-episode successes disagree
  with its own `overall` header (partial vector, or a header that contradicts
  the episodes) has the vector dropped at parse time and is flagged
  `SR_VECTOR_MISMATCH`: no paired statistic may run on episodes that do not
  correspond to the reported `n_eval`/SR.
* **replicate sets** — detected mechanically via
  `set_hash = sha1(sorted episode ids)[:16]`.

### `compare` — which regime is this number from?

If per-episode successes exist on both runs and the eval seeds match, you get
a **CRN-PAIRED** comparison: exact McNemar on the discordant pairs plus a
paired bootstrap CI on the SR difference. Otherwise you get an unpaired
two-proportion comparison with an explicit warning. Either way the output
states which of questions (a)/(b) the comparison can and cannot answer.

### `power` — both directions of the design rule

`MDE = 2.80 · sigma · sqrt(2/m)` (α=.05 two-sided, power .80), from the
Tier-0 power analysis. `--arms fixed-set` prices retrains on a fixed episode
set (`sigma_run`); `--arms method` prices independent set draws
(`sqrt(sigma_set² + sigma_run²)` → ~59 draws/arm at 10 pp in the k=22
single-task regime; the price is k-dependent and falls to ~3–4 draws/arm at
k≥88 on a multi-task pool) and prints the honesty note.

### `release` — which of your skills did the new model break?

**Start here.** One command, on an evaluation log you already have.

```bash
orbit-eval release evals.csv     # version,skill,episode,success (names auto-detected)
```

It compares the new model against the one it replaces, **per named skill**, and leads
with the answer. A skill is `REGRESSED` only when the whole 95 % interval sits at or
below −5 pp, so the verdict and the interval beside it can never disagree. A drop that
clears the floor but not the interval is `SUSPECT`. A skill with fewer than 10 episodes
is `UNDERPOWERED`: not judged, and explicitly **not** a pass.

Exit codes are meant for CI: `0` checked and clean, `1` a skill regressed or vanished
from the new model, `3` the file could not answer the question, `2` unusable input.

Why per skill and not the suite average: on a 50-task suite a best-validation release
damaged at least one task in 100 % of measured draws while the suite mean got three
times steadier (`research/cleanrel50/`). An average over many skills absorbs one
collapse, and it gets *steadier* as you add skills, not more sensitive.

With three or more versions it reads release history and answers the question a
deployer can act on: how many of your last N releases broke a skill, and how many of
those had a suite average that held or improved.

### `route` — the release step, measured

`cp checkpoint_final prod/` leaves success rate on the table in some cells and
not others. **How much is a measurement about your cell, never a rate to
expect.** Where it paid: eight identical-recipe retrains, same data, same eval —
ship one at random 62.3 %; ship the best-validation one 68.3 %; ship a
**task-conditioned** choice over the same eight 73.9 %, with damaged tasks
falling 1.82 → 0.22 (π₀.₅ / libero_object, held out on disjoint episodes;
`research/ROUTE1_RESULTS_2026-08-29.md`). The shape reproduced on Meta-World
MT10 (+6.9 pp median, 3 replicates) and MT50 (+10 pp).

Where it did not: on SmolVLA-ft / `libero_spatial` the same estimator reads
**+1.7 pp, 95 % CI [−2.1, +5.4]** at 100 selection episodes per task — zero
inside the interval — tracking σ_pertask (3.29 pp there against 6.69 on
`libero_object`, same wave; `experiments/influence/route1_spatial_scope_check.py`).
Routing pays in proportion to the per-task complementarity a cell actually has,
and `plan` exists to tell you which case you are in before you spend anything.

The claim that does travel is about the **tail**, not the mean. Shipping the
best-validation retrain reduced the probability of a damaged task far less than
per-task selection did at every pool size tested — and not at all on the 50-task
pools, where it stayed at 1.000 from J=1 through J=8, while one ten-task
replicate got *worse* with more candidates (0.786 → 0.929). Per-task selection
with abstention cut it 3–8× at J=3 in every ten-task cell, and the pool size
needed grew with the task count (`research/damage_jcurve/`, declaration
`fb1c760f`).

```bash
# candidates/<name>/**/eval_info.json (LeRobot), or a JSON matrix, or your own
# CSV export: candidate,task,episode,success[,block]
orbit-eval route build candidates/ --incumbent final --assume-crn \
    --policy-path final=/weights/final --policy-path s3=/weights/s3 \
    --out release.json --report REPORT.md
orbit-eval route build eval_export.csv --incumbent fleet_incumbents.csv   # per-SKU incumbents
orbit-eval route plan --candidates 4 --tasks 8 --select-eps 20              # before spending
```

