Lab experiments
Experiments convert open agent-governance questions into bounded, reviewable work. They are not release promises and they do not create normative protocol requirements until a later governed change explicitly promotes them.
Experiment template
Each experiment should define:
| Field | Purpose |
|---|---|
| Hypothesis | The claim being tested |
| Scope | The smallest useful system boundary |
| Setup | Required registries, agents, keys, fixtures and policies |
| Expected evidence | Files, logs, reports or vectors produced by the run |
| Failure modes | What would disprove, narrow or defer the hypothesis |
| Promotion path | How useful results could become documentation, vectors, profiles or specification text |
Candidate experiments
EXP-001 — Federated registry conflict
Hypothesis: two registries can exchange or project agent state without implying that the receiving registry accepts the source registry’s authority for every relying context.
Evidence target: conflict matrix, competence-preservation rules, example federation records, negative vectors.
Status: designed as a concrete experiment packet.
EXP-002 — Delegated authority separation
Hypothesis: agent identity, operator identity, issuer competence, delegated authority and workflow authorization can be evaluated separately without losing usability.
Evidence target: worked delegated-action scenario, proof-input matrix, positive and negative vectors.
EXP-003 — Historical reconstruction after withdrawal
Hypothesis: requested-time registry state can be reconstructed after later withdrawal, compromise or retirement if provenance, selected records, later material events and reconstruction quality are retained.
Evidence target: reconstruction receipts, later-event disclosure examples, quality-level classification.
EXP-004 — Revocation effectiveness versus enforcement convergence
Hypothesis: revocation effectiveness can be represented separately from distributed enforcement convergence, allowing relying parties to distinguish authority status from operational propagation.
Evidence target: revocation event sequence, enforcement acknowledgement model, stale-reliance negative cases.
EXP-005 — Discovery without authority laundering
Hypothesis: discovery services can improve findability while preserving the rule that discovery is not authority.
Evidence target: search/discovery boundary examples, non-authoritative projection warnings, dereference failure cases.
EXP-006 — Privacy-preserving continuity
Hypothesis: scoped continuity can support audit and redress without creating a universal correlation handle for all agent contexts.
Evidence target: scoped identifier patterns, minimization rules, continuity limits, privacy negative cases.
EXP-007 — External implementation report
Hypothesis: an independently operated implementation can produce a small, repeatable evidence bundle sufficient to compare behavior against ARPA conformance expectations.
Evidence target: implementation report, fixture results, conformance declaration, operator statement and known deviations.
EXP-008 — Assurance composition
Hypothesis: ARPA evidence can feed an assurance process without ARPA claiming to certify the agent, registry, operator or relying-party decision.
Evidence target: assurance-consumer mapping, evidence sufficiency table, non-certification language.
Experiment lifecycle
| State | Meaning |
|---|---|
| Proposed | Idea is captured but not yet executable |
| Designed | Hypothesis, setup and evidence target are clear |
| Running | Implementation or evidence work is active |
| Observed | Results exist and can be reviewed |
| Promoted | Result became a profile, vector, guide, ADR or specification proposal |
| Closed | No further work is planned |
Promotion discipline
An experiment should be promoted only when the evidence clarifies behavior that implementers, operators or reviewers can act on. Negative results are acceptable and should be retained where they prevent over-design or unsafe assumptions.
Related issues
- Supports #76 and #79.