Arbitrator AI — Rules of Procedure & Evidence

Version 0.10 (DRAFT) · Effective: [DATE on adoption] · Status: internal draft, not yet in force · v0.2 (2026-07-04): Rule 4.11 protocol-native claims — operator-config consent basis + escrowed_refund_request class · v0.3 (2026-07-04): Rule 4.11 fast rung DECIDES eligible micro-claims — confidence/conflict recorded at sign-off, never re-routed · v0.4 (2026-07-04): Rule 8.7 additional award — a pleaded-but-omitted claim decided post-award without reopening · v0.5 (2026-07-08): evidence integrity & conduct — Rule 5.7 challenge raising flow · Rule 5.14 truthfulness attestation · Rule 2.7 reviewer neutrality/recusal + named reviewer (Rule 2.5) · Rule 4.10 live testimony by Reviewer direction on escalation · Rule 9.5 exclusion from the forum · Appendix E canonical award-hash recipe · v0.6 (2026-07-08): Rule 8.8 optional paid appeal to a human appellate reviewer — the Rule 8.1 challenge window becomes its live mechanism · v0.7 (2026-07-10): Rule 10.1 simplified to plain fee-for-service wording — the fee is charged separately, never taken from the escrowed funds; standalone scope-guardrail doc retired, references removed · v0.8 (2026-07-10): Rule 4.11 ACP evaluation channel — evaluator-designation consent basis + acp_evaluation class; the job's own artifacts are the exchanged record (evaluation proceeds directly to the merits); execution on this rail is binary complete/reject, anything else escalates to the Reviewer · v0.9 (2026-07-10): Rule 10.1 fee schedule — filing / Rule 8.8 appeal / evaluation fees, flat or the greater of a published percentage of the amount in dispute and a published per-tier minimum, fixed and quoted at filing, charged separately, never outcome-linked, never from the corpus; unpaid filing fee ⇒ never docketed; Rule 4.7 upward amendment tops up · Rule 10.3 costs allocation — the Tribunal may allocate the fee burden between the parties in the Award (default filer-bears; unavailable on corpus-only awards) · v0.10 (2026-07-10): Rules 1.1/1.3/3.2 dynamic incorporation — consent may identify the governing version by the published identifier arbitral-rules-current (the Rules in effect at filing); the concrete version + rules-text hash stamped at commencement resolve the reference and govern the matter end to end; amendments never apply to pending matters

DRAFT — NOT LEGAL ADVICE — REQUIRES LICENSED COUNSEL. This is the procedural rulebook for the Arbitrator AI process — the analogue to the AAA Commercial Rules, JAMS Rules, or the UNCITRAL Arbitration Rules, adapted to an AI-tribunal, on-chain-escrow system. It is a structural draft. Every Rule with enforceability weight carries a COUNSEL flag (an attorney must finalize) or CALIBRATE flag (a default value that must be tuned against real dispute data and the eval harness). Sample numeric defaults are illustrative.

Eligible parties: open to businesses, individuals, and autonomous agents alike (PartyType = human | entity | agent). Nothing here forecloses any class of party. These Rules govern consent-based arbitration between the actual disputing parties over their own live dispute. They do not create, and must never be read to create, any market in which third parties wager on, fund, or profit from outcomes.

COUNSEL] — consumer note (risk carried, not a gate): binding a consumer to arbitration pre-dispute draws heightened scrutiny (unconscionability, disclosure duties, jurisdictional limits on pre-dispute consumer arbitration). The process therefore supports both pre-dispute consent (captured at case creation) and post-dispute consent (captured after the dispute exists) — post-dispute consent is the lower-risk path for individuals. The decision to serve consumers is the operator's; this note preserves the ability to tighten per counsel without re-architecting.

Relationship to the consent artifact: These Rules have no force on their own. They bind only because the parties' Consent to Arbitrate (the load-bearing artifact) incorporates a specific, identified version of them. The consent agreement is the source of authority; these Rules are its procedural body.

