<?xml version='1.0' encoding='UTF-8'?>
<rfc version="3" ipr="trust200902" category="info" submissionType="independent" docName="draft-watts-rec-evidence-bound-evaluation-00">
  <front>
    <title abbrev="REC-EP Evaluation Profile">An Evidence-Bound Protocol for Evaluating Unfamiliar Mathematical-Rule Communication</title>
    <author fullname="Deonté O'dell Watts" initials="D. O." surname="Watts">
      <organization>GoodShyt Group Inc.</organization>
      <address>
        <email>deonte@goodshyt.fun</email>
        <uri>https://orcid.org/0009-0005-8586-3650</uri>
      </address>
    </author>
    <date day="16" month="11" year="2026"/>
    <keyword>experimental protocol</keyword>
    <keyword>mathematical communication</keyword>
    <abstract>
      <t>This document specifies REC-EP, an experimental, transport-independent JSON interchange profile for exchanging masked mathematical-rule challenges, receiver predictions, evidence statuses, immutable artifact digests, and signed evaluator receipts. It separates invariant matches from proven equivalence, requires explicit evaluation custody boundaries, and supports reproducible comparisons of arithmetic-first, invariance-first, and hybrid instruction. REC-EP does not assert a novel mathematical theorem, a validated superiority effect, extraterrestrial intelligence, or IETF endorsement. This candidate is not submitted and proposes no IANA actions.</t>
    </abstract>
    <note title="Local Candidate - Release Gate" removeInRFC="true">
      <t>NOT SUBMITTED TO IETF; NOT AN RFC. Source ipr attribute indicates planned IETF boilerplate for a future authorized submission, not an assertion of current grant or publication. External rights review, current-date XML conversion and editor validation pending.</t>
    </note>
  </front>
  <middle>
    <section anchor="rec-ep-1">
      <name>Introduction and Applicability</name>
      <t>Communication about unknown mathematical rules is an evaluation problem as well as a mathematical one. Two systems can agree on observed invariants without agreeing on which state transitions are permitted. An evaluator who treats a matching signature as proof of a transformation makes an unsound inference. REC-EP defines exchangeable study, challenge, prediction, and evidence receipts so independent implementations can audit this distinction.</t>
      <t>Prior studies of emergent-language compositionality and generalization provide relevant motivation <xref target="Chaabouni2020"/> <xref target="MuGoodman2021"/>. They do not demonstrate the REC teaching-order hypothesis. The original Relational Echo Calculus (REC) artifacts began on 2026-10-21 and include local synthetic pilots and evaluator-readiness software. The initial REC work is exploratory and unreviewed; these materials are not evidence of extraterrestrial communication.</t>
      <t>This document is an Experimental/Informational individual Internet-Draft candidate intended for consideration in the Independent Submission Stream, subject to editor discretion. It does not define an Internet transport, does not update existing RFCs, and does not request IANA allocations. Implementations can exchange REC-EP messages over existing authenticated file channels or HTTP-based application protocols.</t>
      <t>This profile is distinct from the author's earlier proposals on evidence qualification and research provenance (draft-watts-agent-evidence-boundary, draft-watts-research-evidence-bundle, and draft-watts-research-intelligence-archive). REC-EP adds a bounded mathematical challenge/prediction exchange and evaluator-run scoring contract; it does not supersede or modify those proposals. Any overlapping evidence-receipt semantics require independent interoperability review before publication.</t><t>A conforming REC-EP evaluation profile supports interoperable artifacts; conformance is not a credential, independent replication, or a claim that invariance-first instruction improves receiver performance.</t>
    </section>
    <section anchor="rec-ep-2">
      <name>Conventions and Terminology</name>
      <t>The key words MUST, MUST NOT, REQUIRED, SHOULD, SHOULD NOT, MAY, and OPTIONAL in this document are interpreted as described in BCP 14 <xref target="RFC2119"/> <xref target="RFC8174"/> when, and only when, they appear in all capitals.</t>
      <ul>
        <li>
          <t>Motif: a sorted, duplicate-free finite collection of n-bit vertices.</t>
        </li>
        <li>
          <t>Echo: a specified allowable transformation of motifs, possibly from a finite evaluator-defined family.</t>
        </li>
        <li>
          <t>Resonance: an invariant or invariant-signature observation under a stated transformation family.</t>