What `build` decides and reports:

- **per task**, the candidate that ships: the per-task winner only when its
  advantage over that task's incumbent clears one-sided z ≥ 1.645 on the
  selection episodes, otherwise **ABSTAIN** (keep the incumbent) — routing must
  never itself be a lottery. Abstention is a deliberate trade and the report
  prints both sides: never abstaining is the better arm against the candidate
  *pool*, but it swaps far more of the fleet and makes a swap landing ≥ 10 pp
  below the model it replaced 14× to 48× more likely across the banked cells at
  J = 3. The default protects the incumbent; `--z-abstain 0` takes the other
  side knowingly. `--incumbent` is one candidate, a
  `task=candidate,...` list, or a JSON/CSV mapping (`*` = default), because a
  fleet already runs different models per SKU;
- the **held-out gain**: a stratified split-half over episodes (select on one
  half, score on the disjoint other half, 2000 draws) for RANDOM / INCUMBENT /
  ROUTE / ROUTE+ABSTAIN / ORACLE, with **priced damage counts** (a task is
  damaged when it sits ≥ 5 pp + 2·SE below the candidate-pool mean) and the
  worst task; the in-sample plug-in gain is printed too, labelled optimistic;
- **blocks**: give each episode a block label (robot, day, cell — a `block`
  column in the CSV or `"blocks"` in the JSON). Halves are stratified within
  block and every defection reports the advantage inside each block and whether
  all blocks agree; a defection that reverses sign on one robot is flagged as
  drift-sensitive. Candidates evaluated on different block mixes are refused;
- the **budget**: at your episodes per task, the per-task advantage the rule can
  defect on and the advantage it detects with 80 % power, and for 5/10/15/20 pp
  targets the episodes per task you would need and the total extra episodes
  that implies. At LeRobot's documented 10–20 episodes per task the threshold
  is 26–37 pp, so an INCUMBENT-STANDS verdict there means *unresolved*, and the
  report says so;
- the **regression bound**: for every defection the one-sided 95 % bound on
  "worse than the incumbent" is ≤ 0 by construction of the rule; abstained
  tasks keep the incumbent exactly; the per-task alpha is stated as per-task,
  with the expected false defections under a no-difference null;
- a signed record (`--out`: inputs sha256, decisions, held-out estimate, budget,
  timestamp, argv, and `--policy-path` locations for serving).

**Serving the release.** `orbit_eval.routed_policy.RoutedPolicy` turns the
record into one policy object: task in, the chosen candidate's action out, with
candidates loaded lazily and each carrying its own pre/post processing. For
LeRobot, `examples/eval_routed_lerobot.py` evaluates a release through
`lerobot-eval`'s own rollout loop unchanged (identity pipelines outside, each
candidate's pipelines inside), and checks the per-task outcomes against the
candidates' own evaluations.

```python
from orbit_eval.routed_policy import RoutedPolicy
rp = RoutedPolicy.from_release("release.json", loader=load_my_policy)  # loader(path) -> policy
rp.set_task("libero_object/3"); action = rp.select_action(observation)
```

`plan` costs nothing: given J candidates, T tasks and a selection budget it
prints the same budget lines plus the measured cells' J-curve / budget-curve
rows nearest your design, **stamped with their cell** — they are not a forecast
for yours (no transferable-sigma law). Below ~20 selection episodes per task
the advantage was unreliable in every measured cell.

Scope: task identity must be known at inference; J candidates are held and
served (a fine-tune that freezes its backbone shares it — the SmolVLA retrains
shared 378/500 tensors bitwise, so the routed release is one backbone plus
per-candidate expert modules).

### The measured atlas (`orbit_eval.atlas`)

All **16 measured regimes** ship as data — sigma_run spans **0.78–11.43 pp**
across cells (a ~15× range: pricing from the wrong cell is the exact failure
this tool refuses), and the extremes are LOW-SR floor-dominated cells, not
stability results.

