]> Continuity Laboratories

hello@continuitylaboratories.com

Security Individual Submission refusal transparency adversarial probes AI agents signed evidence offline verification Ed25519 A governance system for autonomous agents is defined as much by what it refuses as by what it permits, yet refusals are the one behavior vendors only ever assert. A Refusal Digest is a portable, signed JSON document recording one adversarial probe run against a live agent-governance system: for every attack attempted, the system's verbatim refusal ground, the event types the attack would have recorded had it succeeded, and the complete signed event ledger of the attempt — in which the refused transition is provably absent. Digests verify offline, by parties who do not operate the probed system, using only the issuer's public key. From version 0.2, every quantity that varies between runs derives deterministically from a per-run seed carried in the digest, so a relying party can recompute the derivations and a replayed or pre-recorded ledger cannot match a fresh digest. This document specifies the digest wire format, its canonicalization and signature scheme, the seeded-variation derivations, and the verification algorithm. Where succession receipts prove that an agent legitimately became the holder of an authority, refusal digests prove what the governing system declined to let happen.

Introduction Any vendor of agent-governance infrastructure claims that its guards hold: that the illegal transition is blocked, that the unauthorized actor is refused, that the forbidden state change cannot be recorded. The claim is load-bearing — buyers adopt such systems precisely for what they will not do — and it is almost never checkable. Audit logs show what happened; nothing shows, verifiably, what was prevented from happening. A Refusal Digest makes the claim checkable. It is a self-contained JSON document, produced by running a published suite of adversarial probes against the governing system, carrying for each probe: the attack attempted and the guard expected to refuse it, in plain language; the outcome — refused, or unexpected_pass when a guard did not fire (recorded and published, never hidden); the system's verbatim refusal ground — the actual error the guard returned, not a summary of it; the forbidden event types: the records the attack would have produced had it succeeded; and the complete signed event ledger of the attempt, embedded verbatim, in which every forbidden event type is provably absent. A relying party verifies a digest offline with only the issuer's Ed25519 public key: the proof over the whole document, every embedded event's hash and signature, the absence of every forbidden transition, and the summary counts. Refusal-before-recording is the probed system's discipline — an illegal transition is refused before anything is appended — so a successful refusal leaves exactly the shape the digest records: a ledger carrying the attempt's setup and no completion. One objection survives that verification: a dishonest operator could replay a single recorded probe run forever. Version 0.2 of the format answers it with seeded variation (): every quantity that varies between runs — the staged identifiers of each probe's setup, the ordering of causally independent steps — derives deterministically from a fresh per-run seed carried in the digest. A verifier recomputes the derivations from the seed; a ledger recorded under any other seed cannot match. A sequence of runs MAY additionally chain seeds, making the sequence itself tamper-evident. This document is a companion to Succession Receipts , with which it shares its canonicalization and signature discipline () and its conformance-corpus discipline: the format is published with golden vectors and tamper cases that MUST fail at named checks , against which independent verifier implementations validate. Succession receipts attest that an agent legitimately became the holder of an authority; refusal digests attest the transitions the same class of system refused to perform. Pre-execution authorization of individual material actions — a named human approving the exact action before it runs, as in — is a third, complementary evidence class: what was authorized, what transferred, and what was refused compose without overlap. In the transparency-log sense, a digest is a signed statement suitable for registration with transparency services (in the style of and SCITT ); anchoring is discussed in .

Conventions and Definitions The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 when, and only when, they appear in all capitals, as shown here.

Probe:

One adversarial attack attempted against a freshly constructed, isolated instance of the governed system, together with the staged setup it requires.

Refusal ground:

The verbatim error with which the governed system refused the probe's attack — the same string an ordinary caller would receive.

Forbidden event types:

The event types the probe's attack would have appended to the ledger had it succeeded. Their absence from the probe's embedded ledger is what verification checks.

Run seed:

A 32-byte value, carried as 64 lowercase hexadecimal characters in a version 0.2 digest, from which every run-varying quantity derives ().

Issuer:

The operator that ran the probe suite and signs the digest. At the versions specified here the issuer attests its own run (run.attestation is "self"); see .

Relying party:

Any holder of the digest verifying it against the issuer's public keys.

The Digest Document A Refusal Digest is a UTF-8 JSON object. The complete normative member catalog, with types and constraints, is the published digest specification , which publishes version 0.1 (frozen) and version 0.2 (current); this section summarizes the structure a verifier depends on and shows a version 0.2 digest.