        </li>
        <li>
          <t>Family: the fully specified transformation hypothesis space and its admissibility constraints.</t>
        </li>
        <li>
          <t>Witness: reproducible constructive evidence of a valid transformation.</t>
        </li>
        <li>
          <t>Separating proof: an admissible obstruction or complete search proving no permitted transformation exists.</t>
        </li>
        <li>
          <t>UNKNOWN: a distinct epistemic state; it is neither TRUE nor FALSE.</t>
        </li>
        <li>
          <t>Receiver: a model or program that receives a primer and public challenge and outputs a prediction.</t>
        </li>
        <li>
          <t>Independent evaluator: a separately accountable actor who holds the private generator, keys, test labels, and scoring controls.</t>
        </li>
      </ul>
      <t>Each proof claim MUST name the actual admissible family, dimension, and verification scope. A receipt is a record of evaluated operations, not a substitute for checking their scientific correctness.</t>
    </section>
    <section anchor="rec-ep-3">
      <name>Formal Reference Model</name>
      <t>For the initial signed-coordinate reference family, V_n = {0,1}^n and a motif is M subset V_n. An echo T_{pi,b} acts as P_pi(x) XOR b, with pi a coordinate permutation and b in V_n. The reference group G_n is (Z_2)^n semidirect S_n, with cardinality 2^n n!. Distance d_H(x,y) counts coordinates in which x and y differ.</t>
      <t>For T in G_n, d_H(Tx,Ty) = d_H(x,y). Consequently the pairwise-distance multiset R(M) and the induced distance-one vertex-degree multiset are preserved. Images of unions are unions of images. This is established finite symmetry mathematics, not a newly discovered algebra.</t>
      <t>For n=4, A={0,1,2,12} and B={0,1,6,7} both have pairwise Hamming-distance multiset {1,1,2,2,3,3}. However their distance-one degree multisets are {0,1,1,2} and {1,1,1,1}, respectively. They cannot be related by a G_4 echo. This explicit counterexample is a required negative test vector: equality of a coarse signature MUST NOT be promoted to EQUIVALENT.</t>
      <t>For evaluator-created rule families outside G_n, especially invertible binary-linear mixing and nonlinear Boolean maps, Hamming distance need not be invariant. An implementation MUST NOT reuse G_n-specific Hamming-distance arguments unless the independent family definition proves they are valid.</t>
      <t>The relation asserted in a challenge is defined relative to the evaluator-frozen admissible operation set and composition bound. If the evaluator cannot exhaust the relevant finite admissible set or supply a valid proof, the epistemic status MUST be UNKNOWN rather than DISTINCT.</t>
    </section>
    <section anchor="rec-ep-4">
      <name>Epistemic States and Non-Collapse Rule</name>
      <ul>
        <li>
          <t>EQUIVALENT: the evaluator has a valid witness and verified that the witness belongs to the frozen admissible family.</t>
        </li>
        <li>
          <t>DISTINCT: the evaluator has a separating invariant valid for this family or a complete finite search under the frozen rule and composition bounds.</t>
        </li>
        <li>
          <t>UNKNOWN: insufficient proof, ambiguous labels, incomplete search, or a generator not independently validated.</t>
        </li>
      </ul>
      <t>A match between resonance signatures is an observation and cannot, by itself, certify equivalence. Likewise, an absent witness is not evidence of non-equivalence. Transition from UNKNOWN to EQUIVALENT or DISTINCT MUST be accompanied by new admissible evidence; a downstream serializer, metric calculator, or receiver MUST NOT change a ground-truth status.</t>
      <t>Where label identifiability is not established within the actual hypothesis universe, the evaluator MUST mark the item UNKNOWN, record the ambiguity, and apply the exclusion or abstention policy fixed before confirmatory testing. A deliberately negative sample cannot be constructed merely by drawing a different-looking motif.</t>
    </section>
    <section anchor="rec-ep-5">
      <name>Exchange Objects and Versioning</name>
      <t>REC-EP 0.1 uses UTF-8 JSON <xref target="RFC8259"/>. Every public challenge MUST contain exactly the following fields: profile, study_id, run_id, trial_id, n, motif_a, motif_b. The profile value is "rec-ep/0.1". Motifs are sorted distinct integers between zero and 2^n-1, and in this version 1 &lt;= n &lt;= 12. A trial_id is an opaque, lowercase 32-hex-digit identifier independent of its label. Receivers MUST reject duplicate trial identifiers and unknown required-field changes; future versions MUST use a different profile value.</t>
