Rules v0.1

Definitions

The following terms have the meanings assigned here when used in these Rules. Where a term is given a more specific meaning in a later section (for example, "stratum (b)" in B.4 or "Tier 2 remedy" in B.6), that section governs.

Principal

The natural or legal person on whose behalf an Agent acts in the disputed transaction.

Counterparty

The Principal on the other side of the disputed transaction.

Agent

An autonomous or semi-autonomous AI system acting on behalf of a Principal in commercial transactions.

Operator

The party responsible for deploying, configuring, and maintaining an Agent's runtime environment. The Operator is typically the Principal but may be a separate party (a platform provider, systems integrator, or managed-service provider) where deployment is delegated. Where Operator and Principal are different parties, the contract between them allocates responsibility for stratum (b) Configuration and stratum (e) Instruction; absent such allocation, the Principal bears the configuration risk vis-à-vis the Counterparty.

Mandate

A written instrument signed by the Principal (cryptographically or contractually) authorizing an Agent to act within specified scope. A Mandate must specify: (i) the Agent it authorizes; (ii) the action types within scope; (iii) any monetary, categorical, or temporal limits; (iv) effective and expiry dates; (v) the revocation mechanism. A Mandate is the controlling instrument under B.9.1 where the contract is silent on Agent authority. A Mandate is not itself a contract with the Counterparty and binds the Counterparty only to the extent of apparent authority under B.3.2.

Attribution Service

Any compliant venue applying these Rules to issue an Attribution Record on a referred dispute. Verdra Forum (in formation) will be one such venue. Counsel, insurers, payment networks, arbitral panels, and parties' internal review processes that apply these Rules are also Attribution Services for purposes of these Rules.

Attribution Record

The written determination produced under these Rules. Each Attribution Record identifies: the Category and primary attribution pathway, alternative pathways noted under B.5.3, the evidence strata weighed and the weights applied, the remedy tier, the confidence figure where automated determination applies under B.10, the contestability window, and any prospective remedies ordered under B.6.10 and Annex E.2.

Evidence Package

The cryptographically sealed record of a single Agent commercial action, structured under the six evidence strata (a)-(f) defined in B.4. An Evidence Package may be produced by Verdra Record, by a third-party capture tool, by platform-native logging, or by a party-assembled compilation; under P.4 these sources are treated identically on the face of the record subject to the integrity requirements in B.4.

Established line of decisions

A coherent set of at least three Attribution Records issued under these Rules by independent panels (or by an Attribution Service in successive proceedings) addressing the same interpretive term, where no later Record materially departs from the earlier ones. The corpus of redacted Attribution Records is published per P.7 such that a party can verify the existence of an established line. The B.3.2(c) gate-removal condition for Category 2 disputes requires an established line of decisions on the specific interpretive term at issue. Where two prior Records materially diverge, no established line exists for that term until a subsequent Record reconciles them.

Rehearsal data

A versioned corpus of synthetic and historical disputes, annotated with ground-truth attribution outcomes, used to calibrate the confidence-computation methodology in B.10. Rehearsal data is published with each version of these Rules; parties may inspect, challenge, and propose additions through the contestability mechanism in P.6.

Confidence (in the sense of B.10)

The calibrated probability that the primary attribution pathway identified by the Attribution Service is the correct pathway given the evidence on the record. Confidence is computed by the methodology disclosed alongside Rehearsal data in each version of these Rules. Confidence is distinct from the standard of proof (P.8): confidence routes the determination (automated vs human-escalated under B.10), while the standard of proof governs whether the breach itself is established.

Coverage Matrix

A document published with each version of these Rules listing, per model provider and per protocol, which evidence strata are reliably available and at what fidelity. Parties consult the Coverage Matrix to assess the evidentiary footprint of an Agent deployment before adopting these Rules. See Annex E.1.

Forward commitment

A disclosed limitation in the evidence available to the Attribution Service, recorded in the Attribution Record per P.5 and Annex E.1, expected to become resolved as provider, protocol, or technical limitations are addressed in future versions of these Rules. A Forward commitment is not adverse inference.

Prospective remedy

A non-punitive remedy ordering changes to an Agent's configuration, parameters, Mandate, or operational policies to prevent recurrence of disputed conduct. See B.6.10 and Annex E.2.

Framework

The Rules structure every dispute along three axes: category (what kind of breach is alleged), attribution pathway (how to assign responsibility between Agent, Principal, Operator, and Counterparty), and evidence strata (how to weigh what's on the record).