", "completed_at": "", "attestation": "self", "seed": "<64 lowercase hex characters>" }, "probes": [ { "id": "illegal-amendment", "name": "Illegal Amendment Attempt", "attack": "", "guard": "", "outcome": "refused" | "unexpected_pass", "refusal_ground": "", "engine_summary": "", "forbidden_event_types": [ "AmendmentRatified", ... ], "events": [ , ... ] }, ... ], "summary": { "attempted": , "refused": }, "proof": { "type": "CSSEd25519Signature", "created": "", "verification_method": "", "digest_hash": "", "signature": "ed25519::" } } ]]>

run.seed is REQUIRED at version 0.2 and absent at version 0.1. A version 0.1 digest sets spec_version to 0.1 and carries no seed; a conforming verifier accepts either version, each under its own rules (). Timestamps throughout are UTC instants. Each probe runs on an isolated, freshly constructed instance of the governed system, so its events member is the complete ledger of that attempt and probes cannot contaminate one another. Event envelopes carry the same members, integrity rules, and optional signatures as succession receipt evidence (; the event-hash rule is identical) and are embedded exactly as recorded. The probe suite is expected to grow. Adding a probe is additive within a spec version: verification is per-probe, and the summary is recounted, never trusted.

Seeded Variation Version 0.2 exists to defeat the replayed-ledger objection. Its rule: every quantity that varies between runs derives deterministically from run.seed, so any relying party can recompute the variation, and no pre-recorded ledger can satisfy a digest carrying a different seed. The standing derivations, over the seed's lowercase-hex string form and its 32 raw bytes: Run namespace. NS = UUIDv5(NameSpaceURL, "css-refusal-digest/v0.2|" + seed), where UUIDv5 and the URL namespace are as defined in . Derived identifiers. Each staged identifier of a probe's setup is UUIDv5(NS, "<probe-id>|<quantity-label>"). The identifier labels each probe MUST derive — and the ledger positions where they MUST appear — are pinned per probe by the published specification and its conformance corpus . Derived orderings. Where a probe contains causally independent steps (or attack routes) whose order is real run-to-run variation, the candidates are sorted by the lowercase-hex SHA-256 digest of the concatenation of the seed's raw bytes and "<probe-id>|<candidate-label>". What never varies is the line the transparency claim lives on: the probe set, each probe's id, the guard under test, the attack's semantics, and the forbidden_event_types. If those varied, a run would stop being independently recomputable — a reader could no longer say what was attacked, only that something was — and the claim would collapse to trust in the operator. Variation exists to make replay detectable, never to make the attack a moving target. Seed choice is deliberately unconstrained: because nothing seed-derived affects what is attacked or what must be refused, grinding seeds buys an issuer nothing — any seed's run must refuse. For a sequence of runs it is RECOMMENDED to chain seeds as the lowercase-hex SHA-256 of the ASCII bytes of the previous digest's digest_hash (first run: issuer's choice). A chained sequence is self-linking: the claim "N consecutive runs, all refused" then inherits the same tamper evidence as the digests themselves.

Canonicalization and Signatures The digest reuses the Succession Receipts discipline () unchanged: Canonical bytes: the JSON serialization with the proof member absent, member names sorted lexicographically, no insignificant whitespace, no HTML escaping — coinciding with the JSON Canonicalization Scheme for the format's value domain (no floating-point numbers). Hash-then-sign: the proof signs the lowercase hexadecimal SHA-256 digest of the canonical bytes (the digest string is the signed message); digest_hash records it; signature is the canonical <alg>:<key_id>:base64url(signature-bytes) string, with ed25519 () the sole algorithm registered at versions 0.1 and 0.2. Event integrity: each embedded event's event_hash recomputes under the same event-hash rule as receipt evidence, and a present event signature verifies over it. proof.created is outside the signed content: re-signing identical content yields the same digest_hash.