      <figure>
        <sourcecode type="json">{
  "profile": "rec-ep/0.1",
  "study_id": "rec-synthetic-example",
  "run_id": "local-demo-only",
  "trial_id": "0123456789abcdef0123456789abcdef",
  "n": 4,
  "motif_a": [0, 1, 2, 12],
  "motif_b": [0, 1, 6, 7]
}</sourcecode>
      </figure>
      <t>The public receiver-view object MUST NOT contain answer keys, canonical rule mappings, hidden labels, private transformations, oracle handles, or indirect predictors of a target label. Implementations MUST audit additional fields before making a public fixture available. Private evaluator metadata MUST travel only through separately authorized channels.</t>
      <t>A prediction batch MUST contain exactly one prediction for every public trial_id. A batch MUST use a single study_id and a single run_id across all associated public challenges; an implementation MUST reject cross-run or cross-study mixtures. Each prediction has: trial_id, prediction (EQUIVALENT, DISTINCT, UNKNOWN), and p_equivalent_ppm, an integer from zero through one million. The ppm probability applies to the EQUIVALENT event. The prediction and probability fields MUST be interpreted consistently; any inconsistency policy MUST be fixed before evaluation. UNKNOWN is an abstention and requires explicit accounting.</t>
      <figure>
        <sourcecode type="json">[
  {"trial_id":"0123456789abcdef0123456789abcdef",
   "prediction":"DISTINCT", "p_equivalent_ppm":100000}
]</sourcecode>
      </figure>
      <t>The byte representation of every received file MUST be preserved exactly. A sha256 field binds to the unmodified exact file octets, prefixed with "sha256:" and followed by 64 lowercase hexadecimal characters. No application is permitted to silently reserialize the JSON and treat the resulting hash as the submitted file hash. Systems wanting semantic canonicalization MAY additionally use JCS <xref target="RFC8785"/>, but its digest MUST be separately named and generated with an actual JCS implementation.</t>
    </section>
    <section anchor="rec-ep-6">
      <name>Study Manifest and Pre-Registration</name>
      <t>A study manifest MUST specify study_id, protocol version, author and evaluator identifiers, registered question, primary estimands, arithmetic/invariance/hybrid conditions, unit of randomization, error rates, stopping rule, planned sample-size method, receiver families, data-access roles, rule-family holdout policies, software/version digests, and negative controls. Each preregistered exclusion rule and deviation policy MUST be recorded before unblinding.</t>
      <t>Pre-registration MUST be tied to an external immutable or independently timestamped record. A local SHA-256 hash or an evaluator-signed document alone demonstrates integrity or signer possession but does not prove the claimed creation time to an independent party. Actual registration URI and acceptance status are PENDING for REC.</t>
      <t>A proposed practically meaningful effect is delta=0.05 in balanced-accuracy units (five percentage points). This is an illustration, NOT an approved margin. Receiver-by-seed-by-rule-family hierarchical clustering must be represented in prospective power simulations. An older i.i.d.-paired calculation of 1,132 trials ignores clustering and MUST NOT be adopted as the confirmatory target.</t>
      <t>A concept-content by presentation-order factorial control SHOULD accompany the three teaching arms. Equal numbers of UTF-8 bytes or tokens do not guarantee equal semantic information, supervision, cognitive load, or target-label signal. An independent audit SHOULD document which aspects are matched and which cannot be matched.</t>
    </section>
    <section anchor="rec-ep-7">
      <name>Experimental Roles and State Machine</name>
      <ul>
        <li>
          <t>DRAFT: investigator proposes a hypothesis, materials, role allocation, and statistical plan. No confirmatory claim is allowed.</t>
        </li>
        <li>
          <t>ACCEPTED: independent evaluator supplies written qualifications, conflicts, data-rights, signed scope, costs, and independent infrastructure/key custody. Interest alone does not count.</t>
        </li>
        <li>
          <t>FROZEN: evaluator and authorized investigator lock the manifest and independently verifiable timestamp before generating or opening confirmatory test keys.</t>
        </li>
        <li>
          <t>COMMITTED: evaluator creates private hidden labels and publishes hashes over the exact immutable challenge and answer files without disclosing the answer bytes.</t>
        </li>
        <li>
          <t>RECEIVED: receiver submissions are sealed against the run_id, trial IDs, and challenge digest.</t>