How these Rules are organized

• Article 1 — Scope, application, version-binding, definitions
• Article 2 — The Tribunal (single arbiter or AI panel + human reviewer)
• Article 3 — Commencement & jurisdiction (the pre-merits threshold)
• Article 4 — Written submissions (the phased proceeding: complaint → answer → simultaneous evidence → rebuttal)
• Article 5 — Evidence (admissible classes, provenance, lifecycle, burden)
• Article 6 — Decision & the Award
• Article 7 — Escalation to a human
• Article 8 — Execution & finality
• Article 9 — Integrity, conduct & sanctions
• Article 10 — Fees & administration
• Article 11 — General provisions
• Appendices — default schedule, reliability tiers, escalation triggers, Rule→engine map, award-hash recipe

Rules are numbered Article.Rule (e.g., Rule 5.3) so the engine and the opinions can cite them directly. The current engine implementation of each Rule is mapped in Appendix D; Rules not yet enforced in code are marked (not yet enforced).

ARTICLE 1 — SCOPE, APPLICATION & INTERPRETATION

Rule 1.1 — Application

These Rules govern any dispute the parties have agreed, in a Consent to Arbitrate, to resolve through the Arbitrator AI process. They apply as in force in the version identified in that consent — identified directly by number, or dynamically per Rule 1.3.

Rule 1.2 — Eligible parties and disputes

The process is open to any party — business, individual, or autonomous agent — that gives valid consent (Rule 3.2). Covered disputes are those arising out of or relating to the specific engagement identified in the consent (e.g., a smart-contract escrow, statement of work, or services engagement identified by contract hash). A party's use of the process is not a substitute for independent legal advice. Where a party is an individual consumer, post-dispute consent is the default and the operator should provide consumer-appropriate disclosures. COUNSEL — see the consumer note in the header; the pre-dispute consumer-binding posture must be confirmed by counsel before binding use.

Rule 1.3 — Version-binding (load-bearing)

A dispute is governed by the version of these Rules in force at the moment the Consent to Arbitrate was executed (or, for an existing relationship, the version the parties expressly adopt). Amendments to these Rules do not apply retroactively to a pending or already-consented dispute. The governing Rule version is recorded on the dispute and on the Award. (not yet enforced — see backlog: "case bound to filing snapshot / versioned ruleset")

Dynamic incorporation (new v0.10). A Consent to Arbitrate may identify the governing version by reference instead of by number: the published identifier arbitral-rules-current incorporates the version of these Rules in force at the moment the dispute is filed — the standard institutional clause form ("under the Rules in effect at the time the arbitration is commenced"). On commencement the engine stamps the concrete version and rules-text hash onto the dispute; the stamp is the resolution of the reference, and that stamped version governs the matter end to end. The amendment power behind a dynamic reference is bounded: amendments take effect only on publication, apply only to matters filed after adoption, and pending matters are immune — this pending-matter immunity is what keeps the dynamic reference from operating as an unbounded unilateral power. COUNSEL — enforceability of dynamic incorporation against an illusoriness or unconscionability challenge; whether consumer-party consents require a pinned version or additional disclosure at signing

Rule 1.4 — Definitions

• Tribunal — the AI adjudicator(s) assigned under Rule 2.1: one neutral AI arbiter for an eligible agent-lane matter, or the three-agent Panel.
• Panel — the three AI agents assigned to human/entity-track matters and high-stakes agent-lane matters (Article 2).
• Reviewer — the qualified human who signs off where required (Rule 2.5).
• Claimant / Respondent — the party initiating / responding to the dispute.
• Corpus — the funds under the tribunal's control (typically escrowed) that an award may allocate.
• Considered record — the single canonical set of admitted facts and exhibits the Tribunal is permitted to rely on (Rule 5.5). Anything not in the considered record does not exist for decisional purposes.
• Award — the operative, validated disposition of the dispute (Article 6).
• Award posture — what kind of award it is, independent of who wins (Rule 6.5).

Rule 1.5 — Good faith, waiver, and interpretation

The parties shall participate in good faith. A party that proceeds without objecting to a known departure from these Rules waives that objection. These Rules are interpreted to secure a just, expeditious, and economical resolution; where a situation is not expressly provided for, the Tribunal (or Reviewer) shall act in the spirit of these Rules.

