Expand description
Declared reliance & judgement provenance as a DAG — the evidentiary substrate that keeps a natural person’s authorship and responsibility intact while making their toolchain auditable.
(Timothy, 2026-07-03; see docs/plans/adr-authority-attestation-guardianship-model.md §8/§9/§10.)
When a natural agent forms a judgement for another person, they must declare which other agents
informed it. A Reliance records who/what contributed and how; because a consulted agent carries
their own JudgementProvenance (nested via Reliance::sub_provenance), informed_by is not a
flat list — it is a directed acyclic graph of contributions.
Load-bearing distinctions:
- Selfhood ≠ personhood, authorship stays with the natural person. A
AgentType::SoftwareAgentis tooling & provenance — it never bears liability, but its use MUST be disclosed and scoped. The natural personresponsible_agentbears authorship + responsibility. - Dual-timed veracity. Each contribution carries its veracity as reasonably assessed at the time
(
Reliance::veracity_at_time, against theJudgementProvenance::epistemic_horizon) and as later determined by forensic review (Reliance::veracity_determined). The gap between the two is exactly where the malice / negligence / honest-error distinction lives (§10). - Epistemic horizon is a content-addressed reference (Merkle root / checkpoint hash) to the information-state available to the agent at decision time — the hindsight-resistance mechanism.
- Standing declaration. A
RelianceDeclarationlets an agent declare, up front, which tools they MAY use, so a principal can consent to the toolchain before relying on the agent. An undeclared software agent discovered later is a serious integrity breach — seehas_undeclared_ai. - Disclosure is two independent axes (subject vs other agents) over a proof modality; the process subject’s right-to-know is the first-class default (§9).
Structs§
- Agent
Ref - A reference to a contributing agent.
capacityis a role/credential term id (e.g. a professional registration URI);versionpins a software agent or instrument to a specific version/configuration so result variation is attributable (§10). - Disclosure
Policy - Per-node disclosure policy over two independent axes (§9): what the process subject (the natural person the decision is about) may see, versus what other agents / institutions may see. The subject’s right-to-know is the first-class default; others get proportionate disclosure.
- Judgement
Provenance - The provenance of a single judgement: the natural (or other) agent responsible for it, the ordered
set of contributions that informed it, an optional epistemic horizon, and free-form procedure notes.
Because each
Reliancemay carry a nestedsub_provenance, this type is recursively a DAG node. - Reliance
- One contribution to a judgement — an edge in the provenance DAG. Records the contributing agent, the
nature of reliance (
"diagnostic-support","drafting","data-source","consult", …), whether the agent/tool was used within its validated/insured scope, dual-timed veracity, and — when the contributor is itself an agent that formed a judgement — that contributor’s own nested provenance (sub_provenance), which is what makesinformed_bya DAG rather than a flat list. - Reliance
Declaration - A standing declaration (capacity-level) of the agents/tools an agent MAY use, so a principal can
evaluate and consent to the toolchain before relying on the agent (§8.2). This is not the per-judgement
record; it is the up-front permission surface against which
has_undeclared_aichecks actual reliance.
Enums§
- Agent
Type - The kind of agent that contributed to a judgement. Weight and liability differ per kind (§8.1):
a natural person bears authorship + responsibility; a software agent is tooling whose use must be
disclosed and scoped but which never bears liability; an instrument has calibration/validation state;
a dataset/source is provenance for input data. Modelled as an enum here (a closed, well-known set);
finer agent-kind facets live in the extensible
taxonomy. - Disclosure
Modality - The proof modality by which a provenance node (or the DAG) is disclosed (§9).
- Input
Veracity - Veracity of a contributed input, over the extended epistemic vocabulary (§10). Used dual-timed:
once for the assessment at the time (against the epistemic horizon) and again for the later,
forensic determination. The default is
InputVeracity::Uncertain— nothing is assumed accurate.
Functions§
- has_
undeclared_ ai - The integrity-breach detector (§8.3): returns
trueif anyAgentType::SoftwareAgentappearing injp.informed_by(recursively, through nestedsub_provenance) is not covered bydeclaration’s permitted set (matched byAgentRef.id). An undeclared AI reliance discovered after the fact is a serious integrity breach — the omission is itself evidence.