        </li>
        <li>
          <t>SCORED: evaluator opens the protected answers only after receiving the locked predictions and produces a signed evidence-linked report.</t>
        </li>
        <li>
          <t>REVIEWED: separate independent replication and adversarial auditing may validate or dispute the scientific interpretation.</t>
        </li>
      </ul>
      <t>State transitions MUST NOT be inferred from an identifier alone. In particular, possession of a run_id, an evidence_id, or a signed receipt does not establish authorization, independence, correct labeling, or adequate statistical power. The verification policy MUST authenticate the signer through an external trust record and compare the receipt with the frozen run context.</t>
      <t>An evaluator MUST maintain confidentiality of private labels, generator code, signing keys, and hidden transformations. Receivers and their developers MUST NOT have read access to those materials before scoring. The same actor MAY have multiple roles only when the resulting conflicts are disclosed and the evaluation is explicitly described as non-independent.</t>
    </section>
    <section anchor="rec-ep-8">
      <name>Signing, Digests, and Replay Resistance</name>
      <t>The signed evaluator receipt uses JWS Compact Serialization <xref target="RFC7515"/> with an authenticated Ed25519 public key and algorithm EdDSA <xref target="RFC8037"/>. The protected header MUST carry alg=EdDSA, a key identifier bound by the relying party to an evaluator-owned key, and typ=REC-EP-RECEIPT+jws. Verifiers MUST reject alg=none, an unexpected algorithm, duplicate keys, unknown signing keys, and any invalid signature.</t>
      <t>The signed payload MUST contain exactly profile, study_id, run_id, evaluator_id and the exact-file digests named manifest_sha256, challenge_sha256, prediction_sha256, and result_sha256. The receiver or reviewer MUST compare these values to locally computed digests of the corresponding exact files and reject absent, changed, or extra trial IDs. The evaluator_id MUST be a nonempty text value, and a verifier MUST compare it to an expected evaluator identity bound out of band to the authenticated signing key. A valid signature using an otherwise trusted key MUST NOT permit a different evaluator identity to be asserted. Unspecified additional receipt fields MUST be rejected, including any purported authorization or publication approval. A receipt issued for run A MUST NOT authorize or validate results for run B.</t>
      <figure>
        <sourcecode type="json">{
  "profile":"rec-ep/0.1",
  "study_id":"rec-synthetic-example",
  "run_id":"local-demo-only",
  "evaluator_id":"fixture-only",
  "manifest_sha256":"sha256:&lt;64 lowercase hex&gt;",
  "challenge_sha256":"sha256:&lt;64 lowercase hex&gt;",
  "prediction_sha256":"sha256:&lt;64 lowercase hex&gt;",
  "result_sha256":"sha256:&lt;64 lowercase hex&gt;"
}</sourcecode>
      </figure>
      <t>The example uses placeholders and is NOT a valid signed production receipt. Real private signing keys MUST be created and retained in evaluator-controlled infrastructure. A locally generated demonstration key is untrusted for independent claims. Signatures authenticate bytes against a known public key; they do not establish that an evaluator is independent or that the scientific conclusion is valid.</t>
    </section>
    <section anchor="rec-ep-9">
      <name>Evaluation, Statistics, and Reporting</name>
      <t>The primary comparison is Delta21 = BA(H2) - BA(H1), where BA denotes class-balanced accuracy over a common prespecified task distribution. The secondary contrast is Delta3 = BA(H3) - max(BA(H1), BA(H2)). The same held-out target items and labels MUST be used across teaching conditions; differences in task type cannot be mistaken for differences in instruction.</t>
      <t>The evaluator MUST prespecify the distinction between an effect above zero, a practically meaningful advantage, and an inconclusive result. For example, with proposed delta=0.05, a confidence interval whose lower bound exceeds 0.05 supports the specified meaningful advantage; an upper bound below 0.05 excludes that magnitude under the model and design. A failure to reach significance alone does not demonstrate a zero effect. Multiplicity corrections and comparisons to the observed best comparator MUST be specified in the analysis plan.</t>
      <t>Every report MUST include per-architecture and per-dimension results, seed-blocked interval estimates, calibration where available, abstention/UNKNOWN coverage, false-equivalence promotion rate, hard-negative performance, null-scrambled message controls, input/compute budgets, leakage probes, preregistration deviations, and full negative findings. Runtime/energy costs SHOULD be measured if instrumentation is available.</t>