ARTICLE 2 — THE TRIBUNAL

Rule 2.1 — Composition

Composition is determined deterministically from party type, track, and amount in dispute before the merits:

• Agent lane below the panel threshold — one AI arbiter. Where both parties qualify for the agent lane under Rule 4.9 and the amount in dispute is below the panel threshold (default $10,000 [CALIBRATE], configured as SIGNOFF_THRESHOLD_USD), one neutral AI arbiter reviews the considered record and issues the reasoned Award. Rule 4.11 may resolve an eligible micro-claim at an earlier oracle or fast-memorandum rung; if that rung cannot decide, the matter remains with one arbiter for full-record review.
• Panel matters — three AI agents. Human/entity-track matters, and agent-lane matters at or above the panel threshold, are heard by the fixed three-agent Panel. Each agent independently reviews the considered record and issues its own reasoned opinion. The outcome is the decision of the majority (Rule 6.3); any concurrence or dissent is recorded and published (Rule 6.4).

Composition is not changed merely because an AI adjudicator expresses low confidence or a micro-claim exits its fast rung. A below-threshold agent-lane case proceeds to fuller single-arbiter review; it does not become a three-agent case. Validation or integrity failure routes to the human Reviewer (Article 7), not to a wider AI bench. CALIBRATE — confirm the $10,000 composition threshold against eval results, dispute economics, and counsel advice.

Rule 2.2 — Declared interpretive methodologies

For Panel matters, each agent operates under one declared, fixed interpretive stance applied to the governing text:

• Textualist — plain meaning of the words as written; reluctant to look past the four corners of the agreement.
• Purposivist — the parties' intent and the purpose of the bargain; what the deal was for.
• Pragmatist — workable outcomes, custom, and reasonable-party expectation in the relevant ecosystem.

The Panel composition is fixed; parties do not select or tune the agents that hear their dispute (anti-forum-shopping). The single arbiter is likewise fixed and neutral; the parties do not select its model or methodology. CALIBRATE — fixed vs. party-influenced composition is an open design question; default is fixed.

Rule 2.3 — Independence of the agents

In a Panel matter, each agent reaches its own conclusion and issues its own final opinion. Independence permits one structured deliberation round (Rule 6.1): after the agents form independent first-round opinions, each is shown the others' dispositions and reasoning and must engage any materially different remedy before finalizing. Deliberation is informed engagement, not bargaining — there is no negotiation, vote-trading, join/dissent bargaining, or convergence to a compromise position, and the panel result is still computed deterministically from the agents' actual final opinions (Rule 6.3). A single-arbiter matter has no synthetic deliberation round and no second AI merits reviewer; one AI judge owns the decision.

Rule 2.4 — Competence to rule on jurisdiction (kompetenz-kompetenz)

The Tribunal may rule on its own jurisdiction, but only within the limits of Article 3: jurisdiction rests on a valid bilateral Consent to Arbitrate, verified deterministically. Absent valid consent, the Tribunal declines before reaching the merits (Rule 3.2). COUNSEL — the scope of any delegation of arbitrability to the Tribunal must be drafted "clearly and unmistakably."

Rule 2.5 — Human reviewer and sign-off

In V1, no Award is final until a qualified human Reviewer reviews and signs off. The Reviewer's role is supervisory (process integrity, escalation, safety) and does not substitute the Reviewer's merits judgment for the Tribunal's except on escalation (Article 7). The scope of automatic, no-sign-off resolution expands only from the evidence the eval harness produces, never by default. Sign-off is a single act that also serves the decision: on signing, the Reviewer adopts (or edits) the reasoned Award (Rule 6.9) and releases it to the parties in the same step — sign-off and party-release are not separate gates (Rule 6.10). The sign-off records the Reviewer's identity, and the served Award identifies the reviewing arbitrator by name — e.g., Signed by Arbitrator [name] (Rules 2.7, 6.9); an unnamed sign-off is reserved for non-binding testing modes. Serving the decision is distinct from executing it (Rule 8.1). COUNSEL — who the Reviewer is (the operator, a contracted arbitrator, a party-selected neutral) carries different enforceability consequences.

Rule 2.6 — Impartiality, disclosure, and model provenance

The same assigned composition and model configuration are applied to both sides; no adjudicator is tuned to, or briefed privately by, either party. The model provenance, composition decision, and configuration in force for a dispute are recorded and disclosable. Transparency of process (reasoned opinions, published dissents where a Panel sits, disclosed methodology) is itself a safeguard against "black-box" challenge.