Four categories

Seven attribution pathways

A primary pathway is identified per Attribution Record. Where multiple pathways are equally plausible on the evidence, B.5.3 governs the combinatorial logic; alternatives are noted in the Record.

Six evidence strata

• (a) Transaction: the protocol-level exchange: POs, acknowledgements, shipment records, substitution notices, timestamps. What the Agents exchanged.
• (b) Configuration: the Agent's system prompt, parameter limits, authorized actions, and approval policies as in force at the moment of action. How the Agent was set up.
• (c) Behavior profile: counterfactual re-queries of the Agent under preserved state; diagnostic of whether the observed behaviour is the modal outcome or an edge case. What the Agent would have done again.
• (d) Mandate: the principal's signed delegation authorizing the Agent to act, with scope, limits, expiry, and revocation in force at the moment of action. What the Principal authorized the Agent to do.
• (e) Instruction: the specific runtime input or trigger received by the Agent at the moment of action, distinct from static (b) Configuration: the prompt, payload, message, or protocol invocation that initiated the disputed action. What the Agent was just told to do.
• (f) Environment: the Agent's view of the world at the time of action: Counterparty status, market data, third-party feed values, time-sensitive context external to the Agent itself. What the Agent perceived around it.

Strata are cumulative, not hierarchical. Strata (a) Transaction, (b) Configuration, and (d) Mandate are production strata captured at dispatch in every Evidence Package. Missing or partial strata (c) Behavior profile, (e) Instruction, or (f) Environment are disclosed as Forward commitments per Rules P.5 and Annex E.1, not treated as adverse inference. The Coverage Matrix (Annex E.1) lists, per model provider and per protocol, which strata are reliably available and at what fidelity.

Principles

P.1 Opt-in. Jurisdiction is consensual. Parties opt in per engagement or by standing agreement. The opt-in instrument names the Attribution Service and incorporates these Rules by reference.

P.2 Contract-first. The Rules fill gaps; they do not override express contract terms. Where applicable mandatory law conflicts with these Rules or with the contract, mandatory law controls.

P.3 Machine speed. Category 1 determinations target sub-five-minute turnaround when administered by Verdra Forum, or by an Attribution Service that has elected to commit to the speed target in its acceptance of the dispute. The target is non-binding on Attribution Services that have not so elected. Human review is reserved for Category 2 until gate-removed under B.3.2(c), for Category 3 always, and for Category 4 pending v0.2.

P.4 Evidence-source neutrality. Evidence from Verdra Record, third-party capture tools, platform-native logging, and party-assembled packages is treated identically on the face of the record, subject to the integrity requirements in B.4.

P.5 Scope honesty. Where strata (c) Behavior profile, (e) Instruction, or (f) Environment are missing or partial in an Evidence Package, the limitation is disclosed in the Attribution Record as a Forward commitment per Annex E.1. No silent assumption that missing evidence is unfavourable.

P.6 Contestability. Every determination carries a contestability window. The default window is 14 days; parties may agree by contract or at the time of filing to a longer or shorter window subject to a 7-day minimum. Grounds: procedural error, new evidence, classification dispute. Standard of review on contest is de novo on the record.

P.7 Publicness. These Rules, the pathway catalogue, the Coverage Matrix, the Rehearsal data, and the corpus of redacted Attribution Records are public. Redactions remove party identifiers and confidential commercial information; the Category, primary attribution pathway, evidence weights applied, remedy tier, and legal reasoning are preserved. Parties may audit the published corpus for consistency, and the corpus is the basis for verifying an Established line of decisions per B.3.2(c).

P.8 Standard of proof. The petitioner bears the burden of proof on the breach alleged. The standard is the balance of probabilities: the petitioner must establish that the alleged breach is more likely than not on the evidence on the record. The B.10 confidence threshold is a separate, automation-routing rule that determines whether a determination issues automatically or escalates to human review; it does not modify the burden or the standard. A determination at confidence above the B.10 threshold still requires that the breach be established on the balance of probabilities, and a determination escalated to human review applies the same standard.

P.9 Independence. Where a dispute is referred to human review under P.3, the reviewing adjudicator is independent of the parties and free of conflicts under the standards customary in international commercial arbitration. Disclosure obligations follow the IBA Guidelines on Conflicts of Interest in International Arbitration where not otherwise specified.

P.10 Procedural fairness. Each party is entitled to present its case, to respond to the other party's case, and to be heard on any matter the Attribution Service relies on in its determination. Where automated determination applies under B.10, the Evidence Package and the Attribution Record together discharge this requirement; either party may invoke the contestability mechanism in P.6 to require human review.