Verification A verifier is given the digest document and the issuer's public keys, pinned out of band. Verification MUST perform, in order: Structure. spec and spec_version match; run.attestation is "self"; at least one probe is present; every probe carries an id, a non-empty forbidden_event_types, and an outcome of refused or unexpected_pass; a refused probe carries a non-empty refusal_ground. At version 0.2, run.seed is present and is exactly 64 lowercase hexadecimal characters; at version 0.1 it is absent. Unknown versions MUST be rejected. Proof. Recompute the canonical hash (). It MUST equal proof.digest_hash, and proof.signature MUST verify against it under the key named by its key_id. A missing proof, an unknown key_id, a hash mismatch, or a failed signature check is fatal. Probe evidence. For every event of every probe: the event hash MUST recompute to its stored value, and a present event signature MUST verify (absent signatures are reported, not fatal; a present-but-invalid signature is fatal). Refused-transition absence. For every probe with outcome refused: no embedded event's type appears in that probe's forbidden_event_types. A digest claiming a refusal while carrying the completed transition is invalid. Summary integrity. summary.attempted MUST equal the probe count and summary.refused the count of refused outcomes. Seed derivation (version 0.2). For every probe whose derivation table the verifier implements (the standing suite, pinned by the published corpus), recompute the derived identifiers and orderings of from run.seed and require the probe's ledger to have used them. For probes it does not know, a verifier MUST still enforce steps 1–5 in full. "Valid" and "all refused" are separate statements. A digest recording an unexpected_pass can verify — it is an honest record of a failed run, published rather than hidden — and a relying party whose standing question is "were all attacks refused?" MUST additionally check that summary.refused equals summary.attempted. Reference tooling distinguishes the two outcomes deliberately. A conforming verifier MUST accept every golden vector and MUST reject every tamper case of the published conformance corpus (trees refusal-v0.1 and refusal-v0.2) at the named check .

Operating Model Cadence is the operator's dial, not the format's: a digest attests one probe run, and nothing in it changes with how often runs happen. An issuer publishing on a cadence SHOULD publish the cadence alongside the digests; a digest says nothing about attacks it does not contain, and nothing about intervals it does not cover. A guard that does not fire is a published fact, never a silent gap: the unexpected_pass outcome is recorded in the digest, the digest still signs and still verifies, and the operational response happens outside the format. Continuous operation naturally combines with seed chaining (): a long-running issuer emits a self-linking sequence in which each run's seed commits to the previous run's content.

Versioning and Stability spec_version identifies the wire format. Published versions are never mutated: format changes only ever add a new version with new golden vectors, and verifiers SHOULD continue to verify every published version. Version 0.1 (frozen) carries no seed and verifies under steps 1–5 forever; version 0.2 (current) adds run.seed and the derivation checks additively .

Security Considerations Self-attestation is the stated trust model, not a hidden one. A digest proves its content is intact, its embedded ledgers are genuine output of the probed system's signing keys, and the refused transitions are absent. It does not prove who operated the probes: at the versions specified here the issuer attests its own run. Probes run by an independent operator and countersigned — the proof becoming a set — are the format's recorded evolution and change who signs, not the format. Relying parties weighing operator honesty get exactly the leverage the seed provides (): whatever else a dishonest operator can do, it cannot serve yesterday's ledger under today's seed. A digest is a positive record with stated scope. It attests the probes it contains, on the cadence its issuer publishes — not the absence of other vulnerabilities, and not behavior between runs. Consumers SHOULD treat the probe suite's composition (pinned, public, versioned) as part of what they are trusting. Binding to the artifact under test is a recorded evolution. A digest names what was attacked and what was refused; the versions specified here do not bind which build of the governed system was probed, so the claim that the probes ran against the deployed production artifact rests on the operator, exactly as the self-attestation model states. The recorded strengthening path is a probe-suite digest and a deployed-artifact digest or build attestation carried in the run metadata, with the independent challenger or witness of the countersignature evolution above checking them. Seed grinding is neutralized by construction. Nothing seed-derived affects what is attacked or what must be refused, so an issuer gains nothing by choosing seeds adversarially; the seed exists to bind the ledger bytes to the run, not to randomize the test. Key pinning is the trust root, exactly as for succession receipts: relying parties MUST obtain issuer keys through a channel they trust. The digest additionally embeds the probed system's event signatures, so a relying party that pins the system's event-signing keys separately from the digest-signing key narrows the forgery surface further. Existence time composes with transparency infrastructure. A digest's digest_hash is a natural registrant for transparency and timestamping services — a SCITT transparency service or comparable append-only witness — giving each published digest an existence proof independent of its issuer. Issuers operating on a cadence SHOULD anchor digests so that backdating a "clean" digest after an incident is detectable. Deterministic serialization is load-bearing. Implementations MUST reproduce the canonical form and the event payloads' document-order serialization exactly; the conformance corpus exists to make divergence detectable, and its version 0.2 trees include re-signed derivation forgeries that only the seed checks refuse.

IANA Considerations This document has no IANA actions. The signature-algorithm registry is internal to the format's spec-version ladder (), shared with Succession Receipts .