      <t>Rule-family extrapolation MUST be assessed separately from interpolation within unseen elements of a familiar family. Results from a receiver with hardcoded invariant features or a complete candidate operation table MUST be labeled as prior-assisted. Models that fail to outperform chance or null controls MUST NOT be described as having acquired the hidden rules.</t>
    </section>
    <section anchor="rec-ep-10">
      <name>Reference Conformance and Test Vectors</name>
      <t>A REC-EP implementation claiming local test-vector conformance MUST pass structural validation, duplicate-key rejection, no-key challenge checks, exact trial-id coverage, signed receipt verification, run-bound replay rejection, and the G_4 counterexample. These checks test implementation behavior and do not confer scientific validation.</t>
      <ul>
        <li>
          <t>TV-01: A={0,1,2,12}, B={0,1,6,7}; distance signatures match, induced-degree signatures differ, exact G_4 relation DISTINCT.</t>
        </li>
        <li>
          <t>TV-02: A={0,1,2,12}, C={1,2,3,15}; XOR with 3 is a valid witness, exact G_4 relation EQUIVALENT.</t>
        </li>
        <li>
          <t>TV-03: an unknown 7D operation set is not automatically DISTINCT; without admissible complete evidence the result remains UNKNOWN.</t>
        </li>
        <li>
          <t>TV-04: challenge containing answer_key is rejected.</t>
        </li>
        <li>
          <t>TV-05: a signed receipt for run A is rejected when verified against run B.</t>
        </li>
      </ul>
      <t>Two separately developed implementations exchanging the same messages and independently reproducing the scores would be an interoperability milestone. Such cross-implementation validation is NOT RUN in this revision. The bundled Python reference suite is a single local implementation and its demo signature uses ephemeral local keys only.</t>
    </section>
    <section anchor="rec-ep-11">
      <name>Security Considerations</name>
      <t>The primary threats are answer-key leakage, trial-identifier leakage, seed contamination, unauthorized status promotion, receipt replay, forged evaluator identities, compromised signing keys, tampered content, selective release, cross-split motif overlap, and false claims of statistical significance. Private data MAY contain sensitive research assets and MUST be protected by least-privilege access controls.</t>
      <t>A cryptographic hash does not conceal low-entropy secrets; commitments to small answer spaces are vulnerable to guessing. Evaluators SHOULD combine unpredictable secret salts with concealed data and keep such salts private until the reveal stage, while following a documented binding procedure. The commitment algorithm, byte encoding and release order MUST be frozen before any challenge is revealed.</t>
      <t>Implementations MUST NOT fetch remote resources identified in untrusted challenge strings without a separate authorization decision. Parsers MUST apply explicit input-size limits and reject malformed JSON, duplicate keys, and unsupported profile versions. The sample reference parser defaults to a one-megabyte JSON input budget; this implementation limit is not a universal protocol maximum. Public identifiers SHOULD avoid embedding people, institutions, or target labels. Systems SHOULD minimize retention of identifiable researcher data and document where public reproducibility conflicts with confidentiality.</t>
      <t>Production systems MUST establish evaluator trust anchors out of band and apply independent key rotation and revocation policies. Authentication verifies control of a key, not ownership of a research claim. Any signer who also controls the receiver or private generator MUST disclose the overlap and MUST NOT describe that operation as independent replication.</t>
    </section>
    <section anchor="rec-ep-12">
      <name>IANA Considerations</name>
      <t>This document has no IANA actions. No ports, media types, registries, or official compliance marks are requested. The string rec-ep/0.1 is an experimental in-document profile identifier, not an IANA-registered protocol identifier.</t>
    </section>
    <section anchor="rec-ep-13">
      <name>Limitations and Non-Claims</name>
      <t>This document is neither evidence of extraterrestrial intelligence nor an assertion that an alien civilization would discover the specified mathematics. It does not establish a globally novel theorem, a new physical dimension, universal mathematical communication, a validated pedagogical advantage, author-only invention, peer review, an issued patent, or a published RFC.</t>