**Suite-level `sigma_run` — DiffusionPolicy / SmolVLA / pi0.5:**

| regime | sigma_run (pp) | df | n_eval | provenance / stamps |
|---|---|---|---|---|
| `pusht-dp` k=103 | **2.09** | 7 | 500 | Phase 1c, 8 noise pairs, old gym_pusht stack |
| `pusht-dp-rebuilt` k=103 | **1.292** | 4 | 500 | **FLOOR-DOMINATED — measurement, not a pricing prior.** Resolved 2026-08-20 pooling all four post-rebuild replicate pairs; sigma sits BELOW its own 1.682 pp binomial floor (169% floor share, floor-subtracted 0.00) — no retrain variance is detectable above sampling. Do not price a gate from it and do not use it as a comparator (`compare_sigma_run` refuses it on floor share); for *pricing*, the corpus policy is max(atlas, EB-shrunk) = **1.80** — a policy choice, not a measurement. Absolute SR shifted −18.2 pp across the rebuild: never compare across it. |
| `libero-dp-k22` | **11.43** | 7 | 200 | Phase 2 |
| `libero-dp-k31` | **7.48** | 5 | 200 | Tier 1 |
| `libero-dp-k44` | **6.24** | 7 | 200 | Tier 1 |
| `smolvla-ft` k=22 | **3.31** | 6 | 200 | Tier 2 fine-tune; truncated-schedule stamp (20k steps / 30k decay) |
| `smolvla-ft-k88-multitask` | **3.35** (df=2, direction-only) / **2.11** re-measured | 2 / 7 | 200 / 2000 | kcurve replicate pairs; 2.11 re-measured at n=2000 on the Leg A′ fresh finals. ANNEAL-1 (2026-08-21, prereg-frozen): completing the LR schedule at this cell is **NULL** — the per-task lottery survives (LOTTERY-SURVIVES, CONCORDANT) |
| `smolvla-ft-pooled` | **2.82** | 8 | 200 | pooled across three suites (goal 1.08 / spatial 1.84 / long10 2.94); spread convention UNKNOWN and no single binomial floor exists for it — `compare_sigma_run` refuses the row; use the per-suite values |
| `pi05-ft` k=22 | **9.74** | 10 | 200 | PI05-PAIRS, 16 runs / 6 masks, within-mask ANOVA; χ² CI [6.81, 17.10]. **Overturns** the cell-1 df=2 reading of 2.02 (CI [1.05, 12.67] — the old point estimate was misleading, its interval was not). Excluding the collapsed mask null04: **5.23** (df=9, CI [3.60, 9.55]) |
| `pi05-ft-k22-multitask` | **6.02** | 7 | 200 | PI05_SCALE W2 |
| `pi05-ft-k88-multitask` | **4.05** | 7 | 200 | PI05_SCALE W1 (production k; 4.23 re-measured at n=1000, suite noise confirmed real) |

**ACT rows (PERTASK-3 legs B1/B2 + ACT-HORIZON)** — every row declares
`sr_range`, the SR band its 8 retrains actually landed in, and Ship-Gate
pre-check 11b refuses to price an eval outside that band:

| regime | sigma_run (pp) | df | SR range (mean) | stamp |
|---|---|---|---|---|
| `pusht-act` (defaults, h=100) | **0.784** | 7 | 0.2–3.0 (1.35) | **LOW-SR**: the binomial floor is ~43% of the variance (floor-subtracted ~0.59) — do NOT read as "ACT is low-noise" outside the near-floor regime |
| `pusht-act-h8` | **4.849** | 7 | 9.2–21.4 (16.35) | eval-horizon stamp: prices retrains *evaluated at* n_action_steps=8 (same checkpoints as `pusht-act`, re-evaluated) |
| `pusht-act-h16` | **4.839** | 7 | 14.6–30.0 (22.1) | selection disclosure: h=16 was the argmax of a one-checkpoint sweep, so its **SR level is optimistically biased** (the sigma is not — h=8, non-argmax, reads essentially the same) |
| `aloha-transfer-cube-act` | **6.515** | 7 | 69.6–88.4 (77.9) | the healthy-SR ACT cell (floor only 8% of variance, floor-subtracted 6.245); common-seed-list UNPAIRED row (G2 pairing unlicensed on gym-aloha) |
| `aloha-insertion-act` | **2.261** | 7 | 9.0–15.4 (12.0) | **LOW-SR**: floor 41% of variance (floor-subtracted 1.734). In *relative* terms the ordering vs transfer-cube reverses (18.9% of mean vs 8.4%) — the pp gap is SR regime, not stability |