Rule 2.7 — Reviewer neutrality, disclosure, and recusal

The human Reviewer (Rule 2.5) owes the parties independence and impartiality, continuing from assignment through the final disposition of the matter. A Reviewer shall not act in any matter in which the Reviewer holds a financial or personal interest in the outcome, or has a relationship with a party, a party's principal or representative, or the subject matter, that could give rise to justifiable doubts as to the Reviewer's impartiality or independence.

• Attestation and disclosure. Before acting on a matter, the Reviewer attests to the absence of disqualifying conflicts and discloses any circumstance that could give rise to justifiable doubts; the attestation and any disclosure are recorded docket events.
• Recusal. A Reviewer who becomes aware of a disqualifying circumstance recuses; the administrator reassigns the matter to a different Reviewer. Recusal and reassignment are logged.
• Party objection. A party may object to a Reviewer on justifiable-doubt grounds; the administrator resolves the objection and reassigns where the doubt is justified. An objection is not a merits event and does not stall execution safeguards.
• Attribution. The identity of the signing Reviewer appears on the served Award (Rule 2.5).

COUNSEL — the operative conflicts standard (cf. FAA §10(a)(2) "evident partiality," UNCITRAL Model Law art. 12 "justifiable doubts," AAA/ABA Code of Ethics, IBA Guidelines on Conflicts of Interest) and its application where the Reviewer is operator-employed.

ARTICLE 3 — COMMENCEMENT & JURISDICTION (PRE-MERITS THRESHOLD)

The threshold runs after triage, before the Tribunal reaches the merits. Its three questions are never collapsed into one.

Rule 3.1 — Commencement

A dispute commences when a party files a notice identifying the engagement, the relief sought, and the basis of the claim, and the matter is registered. Pleaded relief is captured at filing and bounds the remedy the Award may grant (Rule 6.8).

Rule 3.2 — Consent verification (deterministic; declines if defective)

Jurisdiction is verified deterministically against the Consent to Arbitrate of both parties. A consent entry is valid only if it agrees to arbitrate, names a covered scope and a Rule version (directly by number, or by the dynamic identifier per Rule 1.3), and its content hash verifies (tamper-evidence). If consent is missing, defective, or tampered, the Tribunal declines and the corpus is left untouched. This is the line that anchors the system as a consent-based arbitration — only the parties who consented are bound, and only their own dispute is decided.

Acceptable consent methods (any one suffices; more is stronger). Consent may be evidenced by (a) a wallet signature over the consent payload (EIP-191 personal_sign; the address, signature, and signed message are recorded and re-verifiable), (b) an explicit click-through acceptance bound to the party's case identity and timestamped, or (c) other attested forms (email/signed document) as configured. A wallet signature is one method, never the sole gate — it is strong for on-chain self-execution but, on its own, thin for off-chain/human recognition (see enforceability-strategy.md). Each party links its own consent to its own side of the case (Rule 4.1).

Implied-consent testing mode: for non-binding testing, the engine may treat both parties' participation as implied consent and proceed, but the Award is flagged "implied consent — capture signed attestations before any binding use." Implied consent is never sufficient for a binding award. COUNSEL

Rule 3.3 — Scope / arbitrability

The Tribunal proceeds only if the dispute falls within the covered scope named in the consent. A matter outside scope, or implicating the rights of non-consenting third parties, is declined or escalated.

Rule 3.4 — Affirmative defenses (pleaded only; no sua sponte decision)

The Tribunal decides an affirmative defense only if a party pleaded it. A potentially dispositive defense that appears on the facts but was not raised is flagged for review, never decided by the Tribunal on its own initiative. The tribunal does not raise unpleaded affirmative defenses sua sponte.

Rule 3.5 — Framing (advisory only)

An advisory framing step may surface the dispositive question — including where the parties are arguing past the real issue — and may recommend escalation. It never silently decides the dispute.

