Accountability infrastructure

Accountability for
autonomous AI.

We’re building infrastructure for DeFi AI-agent marketplaces to check agent actions against agreed rules and document evidence, human review and decisions.

In development · Seeking pilot partners

The accountability lifecycle

01 / Define & accept

Agree on the system’s obligations, responsible parties and verification criteria. Record them in a versioned Rulebook: the agreed rules and criteria used to check an event.

Initial MVP scope

02 / Observe

Collect relevant events and supporting evidence. Assemble an Evidence Bundle: the records and sources used to check an event. Record what each source can establish.

Initial MVP scope

03 / Verify & review

Checks return PASS, FAIL or INCONCLUSIVE. A potential violation is recorded as an Accountability Claim for human review. The claim can be challenged before a decision is recorded.

Initial MVP scope: human review

04 / Record & report

Keep the rules, evidence, findings and decisions in a traceable history that reviewers can inspect.

Initial MVP scope: case history
From agreed rules to documented decisions.

01 / The missing layer

Monitoring records actions.
Accountability checks obligations.

A log tells you what happened. Reviewing an action also requires the agreed rule, supporting evidence and a way to challenge the finding.

02 / Illustrative scenario

When an agent exceeds its trading limit.

A synthetic example of the proposed workflow, not a live case or integration.

Obligation
A trading agent must keep each trade within an agreed spending limit.
Observed event
A recorded trade exceeds the limit.
Evidence Bundle
The applicable Rulebook version, trade record and transaction data from an external source.
Automated check
FAIL The recorded amount exceeds the rule’s limit.
Human review
Review the evidence and any challenge before making the final decision.

The proposed process

Four stages, from rules to records.

The initial MVP scope connects agreed rules, evidence linked to its sources, automated checks and human review. Independent corroboration is used where available.

  1. 01

    Define & accept

    Agree on the system’s obligations, responsible parties and verification criteria. Record them in a versioned Rulebook.

  2. 02

    Observe

    Collect relevant events and supporting evidence. Record each source and what it can establish.

  3. 03

    Verify & review

    Check the evidence against the applicable rules. Record the result and submit potential violations for human review.

  4. 04

    Record & report

    Keep the rules, evidence, findings and decisions in a traceable history that reviewers can inspect.

Automated results are not final decisions.

PASS means the observed event meets the checked rule. FAIL indicates a potential violation. INCONCLUSIVE means the available evidence does not support a reliable conclusion.

Potential violations are recorded as Accountability Claims for human review. A claim can be challenged before the final decision is recorded.

Beyond the initial MVP

The initial scope focuses on checks, human review and case history. Arbitration, financial remedies and trust scoring are part of the longer-term roadmap.

03 / Who it’s for

Starting with
DeFi AI-agent marketplaces.

The wider vision includes AI system providers, enterprises and operators.

First application focus

DeFi AI-agent
marketplaces

Make agent obligations and the evidence behind statements about their behavior inspectable.

  • A consistent review process across agents
  • Traceable findings and challenges
  • A history grounded in evidence
Broader vision

AI system
providers

Show whether your systems meet agreed rules.

  • Evidence to support statements about system behavior
  • Context for responding to complaints
  • Records tied to specific rule versions
Broader vision

Enterprises
& operators

Give teams a common way to review autonomous AI activity.

  • Clear operational obligations
  • Evidence available for human oversight
  • Documented findings and decisions

04 / Our team

The people behind DelfiN.

05 / Start a conversation

Help shape the first
DelfiN pilots.

We’re looking for pilot partners working with autonomous AI agents, initially in DeFi. Together, we’ll define a small set of rules, identify evidence sources and agree on a review process.

Discuss a pilot

Interested in investing in DelfiN? Start an investor conversation