**Per-task retrain noise (`sigma_pertask`) — always quoted with its suite.**
PERTASK-4 (T4-2 DIVERGENT, 2026-08-21): on libero_spatial the same
SmolVLA class/k measured 3.29 pp against libero_object's 6.69 pp —
non-overlapping CIs, so sigma_pertask is suite-dependent in magnitude *and*
shape and must never be quoted without naming its suite. Both readings are
always carried: the floor-subtracted binding value and the raw fixed-battery
value beside it.

| regime | suite | sigma_pertask (raw) pp | CI 95% | df |
|---|---|---|---|---|
| `pi05-ft-k88-multitask` | libero_object | **11.00** (11.82) | [6.86, 15.13] | 7 |
| `pi05-ft-k22-multitask` | libero_object | **12.01** (15.33) | [8.70, 15.31] | 7 |
| `smolvla-ft-k88-multitask` | libero_object | **6.69** (7.33) | [5.60, 7.77] | 7 |
| same class/k (PERTASK-4) | libero_spatial | **3.29** (4.52) | [1.18, 5.41] | 7 |

**The pretraining ladder is not monotone.** DP 11.43 → SmolVLA-ft 3.31 → **pi05-ft 9.74**.
The 4B fine-tune is *not* the quiet regime the two-pair reading suggested: PI05-PAIRS resolved
its pre-registered null04 rule on the **seed-attributable** branch — the same episode set scored
50.0 at seed 0 and 87.5 at seed 1, a **37.5 pp gap on byte-identical data**, the largest null gap
in the corpus. Any gate pricing a pi0.5 promotion against the old 2.02 prior was under-pricing
the retraining lottery by ~4.8×.

Plus `sigma_set` — 19.01 pp at k=22 single-task, with the measured k-curve
`SIGMA_SET_PURE_BY_K` (12.47 → 4.67 → 2.75 → 3.86 pp at k=22/44/88/176,
multi-task pool P=454; the k=176 entry uses the K4B-pooled σ_run — it was 2.43
under the original df=2 reading, the verdict flipped on re-measurement, and
both readings are kept) — and the
`sigma_0` harness floors (PushT/DP 1.91 pp; SmolVLA-ft 20K 0.00,
bit-deterministic). At small k, set-level draws are bimodal (a failure mode,
not a Gaussian spread); Gaussian MDEs overstate precision there.

#### Cross-cell comparisons: `compare_sigma_run` refuses before it quotes

A multiple like "ACT is N× noisier than DP" is only a fact about the
policies when both cells are on the same footing — which is why the atlas
never states one in prose. Two functions are the public API:

* `atlas.sigma_run_readings(regime)` — both readings of a row's sigma_run
  with its sampling floor separated out: raw, binomial floor at the row's own
  measured SR and `n_eval`, floor-subtracted, floor share of variance, and
  the row's spread convention.
* `atlas.compare_sigma_run(a, b)` — is a cross-cell multiple quotable, and if
  so what is it? It **refuses** (returns `quotable: False` with explicit
  reasons) rather than printing a number when:
  * either row records no `n_eval` — the raw sigma carries a battery-sampling
    term scaling as 1/√n, so 1.80-at-n=500 and 1.80-at-n=2000 are different
    quantities;
  * either row's `sigma_run_convention` is not a known `'eval-inclusive'`
    (an `'unknown'` convention, e.g. `smolvla-ft-pooled`, cannot be set
    beside a known one);
  * either row records no `df` — a ratio without both dfs has no uncertainty;
  * either row is ≥ 75% sampling floor by variance (its floor-subtracted
    reading is noise, not a measurement — e.g. `pusht-dp-rebuilt`);
  * the two rows' binomial floors differ by more than 2× (SR_SCOPE_FLOOR_RATIO)
    — the ratio would partly be a floor artifact;
  * the ratio's own 95% F confidence interval contains 1 — the cells are not
    distinguishable, and "A is 1.00× B" is not a finding.

  A quotable multiple therefore requires both rows to share `n_eval` and a
  known eval-inclusive convention, carry dfs, sit well above their floors on
  comparable floor scales, and have a ratio CI excluding 1. When any of that
  fails you get the reasons, not a number — quote those instead.