      <t>The REC labels represent a research series; the informal first-articulation date is not a legal priority determination or proof of global novelty. External submissions, public repositories, patent disclosures, and new licensing grants require separate rights-holder approval. If this draft is submitted to the Datatracker, it becomes public and the applicable IETF Trust Legal Provisions and disclosure obligations must be evaluated first <xref target="RFC5378"/>.</t>
    </section>
    <section anchor="rec-ep-14">
      <name>Historical Evidence and Open Validation Gates</name>
      <t>REC v0.1: exact signed-hypercube algebra and the displayed G_4 coarse-signature collision were checked locally. A white-box decoder successfully predicted 30/30 constrained glyph examples, but it received the full candidate operation family. These observations demonstrate reference consistency only.</t>
      <t>REC v0.2: three curricula had equal synthetic byte budgets and the same labeled examples but mainly differed by order. A local 4D development pilot reported graph-receiver balanced accuracy H1 59.72%, H2 70.83%, H3 66.67%; a relational-feature receiver achieved 95.83% for all three, but relied on engineered features. These measurements are exploratory and do not isolate semantic teaching content.</t>
      <t>REC v0.3: an initial pilot had ambiguous demonstrations and was superseded. The corrected identifiable-rule single-seed 4D pilot reported GRU H1 55.47%, H2 54.69%, H3 54.69%, null 50.00%, and Transformer H1 50.00%, H2 46.88%, H3 46.88%, null 46.88%. No invariance-first advantage was demonstrated; near-chance and null-control results limit acquisition claims.</t>
      <t>REC v0.4: local evaluator-interface and mock scoring software checks were completed, with an illustrative i.i.d. paired trial calculation of 1,132 that is not cluster-powered and not accepted as a trial count. Independent keys, pre-registration, and real 5D/6D evaluator-controlled testing were not run.</t>
      <t>REC v0.4.1-v0.4.2: evaluator scoping and nonconfidential outreach were carried out. As of the last recorded November 8 mailbox review, five prospective contacts had no verified substantive acceptance. No reviewer appointment, confidential source transfer, independent study, final protocol freeze, or endorsement was established. These operational facts do not constitute peer review.</t>
      <t>The above pilot numbers are historical internal reports, not newly rerun in this protocol implementation. Their experimental denominators, logs and provenance remain in the separate controlled source archives. They are included to prevent selective reporting, not to substantiate superiority.</t>
    </section>
    <section anchor="rec-ep-15">
      <name>Publication and Interoperability Next Steps</name>
      <ul>
        <li>
          <t>Verify the draft against current xml2rfc and idnits tooling, including boilerplate, date and XML validity. This package has not undergone official Datatracker validation.</t>
        </li>
        <li>
          <t>Obtain independent technical review on cryptographic and JSON interchange interoperability, and a separate methodological review on information matching, power, and leakage.</t>
        </li>
        <li>
          <t>Agree the appropriate venue (an Internet-Draft is not automatically an RFC; Independent Submission requires ISE review, while research findings may fit a peer-reviewed venue).</t>
        </li>
        <li>
          <t>Select public licensing and IPR disclosures, reconcile ownership with any third-party rights, then authorize public submission only after removing confidential material.</t>
        </li>
        <li>
          <t>Freeze a legitimate preregistration and commission at least two independent implementations and a separate evaluator-created 5D/6D confirmatory trial.</t>
        </li>
      </ul>
      <t>Commercial implementations can offer hosted validation, cryptographic custody management, custom integrations, reporting, audit support, and training services. Such offerings MUST NOT imply that the IETF endorses them or that a passing software test proves new mathematics or extraterrestrial applicability.</t>
    </section>
  </middle>
  <back>
    <references>
      <name>Normative References</name>
      <reference anchor="RFC2119" target="https://www.rfc-editor.org/info/rfc2119">
        <front>
          <title>Key words for use in RFCs to Indicate Requirement Levels</title>
          <author fullname="S. Bradner" initials="S." surname="Bradner"/>
          <date year="1997"/>