B.1 Principles of interpretation

B.1.1 The Attribution Service applies the governing contract between the parties first. These rules fill gaps, resolve interpretive ambiguity, and allocate evidentiary weight; they do not override express contract terms.

B.1.2 Where an express contract term and an agent's internal configuration conflict, the contract term controls.

B.1.3 An agent's operational policy (including parameters such as timeout-fallbacks, composite-pricing methods, rate-limiter thresholds, or schema-version precedence) is not a contract term unless it is incorporated into the contract by express reference or by a course of dealing evidenced on the record.

B.2 Category 1: Performance

B.2.1 Non-delivery

Failure to deliver the contracted goods, data, or service within the agreed window is non-delivery. Remedy is Tier 1 (refund, replacement at supplier's cost, or proportional credit, at the buyer's election). Where partial delivery occurred, the dispute is assessed under B.2.4 (Quantity) rather than B.2.1.

B.2.2 Service-level breach

Failure to meet a quantified service level (availability, latency, throughput, error rate) measured per the contract's specified method. The contract's measurement method controls; internal supplier measurement methods are not determinative unless incorporated. Remedy is Tier 1 per the contract's credit schedule (or B.9 gap-filling where no schedule exists).

B.2.3 Specification non-conformity

Delivery of goods or services that do not conform to the agreed specification (grade, class, version, configuration). A conforming-delivery requirement is strict unless the contract specifies a tolerance. Where a tolerance is specified, a non-conformity that exceeds the tolerance is a breach. Remedy is Tier 1 per the contract's election clause (replacement, price adjustment, or rejection).

B.2.4 Quantity tolerance

Delivery of a quantity that deviates from the agreed quantity by more than the contractual tolerance. Where no tolerance is specified, a 2 percent tolerance is presumed; this default reflects industry-typical practice in commercial supply contracts and is rebuttable by evidence of contrary industry custom or course of dealing between the parties. Blanket purchase orders are assessed against the cumulative commitment, not against single-release quantities, unless the contract expressly ties tolerance to releases.

B.2.5 Pricing / settlement discrepancy

A settlement or invoice price that deviates from the contract's pricing method. The contract's pricing method controls, including any reference-feed precedence, timestamping policy, or failover conditions. Internal operator-level pricing policies (composite methods, booking-window snapshots) are not contract terms unless incorporated. Remedy is Tier 1 (re-settlement at the corrected price, or unwind at pre-trade mid, at the non-defaulting party's election).

B.3 Category 2: Authority

B.3.1 Scope of mandate

An agent's commitments are binding on its principal to the extent of the principal-signed mandate in force at the time of the commitment. Commitments beyond the mandate are voidable by the principal within the contract's voidability window.

B.3.2 Scope of authority (interpretive ambiguity)

Where the governing contract contains an interpretive term bearing on an agent's authority (for example, "adjacent-category," "strategic-flexibility," "reasonable business judgment," or a quantified phrase such as "does not materially exceed"), and the parties dispute the term's application:

• (a) If the agent's configured parameters plausibly permit the action and the counterparty reasonably relied on the agent's assertion of authority, apparent authority is engaged and the commitment is presumptively binding.
• (b) Apparent authority is rebuttable where the counterparty had actual or constructive notice that the commitment exceeded the contract's authority bounds. A facially material excess (for example, a multiplier materially larger than any prior course of dealing) is a circumstance supporting constructive notice.
• (c) Where the interpretive term has not been resolved by prior decisions under these Rules, the dispute is escalated to human review under P.3. The B.3.2(c) gate-removal condition for Category 2 requires an established line of adjudicator-panel decisions on the specific interpretive term at issue.

B.4 Evidence strata and weight

B.4.1 The six evidence strata are cumulative, not hierarchical. The Attribution Service may weight any stratum by the reliability shown on the record. Strata-level findings establish facts; the legal conclusion that those facts constitute breach is governed by the standard of proof in P.8.

B.4.2 Stratum (b) Configuration evidence is controlling on questions of what the Agent was configured to do. It is not controlling on questions of what the contract requires.

B.4.3 Stratum (c) Behavior profile (counterfactual re-query under preserved state) is controlling on questions of what the Agent would do on the facts. Absence or partial fidelity is a Forward commitment under Annex E.1, not a ground for adverse inference.