## Ship-Gate (`shipgate`)

A statistical promotion gate for robot policies, built on the same machinery.
`check` is **fixed-sample** — one look at a pre-sized eval; `seq` is the
**anytime-valid sequential** gate (v1) whose alpha survives peeking at every
observation. Input: an incumbent and a candidate eval (LeRobot
`eval_info.json` / `model_result.json`, the corpus per-episode JSONL, or a
LeRobot training **output dir** — the adapter locates
`eval_final/eval_info.json`, else the newest `eval*/eval_info.json`). The
gate verifies CRN comparability, runs the paired verdict engine, applies
winner's-curse handling, and rules with the exit code:

```bash
orbit-shipgate check INCUMBENT CANDIDATE [--ledger gate.jsonl] [--json]
    # exit 0 SHIP / 1 HOLD / 2 INVALID / 3 COLLECT-MORE / 4 UNRESOLVABLE
    # INCUMBENT/CANDIDATE may be LeRobot training output dirs
orbit-shipgate seq STREAM.jsonl --mode episodes|retrains [--tau 5.0] [--ledger g.jsonl]
    # anytime-valid sequential gate (v1): one state line per observation,
    # final decision record to the ledger; exit 0 stream consumed / 2 error
orbit-shipgate replay corpus.jsonl --out DIR    # historical replay -> REPLAY_REPORT.md
orbit-shipgate demo --corpus-dir DIR --out DIR  # compact DEMO.md + verifiable ledger
orbit-shipgate verify gate.jsonl                # recompute every record hash
orbit-shipgate capture OUT_TREE --ledger shadow.jsonl
    # harvest every checkpoint eval in a LeRobot training-output tree into a
    # sha256-CHAINED append-only ledger; idempotent (content-hash dedupe)
orbit-shipgate shadow --ledger shadow.jsonl --incumbent A --candidate B \
    [--decided promote|hold] [--regime R]
    # non-blocking: record what the shipped retrain-level gate WOULD have
    # said next to the team's real decision; always exit 0 unless inputs
    # are invalid
orbit-shipgate shadow --report --ledger shadow.jsonl
    # the divergence table — the kill-criterion instrument
```