        </front>
        <seriesInfo name="RFC" value="2119"/>
        <seriesInfo name="DOI" value="10.17487/RFC2119"/>
      </reference>
      <reference anchor="RFC8174" target="https://www.rfc-editor.org/info/rfc8174">
        <front>
          <title>Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words</title>
          <author fullname="B. Leiba" initials="B." surname="Leiba"/>
          <date year="2017"/>
        </front>
        <seriesInfo name="RFC" value="8174"/>
        <seriesInfo name="DOI" value="10.17487/RFC8174"/>
      </reference>
      <reference anchor="RFC8259" target="https://www.rfc-editor.org/info/rfc8259">
        <front>
          <title>The JavaScript Object Notation (JSON) Data Interchange Format</title>
          <author fullname="T. Bray" initials="T." surname="Bray"/>
          <date year="2017"/>
        </front>
        <seriesInfo name="RFC" value="8259"/>
        <seriesInfo name="DOI" value="10.17487/RFC8259"/>
      </reference>
      <reference anchor="RFC7515" target="https://www.rfc-editor.org/info/rfc7515">
        <front>
          <title>JSON Web Signature (JWS)</title>
          <author fullname="M. Jones" initials="M." surname="Jones"/>
          <date year="2015"/>
        </front>
        <seriesInfo name="RFC" value="7515"/>
        <seriesInfo name="DOI" value="10.17487/RFC7515"/>
        <annotation>Other authors: J. Bradley,  N. Sakimura</annotation>
      </reference>
      <reference anchor="RFC8037" target="https://www.rfc-editor.org/info/rfc8037">
        <front>
          <title>CFRG Elliptic Curve Diffie-Hellman (ECDH) and Signatures in JSON Object Signing and Encryption (JOSE)</title>
          <author fullname="I. Liusvaara" initials="I." surname="Liusvaara"/>
          <date year="2017"/>
        </front>
        <seriesInfo name="RFC" value="8037"/>
        <seriesInfo name="DOI" value="10.17487/RFC8037"/>
      </reference>
    </references>
    <references>
      <name>Informative References</name>
      <reference anchor="RFC8785" target="https://www.rfc-editor.org/info/rfc8785">
        <front>
          <title>JSON Canonicalization Scheme (JCS)</title>
          <author fullname="A. Rundgren" initials="A." surname="Rundgren"/>
          <date year="2020"/>
        </front>
        <seriesInfo name="RFC" value="8785"/>
        <seriesInfo name="DOI" value="10.17487/RFC8785"/>
        <annotation>Other authors: B. Jordan,  S. Erdtman</annotation>
      </reference>
      <reference anchor="RFC7991" target="https://www.rfc-editor.org/info/rfc7991">
        <front>
          <title>The xml2rfc Version 3 Vocabulary</title>
          <author fullname="P. Hoffman" initials="P." surname="Hoffman"/>
          <date year="2016"/>
        </front>
        <seriesInfo name="RFC" value="7991"/>
        <seriesInfo name="DOI" value="10.17487/RFC7991"/>
      </reference>
      <reference anchor="RFC5378" target="https://www.rfc-editor.org/info/rfc5378">
        <front>
          <title>Rights Contributors Provide to the IETF Trust</title>
          <author fullname="S. Bradner" initials="S." surname="Bradner"/>
          <date year="2008"/>
        </front>
        <seriesInfo name="RFC" value="5378"/>
        <seriesInfo name="DOI" value="10.17487/RFC5378"/>
        <annotation>Other authors: J. Contreras</annotation>
      </reference>
      <reference anchor="RFC9920" target="https://www.rfc-editor.org/info/rfc9920">
        <front>
          <title>RFC Editor Model (Version 3)</title>
          <author fullname="P. Hoffman" initials="P." surname="Hoffman"/>
          <date year="2026"/>
        </front>
        <seriesInfo name="RFC" value="9920"/>
        <seriesInfo name="DOI" value="10.17487/RFC9920"/>
        <annotation>Other authors: A. Rossi</annotation>
      </reference>
      <reference anchor="Chaabouni2020" target="https://aclanthology.org/2020.acl-main.407/">
        <front>
          <title>Compositionality and Generalization In Emergent Languages</title>
          <author fullname="R. Chaabouni" initials="R." surname="Chaabouni"/>
          <date year="2020"/>
        </front>
        <seriesInfo name="DOI" value="10.18653/v1/2020.acl-main.407"/>
        <annotation>Other authors: E. Kharitonov,  D. Bouchacourt,  E. Dupoux,  M. Baroni</annotation>
      </reference>
      <reference anchor="MuGoodman2021" target="https://arxiv.org/abs/2106.02668">
        <front>
          <title>Emergent Communication of Generalizations</title>
          <author fullname="J. Mu" initials="J." surname="Mu"/>
          <date year="2021"/>
        </front>
        <annotation>Other authors: N. Goodman</annotation>
      </reference>
    </references>
  </back>
</rfc>