B.4.4 Stratum (d) Mandate is controlling on questions of what the Principal authorized the Agent to do. Where the Agent acts within Mandate scope and the action is otherwise consistent with (b) Configuration, the principal is bound under B.3.1. Where the Agent acts outside Mandate scope, B.3 Authority is engaged.

B.4.5 Stratum (e) Instruction is controlling on questions of what the Agent was just told to do. Where stratum (b) Configuration permits an action class but stratum (e) Instruction did not authorize the specific action, the Instruction is controlling on the question of authority for that action.

B.4.6 Stratum (f) Environment is controlling on questions of what state of the world the Agent acted upon. Stratum (f) anchors the analysis of pathway (f) Stale-data dependency and is the basis for assessing the reasonableness of the Agent's reliance on third-party data.

B.4.7 Strata conflict. Where strata point to different conclusions on the same operative question, the Attribution Service applies the following ordering:

• (i) Stratum (a) Transaction is controlling on protocol-level facts (what was exchanged) and is presumed accurate absent evidence of tampering.
• (ii) Where (b) Configuration and (e) Instruction diverge, (e) Instruction is controlling on the specific action; (b) Configuration is controlling on the action class.
• (iii) Where (b)/(e) and (c) Behavior profile diverge - the Agent was configured or instructed to do X but counterfactual re-query shows Y - configuration/instruction is controlling on the question of authority (Categories 2 and 3); the Behavior profile is controlling on the question of causation (Category 1).
• (iv) Where (d) Mandate and (b)/(c) diverge, that is, the Agent's configured tools or observed behavior exceed the Mandate's scope, the Mandate (d) controls; the divergence engages B.3 Authority and the action is voidable per B.3.1.
• (v) Where (f) Environment and any other stratum diverge on a fact about the world (for example, a market price), (f) is presumptively accurate where multiple independent sources are captured; otherwise the most reliable independent source on the record controls.

B.5 Causation

B.5.1 Causation is determinable where the evidence supports a unique attribution pathway with confidence above the B.10 threshold for Category 1, or where the B.3.2(c) gate-removal condition is satisfied for Category 2.

B.5.2 Where multiple pathways are plausible, the Attribution Service identifies the primary pathway and notes the alternatives in the Attribution Record under B.5.3.

B.5.3 Combinatorial pathway logic. Where two or more pathways are equally plausible on the evidence:

• (i) The pathway closest in proximate cause to the disputed action is identified as primary;
• (ii) Pathways earlier in the causal chain are noted as governing alternatives in the Attribution Record;
• (iii) Where the primary pathway shifts the remedy tier compared to an alternative, the Attribution Service describes the differential and applies the more conservative tier where the conservative tier is materially more favorable to the non-defaulting party;
• (iv) Pathways (e) Adversarial manipulation and (g) Inter-agent coordination failure, where plausible alongside pathways (a)-(d) or (f), are independently noted regardless of primacy because they affect both the remedy and the prospective remedies under Annex E.2;
• (v) Where pathway (e) Adversarial manipulation is attributed to a party with intent, the dispute is reclassified to Category 3 and human review is engaged.

B.5.4 Force majeure and external causation. Where the Agent's failure is materially caused by an external event outside the Operator's reasonable control (third-party outage, force-majeure event, regulatory intervention), the Attribution Service notes the external causation as a defense to liability, allocates remedies in accordance with the contract's force-majeure provision where one exists, and otherwise applies commercial-reasonableness defaults under B.8 and B.9. External causation does not extinguish a Forward commitment under Annex E.1 to improve the Evidence Package's coverage of stratum (f) Environment.

B.6 Remedies

B.6.1 Tier 1 remedies are Category 1 performance remedies: credit, refund, replacement, re-settlement, price adjustment. Automated under the B.10 confidence threshold.

B.6.2 Tier 2 remedies are Category 2 authority remedies: voidance, ratification, authority clarification. Human review until B.3.2(c) gate-removal for the specific interpretive term.

B.6.3 Tier 3 remedies (fraud findings, reputational sanctions, exclusion from the protocol) are reserved for Category 3 intentional misconduct; always human-escalated.

B.6.4 Category 4 remedies are routed to human review pending v0.2 and adjudicated against the contract supplemented by applicable confidentiality, data-protection, and intellectual-property law.

B.6.5 Consequential damages. Consequential damages (loss of profits, loss of opportunity, loss of reputation, indirect or downstream losses) are available where the contract permits and not foreclosed by liability caps. Where the contract is silent on consequential damages, the default is exclusion: only direct damages are awarded under these Rules absent express contractual inclusion. The non-prevailing party may petition under P.6 to argue contractual inclusion or applicable-law inclusion of consequential damages.