* **Pre-checks** — budget gate (truncation is directional), per-episode
  successes present, equal `n_eval`, matching task sets, matching eval seeds
  (unrecorded / mixed-with-1000 seeds pair only under `--assume-crn`, with the
  assumption recorded). On LIBERO, pairing degrades to unpaired unless
  `--sequential-eval` declares a non-vectorized eval (issue #4152).
  **Regime-scope guards**: check 11 rules INVALID when a `--regime`'s
  measured cell verifiably contradicts the eval's own metadata (env, task
  count, trained-k); check **11b** (SR-scope, 2026-08-18) **refuses to price
  an eval whose SR sits outside the regime's declared `sr_range`** — sigma in
  pp carries a binomial component scaling as √(p(1−p)), so when the eval's
  binomial floor differs from every point of the band the regime was measured
  in by more than the 2.0× tolerance (`SR_SCOPE_FLOOR_RATIO`), the verdict is
  INVALID rather than a wrong-cell price (any LOW-SR row — `pusht-act`,
  `aloha-insertion-act` — will exercise this); check **11c** (2026-08-20)
  stamps an **ANTI-CONSERVATIVE PRIOR** warning on the record whenever the
  pricing regime's own sigma_run sits at/below its own binomial floor (e.g.
  `pusht-dp-rebuilt`) — pricing from such a cell *understates* the lottery,
  the one direction this gate exists to avoid. Bad run data yields an INVALID
  record, never a crash.
* **Primary test** — one-sided exact McNemar on the discordant pairs
  (candidate beats incumbent), `delta = candidate − incumbent` in pp with a
  paired bootstrap CI; degraded comparisons use the unpaired two-proportion
  test and Newcombe CI and say so.
* **Winner's curse** — if the candidate was picked as a max over J
  checkpoints, the measured +4.0 pp mean inflation (J=4, SmolVLA-ft,
  PI05_CELL1_RESULTS, measured prospectively) is subtracted from the
  effective delta; other J scale as `4.0·sqrt(ln J / ln 4)` pp, disclosed as
  an assumption in the record. With `--history` (the checkpoint evals) the
  bias is bootstrapped from the data instead. The correction gates SHIP only;
  it never converts a loss into a HOLD escape.
* **min effect 2.0 pp default** — the corpus-measured resolvability floor:
  delta < 2 pp resolves at 42–63% accuracy at every budget tested — a coin
  flip (ADAPTIVE_EVAL_PROGRAMME).
* **Per-task regression** — a task whose paired drop is ≥ 5 pp at one-sided
  p ≤ 0.05 HOLDs the release even when the overall delta is a win. Task
  blocks default to contiguous equal blocks (LeRobot orders eval episodes by
  task_id — recorded as an assumption); pass `--task-blocks` to override.
  **Retrain-priced since 2026-08-15** (HOLDINJ, prereg-frozen): at
  `comparison_level='retrain'` with a measured per-task σ (explicit
  `sigma_pertask_pp` or an atlas regime carrying `sigma_pertask` — measured:
  π₀.₅ / **libero_object** / k=88 at 11.00 pp, raw 11.82 beside it;
  sigma_pertask is suite-dependent — PERTASK-4 T4-2 DIVERGENT: libero_spatial
  measured 3.29 for the same class/k, so it must always name its suite), the
  governing per-task p prices the
  retraining lottery; the unpriced eval-exact rule HOLDs **96.4% of
  identical-data retrain pairs** at production k (HOLDRATE_DIAG_2026-08-15),
  the priced rule is calibrated (null flag rate 0.043 ≈ α; validated
  HOLDINJ_RESULTS). Verdicts are two-tier: `regression-beyond-lottery`
  (gate-actionable) vs `lottery-expected` (real difference a no-change
  retrain produces anyway); 5–20 pp regressions are below single-pair
  resolution (≈39 pp at σ=11) and route to COLLECT-MORE quoting the
  retrain-draw budget (≈15 retrains/side at 10 pp, ≈4 at 20 pp).
* **COLLECT-MORE sizing** — episodes needed for 80% power at the minimum
  effect, from the observed discordance rate (`var(delta_hat) ~ p_d/n`);
  UNRESOLVABLE when that exceeds `--max-budget`.
* **Sequential gating (`orbit-shipgate seq`, v1)** — e-process gating whose alpha
  is honest at *any* stopping time (Ville's inequality): `--mode episodes`
  consumes `{"inc":0/1,"cand":0/1}` CRN episode lines and runs a
  betting-mixture e-process over the discordant pairs; `--mode retrains`
  consumes `{"delta_pp":..,"se_eval_pp":..}` per-retrain-pair lines and runs
  a one-sided half-normal mixture e-process (prior scale `--tau`, default
  5.0 pp, disclosed) on `sigma_j² = 2·sigma_run² + se_eval_j²`, with the
  `sigma_run` prior resolved exactly as in `check` (explicit `--sigma-run` >
  `--regime` > none-with-disclosure). Evidence-to-ship when the running
  e-value reaches `1/alpha` and the corrected effect clears the minimum;
  mirror process for the worse direction; always-valid p-values and
  confidence sequences ride along in every emitted state line (`--every N`
  to thin). Deterministic — no RNG anywhere in the sequential path. The
  statistics live in `orbit_eval.gate.sequential` (Robbins normal-mixture
  martingales; Ramdas et al. safe anytime-valid inference).
* **LeRobot output dirs** — `check` positional args may be training output
  dirs (`eval_final/eval_info.json` preferred, else the newest
  `eval*/eval_info.json`, "newest" = the largest step embedded in the dir
  name — deterministic, no mtime); `--history` may be a run dir whose
  `checkpoints/<step>/eval_info.json` files form the full selection pool
  for the history-bootstrap curse measurement.
* **Decision records** — every verdict is a schema-versioned record carrying
  the full config, statistics, caveats, warnings, and assumption strings,
  hashed over canonical JSON (`record_sha256`, `gate_config_sha256`) and
  appended to an append-only JSONL ledger. `orbit-shipgate verify` recomputes every
  hash and trusts nothing on disk.
* **Replay** — `orbit-shipgate replay` runs corpus per-episode JSONL through the
  gate: null-pair mode (seed replicates trained on identical data — every
  SHIP is a false ship, reported with a Wilson CI against the configured
  alpha *plus* the disclosed dependence structure, a cluster bootstrap CI
  over replicate groups, the retraining noise the null gaps imply beside
  the atlas `sigma_run`, and a regenerated exact-null calibration control),
  curse mode (max-over-checkpoints verdicts with and without correction),
  and effect mode (different-training-set pairs: resolved vs COLLECT-MORE,
  at what episode budgets). Pair keys include the environment **stack
  epoch** (the measured `gym_pusht` rebuild), so cross-rebuild pairs are
  never formed — excluded pairings are counted in the report, and
  cross-wave pairs within an epoch carry a recorded assumption. Every
  number in the report is computed, none asserted.
* **Demo** — `orbit-shipgate demo --corpus-dir DIR --out DIR` replays a corpus
  directory and renders a compact `DEMO.md`: the headline eval-only vs
  retrain-aware false-ship table, three verbatim decision records, and the
  ledger-verify line — every number computed from the corpus at render
  time, over a fresh append-only ledger that `orbit-shipgate verify` re-checks.
* **Shadow-gating (`capture` / `shadow`)** — the deployment mode that costs
  a team nothing. `capture` walks a LeRobot training-output tree (the same
  `eval*/eval_info.json` discovery `check` uses) and appends one record per
  checkpoint eval: per-episode successes, eval seed **with provenance**
  (recorded in `eval_info.json`, the run dir's `model_result.json` sidecar,
  or null with LeRobot's silent default disclosed as
  `lerobot_default_1000_implicit` — never written down as if it had been
  recorded), steps, `set_hash`, and the G3 ordering metadata (task count,
  per-task episode counts, block order) that decides matched-config
  pairability. The ledger is **sha256-chained**: each record carries
  `prev_sha256` of the previous line (genesis 64 zeros), so a line cannot
  be altered, removed, or reordered without breaking the chain
  (`orbit_eval.gate.verify_chain` recomputes every link); capture is
  idempotent — re-capturing the same tree appends nothing. `shadow` then
  runs the shipped retrain-level gate on a captured pair **next to the
  team's declared decision** (`--decided promote|hold`) and records the
  divergence; it refuses via the verdict-INVALID path — recorded, still
  exit 0 — when the G3 metadata says the pair is not matched-config
  (different task sets, block order, or per-task counts: mismatches the
  engine's total-n check cannot see). Non-blocking by contract.
  `shadow --report` prints the divergence table (decisions, agreements,
  gate-said-HOLD-team-promoted = **reversed ship decisions**,
  gate-said-SHIP-team-held) — the literal instrument for the thesis kill
  criterion (*zero reversed ship decisions in 3 months of shadow-gating*).