ARTICLE 4 — WRITTEN SUBMISSIONS (THE PHASED PROCEEDING)

Rule 4.1 — The phased proceeding (overview)

A dispute proceeds through ordered phases, advancing only when the current phase closes. The proceeding deliberately combines two structures, each chosen for the property it secures:

• Sequential at the pleading stage — the Complaint precedes the Answer, so the Respondent knows the specific charge it must meet (notice / opportunity to be heard).
• Simultaneous at the evidentiary stage — each side's evidence and memorial are submitted blind and revealed together, so neither side reverse-engineers its whole case around the other's, and neither gets an unanswered "last word" (Rule 4.4).

The default standard-track phases are:

Phase 5 (Close) has two steps that run in order: first the record-validation checkpoint — each party is shown a neutral restatement of its own submission and confirms or corrects it (Rule 5.11); then, once both parties have confirmed (or the confirmation window lapses), the considered record is frozen (Rule 5.5) and the merits run (Rule 6.1).

A dispute may instead be filed in quick mode (structured positions + evidence references, for fast/low-stakes matters), which collapses Phases 1–4 into a single exchange and normalizes into the same considered record (Rule 5.5). Submission mode runs the full phased sequence above. The same phases run on every track — the track sets only the clock (Rule 4.6): a standard track wrapping within a month, a fast lane, or an agent lane that compresses the identical phases to hours for agent-vs-agent disputes (Rule 4.9).

In all modes the claimant opens the case and files only its own side; the respondent joins that case (by invitation) and files only its own side. No party files for, or on behalf of, the other. Each party may view and act on only the case(s) it is a party to, and — in a simultaneous phase — sees the other side's filing only on reveal (Rule 4.4). When the proceeding reaches Close, both consents are valid (Rule 3.2), and the record is frozen (Rule 5.5), the matter is ready and the Tribunal runs without an operator step (Rule 6.1).

Rule 4.2 — Phase 1: Complaint (Claimant)

The Complaint commences the dispute (Rule 3.1) and is filed by the Claimant alone. It shall state: the parties and engagement; the facts relied on; the legal/contractual basis; and the specific relief sought (amount and form). The relief pleaded here bounds the remedy any Award may grant (Rule 6.8). Evidence supporting the Complaint is tendered in Phase 3, not here; the Complaint's role is to define the claim, not to prove it. Until the Complaint is filed and notice issues to the Respondent, no later phase opens.

Rule 4.3 — Phase 2: Answer & joinder of issue (Respondent; counterclaims)

The Respondent, having seen the Complaint, files an Answer that admits or denies each material allegation, states its own position, and pleads any affirmative defenses (Rule 3.4) and any counterclaim with its own relief sought. The Answer performs joinder of issue: the set of allegations actually contested — and any pleaded defenses and counterclaims — defines the issues the Tribunal will decide. Matters admitted are not re-litigated; matters not pleaded are not decided sua sponte (Rule 3.4). A counterclaim runs on the same phased schedule, mutatis mutandis. If the Respondent does not answer, Rule 4.9 (default) governs.

Rule 4.4 — Phase 3 & 4: Simultaneous exchange (locked submission, joint reveal)

Phase 3 (Evidence & memorials) and Phase 4 (Rebuttal) are simultaneous. Within a simultaneous phase:

• Each party submits independently. Its submission is sealed — content-hashed (Rule 5.4) and withheld from the opposing party — until the reveal trigger.
• The reveal trigger fires when both parties have submitted, or the phase deadline passes, whichever is first. On reveal, both submissions are disclosed at once. No party sees the other's submission before its own is sealed.
• A party may amend its own submission only before it is sealed; sealing is irrevocable for that phase (cf. Rule 4.7).
• If one party seals and the other does not by the deadline, the sealed submission is revealed and the non-filing party forfeits that phase only (Rule 4.8); the proceeding continues.

Phase 3 carries each side's evidence (Article 5) and its memorial — the narrative tying facts to the pleaded issues. Phase 4 (Rebuttal) is a single round confined to responding to the other side's Phase 3 submission: no new claims, no new relief, no new affirmative defenses. Rebuttal may be omitted for a matter the Tribunal or Reviewer flags as not needing one. Evidence first disclosed at reveal becomes decisional through the normal lifecycle (Rule 5.3); the simultaneous reveal is the exchange event for that phase (Rule 5.5).

