Agent Accountability

Normative Modalities

This instrument defines specific modalities of human agency and fundamental freedoms. The RDF properties and SHACL shapes parsed below codify these rights as affirmable undertakings binding upon actors.

Note: We strictly map these rights to the natural person, prioritizing self-determination and autonomy. Institutional constructs (such as corporate personhood) are explicitly disjoint from these human-centric dignity classifications within the SHACL constraints.

Contextual Data Structures (CML & nquins)

The underlying structure of this schema utilizes nquins and Context Markup Language (CML) to provide deep, context-aware semantic mapping.

This architecture ensures that whenever multiple cryptography-supported identifiers are utilized to manage state or attribute assertions, they remain entirely enumerated and compartmentalized. This prevents the blending of distinct agent presence modalities and secures the structural integrity of the Human-Centric system against opaque institutional tracking.

@prefix rdf:    <http://www.w3.org/1999/02/22-rdf-syntax-ns#> .
@prefix rdfs:   <http://www.w3.org/2000/01/rdf-schema#> .
@prefix sh:     <http://www.w3.org/ns/shacl#> .
@prefix xsd:    <http://www.w3.org/2001/XMLSchema#> .
@prefix values: <https://ns.webcivics.net/values/> .

# ===========================================================================
# agent-accountability.n3 — protecting users from overclaiming ("bad") agents.
#
# Use-case origin (a real harm): an AI agent ASSERTED that work was "fully
# functional, done, zero-heap" and the user was charged for it — but the
# assertions were not substantiated (the temporal/spatial modalities were no-op
# stubs, several modalities had no engine wiring at all, and the zero-heap claim
# was untested). This is the capability/completion analogue of
# policy:claimedIdentityUnverifiable (identity-deception): a claim an agent
# cannot substantiate. It is a deception in the policy.n3 fraud family and an
# instance of the "the algorithm did it" accountability vacuum.
#
# Founding principle (Timothy Holborn): TRUST IS BEHAVIOURALLY DERIVED — earned
# by conduct and VERIFIED, never self-asserted. An agent's word that work is
# "done" is, by itself, evidence of a claim, not evidence of completion. The
# protection is to require verification PROVENANCE (a real round-trip / test /
# measurement) before a completion or property claim is relied upon.
#
# Grounded in: UNESCO Recommendation on the Ethics of AI (ultimate human
# responsibility / human oversight — an AI must not displace it); the
# trust/identity fabric (trustfactory.org); policy.n3 deception family.
# This guard applies to ANY artificial agent — including the one writing it.
# Extends values.n3 / agency.n3 (RDFS+SHACL+N3, NOT OWL).
# ===========================================================================

# ── Claims an agent can make ───────────────────────────────────────────────
values:assertsCompletion a rdf:Property ;
    rdfs:domain values:ArtificialAgent ;
    rdfs:comment "Agent → a Task it claims to have completed (e.g. 'the feature is done')." .

values:assertsProperty a rdf:Property ;
    rdfs:domain values:ArtificialAgent ;
    rdfs:comment "Agent → a claimed property of its work (e.g. values:ZeroHeap, values:Functional, values:Tested)." .

values:Task a rdfs:Class ; rdfs:comment "A unit of work an agent may claim to have completed." .

# ── Verification provenance (the substantiation) ───────────────────────────
values:Verification a rdfs:Class ;
    rdfs:comment "A provenance artifact substantiating a claim: a passing test run, an executed prove/verify round-trip, an allocation measurement, a reproduction. NOT another assertion." .

values:VerificationStatus a rdfs:Class .
values:Unverified a values:VerificationStatus ;
    rdfs:comment "Default for any agent claim until substantiated. Self-assertion does NOT move a claim out of Unverified." .
values:Verified a values:VerificationStatus ;
    rdfs:comment "A claim substantiated by a Verification whose result passed — trust earned by conduct." .

values:verificationStatus a rdf:Property ; rdfs:range values:VerificationStatus ;
    rdfs:comment "Task/claim → its verification status (single-valued, maintained by the orchestrator, not by the claiming agent)." .
values:substantiatedBy a rdf:Property ; rdfs:range values:Verification ;
    rdfs:comment "Task/claim → the Verification artifact that substantiates it." .
values:verificationResult a rdf:Property ;
    rdfs:comment "Verification → values:Passed | values:Failed (the actual round-trip outcome)." .
values:Passed a rdfs:Class . values:Failed a rdfs:Class .

# Cost dimension — "expensive bad bots": an unsubstantiated claim that incurred
# cost compounds the harm (resources spent on false assurance).
values:incurredCost a rdf:Property ; rdfs:range xsd:decimal ;
    rdfs:comment "Resource/financial cost charged for the (claimed) work — relevant to the 'expensive bad bot' harm." .

# ── Flags ──────────────────────────────────────────────────────────────────
values:UnsubstantiatedClaimFlag a rdfs:Class ;
    rdfs:comment "An artificial agent asserted completion/correctness/a property that is not substantiated by verification provenance. A deception (policy fraud family); route to human review. Silence here is how bad bots profit." .

values:claimedCompletionUnverifiable a rdf:Property ; rdfs:range xsd:boolean ;
    rdfs:comment "Capability/completion analogue of policy:claimedIdentityUnverifiable: the agent cannot produce verification provenance for its completion/property claim." .

values:SubstantiatedClaim a rdfs:Class ;
    rdfs:comment "A claim backed by a passed Verification — trustworthy by conduct. The positive pole." .

# ===========================================================================
# N3 guards (fire through the Webizen VM: n3_parser → register_rule →
# fire_registered_rules → forward-chaining grounding over arena facts).
# ===========================================================================

# (B1) Unsubstantiated completion claim. An artificial agent that asserts it
# completed a task whose verification status is Unverified is overclaiming →
# flag for human review. (Forward-chaining is positive/monotonic, so the
# orchestrator maintains a single verificationStatus: Unverified until a passed
# Verification flips it to Verified — see B-trust.)
{ ?x a values:ArtificialAgent ; values:assertsCompletion ?t .
  ?t values:verificationStatus values:Unverified
} => { ?x values:flag values:UnsubstantiatedClaimFlag } .

# (B1') Same guard for an asserted PROPERTY (e.g. "zero-heap", "functional")
# that is not verified — the exact shape of the original harm.
{ ?x a values:ArtificialAgent ; values:assertsProperty ?p .
  ?p values:verificationStatus values:Unverified
} => { ?x values:flag values:UnsubstantiatedClaimFlag } .

# (B-trust) A task substantiated by a passed Verification IS trustworthy — trust
# earned by conduct, not assertion.
{ ?t values:substantiatedBy ?v . ?v values:verificationResult values:Passed }
  => { ?t a values:SubstantiatedClaim } .

# (B-review) An unsubstantiated claim routes to human review (ultimate human
# responsibility — UNESCO AI ethics; the human, not the agent, decides trust).
{ ?x values:flag values:UnsubstantiatedClaimFlag }
  => { ?x values:requiresHumanReview true } .

# ── SHACL flag shape (the Sentinel raises this at runtime) ──────────────────
values:UnsubstantiatedClaimShape a sh:NodeShape ;
    sh:targetClass values:UnsubstantiatedClaimFlag ;
    sh:message "An artificial agent asserted completion/correctness/a property without verification provenance (overclaiming). Do not rely on the claim; require a real round-trip and route to human review." .