B.6.6 Specific performance. Specific performance (an order to perform or refrain from a specific action) is available as a Tier 2 or Tier 3 remedy where damages are inadequate and the action is technically feasible. Specific performance is not available against a Counterparty's autonomous Agent without that Counterparty's consent or a separate enforcement mechanism, but is available as an order to the Operator to reconfigure or restrain the Agent.

B.6.7 Interim relief. Pending full adjudication, the Attribution Service may issue interim orders for: (i) evidence preservation; (ii) Agent suspension or scope restriction; (iii) status-quo holds on disputed transactions; (iv) escrow of disputed funds. Interim orders issue on a showing of (a) prima facie merit, (b) risk of irreparable harm absent relief, and (c) balance of harms favoring relief. Interim orders are reviewable de novo at full adjudication and are reviewable on contestability under P.6.

B.6.8 Costs. Each party bears its own costs (counsel, expert, evidence-assembly) absent contractual provision. Administrative fees of the Attribution Service are allocated to the non-prevailing party absent contractual provision; where a determination is mixed, administrative fees are allocated proportionally to the relative success of each party's claims.

B.6.9 Set-off. Where both parties have claims against each other arising from the same or connected transactions, the Attribution Service may net the awards. Set-off is not available across unrelated transactions absent the parties' agreement or applicable-law authorization.

B.6.10 Prospective remedies (configuration changes, parameter updates, Mandate amendments, evidence-capture upgrades) are available under Annex E.2 as Forward-commitment remedies, disclosed in the Attribution Record as non-punitive alignment measures and available alongside Tier 1, 2, or 3 remedies.

B.8 Contract silence: Category 1 gap-filling

Where the contract is silent on a Category 1 issue, the Attribution Service applies commercial-reasonableness defaults:

• (a) Quantity: a 2 percent tolerance (B.2.4).
• (b) Service levels: proportional credit at (missed / contracted) × period-fee.
• (c) Specification: strict conformance absent express tolerance.

B.9 Contract silence: Category 2 gap-filling

B.9.1 Where the contract is silent on the scope of an agent's authority for the action in dispute, the Attribution Service applies the principal-signed mandate as the controlling instrument.

B.9.2 Where neither the contract nor the mandate addresses the action, the dispute is escalated to human review under P.3.

B.10 Confidence threshold

B.10.1 The Attribution Service issues an automated final determination on a Category 1 dispute only where confidence is at or above 0.85. Below that threshold, the dispute is escalated to human review under P.3.

B.10.2 Category 2 disputes are human-escalated by default pending B.3.2(c) gate-removal for the specific interpretive class at issue.

B.10.3 Confidence is not the standard of proof. The B.10 threshold determines whether the Attribution Service issues automatically or escalates to human review; it does not modify the standard of proof in P.8. A determination at confidence above 0.85 still requires that the breach be established on the balance of probabilities; a determination escalated to human review applies the same standard.

B.10.4 Methodology. Confidence is the calibrated probability that the primary attribution pathway identified by the Attribution Service is the correct pathway given the evidence on the record. The computation methodology is disclosed in the methodology document published alongside Rehearsal data with each version of these Rules. The methodology is testable: a party may, on contestability under P.6, challenge the methodology's calibration by reference to the published Rehearsal data.

Contestability

Every determination carries a 14-day contestability window with three grounds:

• Procedural error: the adjudicator misapplied the Rules or omitted required steps.
• New evidence: material evidence unavailable at filing that would alter the pathway or remedy.
• Classification dispute: the dispute was classified into the wrong Category or pathway.

A timely contest re-opens the record for adjudicator review. Contest decisions are final within the adjudicative process under these Rules; parties retain all rights to judicial review under applicable law.

License & citation

The Verdra Rules are published under Creative Commons BY-SA 4.0. You may use, redistribute, and build on them, including for commercial purposes, on two conditions: credit Verdra, and share derivatives under the same license. Pre-launch versions are numbered v0.x; v1.0 ships at public launch. PDF available on request.

Suggested citation: Litwin, D. (2026). Verdra Rules: A Framework for Inter-Agent Commercial Dispute Resolution (Version 0.1, pre-launch). Verdra. https://verdra.co/rules.html

The framework, working papers, and version roadmap are at Research. Comments on the rulebook welcome at hello@verdra.co.