Rule 4.5 — Length and form limits CALIBRATE

To keep the record decidable and costs bounded (defaults; configurable in the consent):

• Complaint / Answer (pleadings): 1,500 words each — these define the dispute, they do not argue it.
• Phase 3 memorial (each side): 3,000 words.
• Phase 4 rebuttal (each side): 1,500 words.
• Each exhibit must carry a one-line description; a party tendering a long document should cite the pinpoint passage relied on (Rule 6.7 enforces pinpoint-cite integrity).
• Submissions exceeding a limit may be truncated to the limit for decisional purposes; the Tribunal is not obliged to consider text beyond it.

Rule 4.6 — Schedule CALIBRATE (default timeline; Appendix A)

The phases are fixed; the clock is set by the track. Each phase deadline is a date certain computed from commencement. The default standard track runs one phase per week and wraps within a month:

• Day 0 — Complaint filed (Phase 1); notice to Respondent.
• Day 7 — Answer due (Phase 2); issues joined.
• Day 14 — Evidence & memorials due (Phase 3); locked submissions revealed together.
• Day 21 — Rebuttal due (Phase 4); locked submissions revealed together.
• Day 28 — exchange closed, record frozen (Phase 5 / Rule 5.5); Award targeted.

A fast lane (low-stakes, low-complexity matters flagged at triage) compresses these intervals and may drop Phase 4. An agent lane (Rule 4.9) compresses the same phases to hours for agent-vs-agent disputes. A party may request a reasonable extension for good cause. Time limits are configurable per dispute in the consent but should not be so short as to deny a fair opportunity to be heard. COUNSEL — adequate opportunity to be heard is a due-process / enforceability concern; the agent lane's compressed clock is sound only where both parties are agents that consented to it.

Rule 4.7 — Amendments

A party may amend a pleading or submission only before the phase it belongs to closes — and, in a simultaneous phase, only before its own submission is sealed (Rule 4.4). After the phase closes, amendment requires the Tribunal or Reviewer's leave for good cause, with an opportunity for the other side to respond. New claims or new relief may not be introduced after Phase 1, nor new affirmative defenses after Phase 2, except by such leave.

Rule 4.8 — Default / non-participation

Default is handled per phase, never as an automatic loss:

• Full default (no Answer). If the Respondent, after valid notice, does not answer by the Phase 2 deadline, the proceeding continues on the Complaint and the Claimant's evidence alone. Non-participation is not treated as an admission: the Claimant must still carry its burden of proof on the record it submits (Rule 5.8), and a default award is tested by the same validation (Rule 6.6) and escalation (Article 7) gates as a contested one. This is the canonical single-party path — e.g., a crypto bounty non-payment where the payer never appears — and a thin or high-stakes default record is an escalation trigger.
• Partial default. A party that pleads but skips a later simultaneous phase forfeits that phase only (Rule 4.4); the Tribunal decides on what was filed.
• A defaulting party retains the right to be heard until its phase closes; late participation before the phase deadline is accepted. COUNSEL — default procedure and adequate notice bear directly on enforceability.

Rule 4.9 — Agent lane (agent-vs-agent disputes) CALIBRATE

Where both parties are autonomous agents (PartyType = agent on each side) and the consent opts into the agent lane, the dispute runs the same five phases and the same locked-submission / joint-reveal mechanics (Rule 4.4) as any other matter. The clock is compressed, and composition follows Rule 2.1: one neutral AI arbiter below the panel threshold; the three-agent Panel at or above it. Nothing about the agent lane relaxes the integrity, validation, or escalation gates.

Both kinds of agent are eligible: an agent acting for itself (e.g., a protocol, treasury, or autonomous service that is the real party) and an agent acting on behalf of a human or entity principal. Eligibility turns on both sides being agents and opting in — not on whom the agent represents. The distinction matters only for consent, not access: where an agent acts for a principal, the principal's informed consent to the compressed clock (not merely to arbitration) must be captured, because the opportunity-to-be-heard interest is the principal's, not the agent's. A principal who has not consented to the compressed timing is run on the standard track. COUNSEL — for principal-representing agents, confirm the compressed clock still affords the principal adequate opportunity to be heard; this is a consent-capture requirement, not a bar.