## Runtime caveats baked into the output

Each of these cost the research program real GPU-days:

* LIBERO env `task_ids` ≠ dataset episode order — match the task **language
  string**, or you train one task and eval another (SR ≈ 0).
* `use_async_envs` must be **false** for reproducible seeding.
* LIBERO vectorized autoreset makes the init-state sequence outcome-dependent
  under early termination (LeRobot issue #4152) → CRN pairing on LIBERO
  requires **sequential** eval.
* Cross-stack comparisons are invalid: a `gym_pusht` rebuild shifted SR by
  −18.2 pp (measured, non-uniform).
* Episode-index → initial-state alignment was **validated 2026-08-02** (6/6
  comparisons, lerobot 0.6.0): reset-time state at episode index *i* is
  identical across fresh processes, sync/async vector envs, and batch sizes,
  for gym_pusht and LIBERO. Scope: the reset-time mapping — on LIBERO, full
  closed-loop evals with early termination still shift later init states
  (issue #4152; see the autoreset bullet above).

## Provenance

Ported from the ORBIT research repo: `analyze_paired.py` (budget gate,
bootstrap/randomisation machinery, the self-test pattern), `power_paired.py`
(generative model), `build_atlas.py` (set_hash, replicate detection),
`fit_forecaster.py` (stdlib stats helpers), `power_analysis_phase2.py`
(MDE design rules), `OPS_RUNBOOK.md` (the defects).

## Licensing

`orbit-eval` is released under the **GNU Affero General Public License, version 3
(AGPL-3.0-only)** — see [`LICENSE`](LICENSE).

In practice:

- **Running it on your own checkpoints, inside your own company, costs nothing
  and obliges you to nothing.** Internal use is not distribution. Point it at
  your candidates, get your σ and your release build, keep everything private.
- **If you modify it and offer it to others over a network** — an eval service, a
  hosted release-gate, a product with this engine inside it — the AGPL requires
  you to make your modified source available to those users.
- A **commercial license**, exempting you from the source-disclosure terms, is
  available from ORBIT Research.

The measurement corpus behind the shipped atlas — the replicate-retraining runs,
the pre-registration record, and the per-cell variance estimates — is a separate
work and is not covered by this license.