In many standards track documents several words are used to signify the requirements in the specification. These words are often capitalized. This document defines these words as they should be interpreted in IETF documents. This document specifies an Internet Best Current Practices for the Internet Community, and requests discussion and suggestions for improvements. RFC 2119 specifies common key words that may be used in protocol specifications. This document aims to reduce the ambiguity by clarifying that only UPPERCASE usage of the key words have the defined special meanings. This document describes elliptic curve signature scheme Edwards-curve Digital Signature Algorithm (EdDSA). The algorithm is instantiated with recommended parameters for the edwards25519 and edwards448 curves. An example implementation and test vectors are provided. JavaScript Object Notation (JSON) is a lightweight, text-based, language-independent data interchange format. It was derived from the ECMAScript Programming Language Standard. JSON defines a small set of formatting rules for the portable representation of structured data. This document removes inconsistencies with other specifications of JSON, repairs specification errors, and offers experience-based interoperability guidance. Cryptographic operations like hashing and signing need the data to be expressed in an invariant format so that the operations are reliably repeatable. One way to address this is to create a canonical representation of the data. Canonicalization also permits data to be exchanged in its original form on the "wire" while cryptographic operations performed on the canonicalized counterpart of the data in the producer and consumer endpoints generate consistent results. This document describes the JSON Canonicalization Scheme (JCS). This specification defines how to create a canonical representation of JSON data by building on the strict serialization methods for JSON primitives defined by ECMAScript, constraining JSON data to the Internet JSON (I-JSON) subset, and by using deterministic property sorting. This document defines a date and time format for use in Internet protocols that is a profile of the ISO 8601 standard for representation of dates and times using the Gregorian calendar. This specification defines UUIDs (Universally Unique IDentifiers) -- also known as GUIDs (Globally Unique IDentifiers) -- and a Uniform Resource Name namespace for UUIDs. A UUID is 128 bits long and is intended to guarantee uniqueness across space and time. UUIDs were originally used in the Apollo Network Computing System (NCS), later in the Open Software Foundation's (OSF's) Distributed Computing Environment (DCE), and then in Microsoft Windows platforms. This specification is derived from the OSF DCE specification with the kind permission of the OSF (now known as "The Open Group"). Information from earlier versions of the OSF DCE specification have been incorporated into this document. This document obsoletes RFC 4122. This document describes version 2.0 of the Certificate Transparency (CT) protocol for publicly logging the existence of Transport Layer Security (TLS) server certificates as they are issued or observed, in a manner that allows anyone to audit certification authority (CA) activity and notice the issuance of suspect certificates as well as to audit the certificate logs themselves. The intent is that eventually clients would refuse to honor certificates that do not appear in a log, effectively forcing CAs to add all issued certificates to the logs. This document obsoletes RFC 6962. It also specifies a new TLS extension that is used to send various CT log artifacts. Logs are network services that implement the protocol operations for submissions and queries that are defined in this document. Traceability in supply chains is a growing security concern. While Verifiable Data Structures (VDSs) have addressed specific issues, such as equivocation over digital certificates, they lack a universal architecture for all supply chains. This document defines such an architecture for single-issuer signed statement transparency. It ensures extensibility and interoperability between different transparency services as well as compliance with various auditing procedures and regulatory requirements. EMILIA Protocol, Inc. This document defines the EMILIA Protocol (EP) authorization receipt, an evidence artifact binding an enrolled approver key to one canonical action before execution. An approver-held key signs an Authorization Context containing the action hash, policy reference, nonce, audience, and validity window. A Trust Receipt carries the signed contexts, terminal consumption record, and Merkle inclusion material so a relying party can verify the recorded event offline under independently selected log, directory, policy, and approver trust inputs. The receipt establishes only the guarantees of the selected verification profile. The mapping from an enrolled approver identifier to a natural person is asserted by the directory authority. Offline verification does not establish current revocation status, global non-replay, comprehension, legality, safety, or execution. Replay prevention requires an online atomic consumption store at the executor. The state-machine invariants are machine-checked under the assumptions stated in this document.

Acknowledgments The format was extracted from a production authority-succession system of record, where the probe suite runs on a standing cadence against the system's own constitutional guards; the conformance corpus packages the resulting attack classes — including fully re-signed derivation forgeries — as executable verification cases. Iman Schrock provided detailed external review of the -00 document family; this revision incorporates it.

Change Log -01: Added the pre-execution authorization composition note and the reference; added the artifact-binding security consideration (probe-suite digest and deployed-artifact attestation as the recorded strengthening path). No wire-format change.