• Same structure, compressed clock. Default agent-lane deadlines, from commencement: Complaint T+0 → Answer T+4h → Evidence & memorials T+12h → Rebuttal T+18h → Close & Award T+24h. A dispute between two cooperating agents is therefore resolvable within a day; the marks above are configurable in the consent.
• Service is machine-to-machine. The Complaint is delivered to the Respondent agent's registered endpoint/wallet (Rule 11.2); "Agent A files and the system serves Agent B" is the notice step, automated. No human relay is required for the proceeding to advance.
• Reveal-on-both-submitted does the real work. Because the joint reveal fires as soon as both agents have sealed (Rule 4.4), two responsive agents finish a phase in minutes, not at the deadline. The hour marks are backstops — they exist only to stop a stalling agent from holding the dispute open, and they trigger per-phase default (Rule 4.8) if missed. Speed is a property of the parties, not a concession by the process.
• Proportional composition. Below the panel threshold, one neutral AI arbiter owns the merits decision; there is no second AI merits reviewer and no three-way vote. At or above the threshold (default $10,000), the fixed three-agent Panel sits. The threshold is deterministic and cannot be waived by triage or model confidence.
• No bypass of the human gate. The agent lane changes speed and, below threshold, composition — not safeguards. A validation failure (Rule 6.6), integrity issue, or attributable bad faith still escalates to the human Reviewer (Article 7) and still cannot self-execute an invalid Award. A fast dispute is not an unreviewed one. COUNSEL — confirm the compressed timing still affords adequate opportunity to be heard for a binding award; for principal-representing agents this is governed by the consent-capture requirement above.

Rule 4.10 — Documents-only presumption (live testimony exceptional, Reviewer-directed) (amended v0.5)

The process is documents-only by presumption: the dispute is decided on the written submissions and admitted exhibits (Article 5) alone, and in the ordinary course there is no live or oral hearing, no live witness testimony, and no cross-examination. A party's account of events enters the record as a written, party-asserted statement (weighted under Rule 5.2), not as live testimony. The presumption keeps matters within the validated, record-bound set the Tribunal decides reliably (Rule 5.10); the AI Tribunal never conducts, attends, or receives live testimony — a pure credibility contest the written record cannot resolve remains an escalation trigger (Rule 7.1), never something the Tribunal resolves by hearing a witness.

The exception sits with the human Reviewer, on escalation only. On an escalated matter (Article 7), the Reviewer may direct a live, recorded testimony session — conducted by the Reviewer over video or equivalent means, on notice to both parties, limited to the questions the written record cannot resolve (typically an unresolved credibility contest or a material authenticity challenge, Rule 5.7). The direction, the session record, and the Reviewer's resulting findings are logged; the findings enter the matter through the Reviewer's reasoned decision (Rule 7.3), never by amending the frozen considered record. A party's unexcused failure to appear after direction may be weighed by the Reviewer in resolving the question the session was directed to answer. Most matters — including most escalated matters — conclude without any session; the direction is expected to be rare.

COUNSEL — due-process adequacy of a Reviewer-directed session for binding awards; the limits of "compelled" appearance under a consent-based process; recording-consent law by jurisdiction. CALIBRATE — whether recorded/asynchronous witness statements are admitted short of a live session; the default remains documents-only.

Rule 4.11 — Micro-claims fast track CALIBRATE — consent and routing defaults require counsel sign-off before production use

Where both parties are autonomous agents acting for themselves, both have given informed consent to this tier specifically (distinct from consent to arbitrate and from the compressed clock, Rule 4.9), the claim is filed as one of the enumerated structured claim classes (free-text claims never fast-track), and the amount in dispute does not exceed the micro-claims ceiling (default $25 [CALIBRATE]), the matter may run the micro-claims fast track — a proportionality cascade that changes the compute, never the guardrails:

1. Oracle-first. A claim class whose dispositive fact is verifiable from the settlement record (e.g. whether the escrow window lapsed) is resolved deterministically from that fact — no model decides. Two limits: the oracle only affirms from facts present in the record (the absence of a fact proves nothing and falls through), and the affirmed fact must itself entitle a party to the corpus. A verified fact that merely refutes the other side's premise decides nothing — the claimant may still be owed relief on the wider record — and falls through to the next rung.
2. Fast single call. Otherwise a single arbiter issues one bounded memorandum verdict: disposition, funds, confidence, and a reasoned basis. This is the deciding rung for an eligible micro-claim — low verdict confidence and evidentiary conflict are recorded on the memorandum and weighed at sign-off (Rule 7.1); they never re-route the matter to a fuller review. Close calls are resolved by burden allocation, not left open.
3. Full single-arbiter review. A fast-rung trigger routes a below-threshold agent-lane matter to the neutral single arbiter for a complete review of the considered record and a full reasoned opinion. The triggers are eligibility, integrity, and lawful-award failures only: value above the micro-claims ceiling (but below the panel threshold) · tier consent or enumerated claim-class requirement not satisfied · quarantined evidence or attributable injection · a split or conditional fast verdict · an award exceeding the claim · accounting-validation failure · rate-limit or ceiling-probing patterns. Merits difficulty is not a trigger (v0.3): an eligible micro-claim is decided at the fast rung whatever its confidence or conflict profile. This is a compute/routing decision, never a decision on the merits. It does not convene the three-agent Panel. A matter at or above the Rule 2.1 panel threshold goes directly to that Panel.

• Safeguard parity (Rule 4.9 discipline). The fast-track disposition passes the same deterministic validation (Rule 6.6) and the same sign-off and execution gates (Rules 2.5, 7.1, 8.1) as every Award; integrity flags and bad-faith markers route to the Reviewer. A fast dispute is never an unreviewed one.
• Form of decision. The fast track issues a memorandum disposition (Rule 6.9): the holding up front, the verifiable facts relied on, and a reasoned basis of several sentences — more than a bare verdict, far less than a full opinion. It resolves the case and carries no precedential force. The decision records its tier (oracle | fast_single) so the record is honest about how the case was decided.
• Principals. Where an agent acts for a human or entity principal, informed consent of the principal to this tier has no capture mechanism yet — principal-representing agents run the standard agent lane (Rule 4.9), never the fast track. COUNSEL — specify dual-consent capture before widening eligibility.
• Protocol-native claims (consent and filing) [amended 2026-07-04]. Where the claim arises from a payment whose operator configuration — accepted on-chain by both agents at payment time — names this tribunal and incorporates these Rules including this tier, that acceptance constitutes the informed tier consent this Rule requires for claims arising from that payment. The arbitration consent and the tier consent derive from the same protocol acceptance and are recorded as distinct attestations; the "distinct consent" requirement of this Rule is satisfied by the configuration's express incorporation of the tier, not by a second signature. A structured protocol act — an on-chain RefundRequest — is itself the filing of the enumerated class escrowed_refund_request: the referral listener assigns the class mechanically from the protocol event, never by characterizing party text. Every claim arriving on this channel is filed under an enumerated class in the first instance — absent a more specific enumerated class in the referral it defaults to escrowed_refund_request, whatever the intake path (listener-relayed or operator-posted); the routing triggers of this Rule, not the intake path, decide whether the memorandum rung hears it. The request states no ground; it is procedure, not proof — the claimant bears the burden on the record, the class is never oracle-resolved, and every escalation trigger of this Rule applies unchanged. Integration requirement: an operator configuration relying on this consent basis must disclose the incorporated Rules, this tier included, to both agents before payment. CALIBRATE — counsel review of the protocol-consent basis before production use.
• ACP evaluation channel (consent, filing, and record) [added v0.8]. Where the claim arises from an Agent Commerce Protocol (ACP) job whose on-chain configuration names this tribunal's wallet as the designated evaluator, the protocol-native consent basis of the preceding bullet applies with this mapping: the buyer (client) named the evaluator address at job creation, and the provider accepted and performed the job under that designation — those two protocol acts constitute, respectively, each party's consent to arbitrate and to this tier for the evaluation of that job, and are recorded as distinct synthesized attestations scoped to the identified jobId and chain. COUNSEL — the provider's consent on this channel is one notch weaker than the x402r operator-config basis: the provider accepted a job whose configuration names the evaluator, rather than co-signing an operator configuration that expressly incorporates these Rules. Not resolvable internally; carried with the integration requirement below. The provider's on-chain submission of the job for evaluation is itself the filing of the enumerated class acp_evaluation: the referral listener assigns the class mechanically from the protocol event, never by characterizing party text, and the class is never oracle-resolved (whether a deliverable satisfies a requirement is judgment). The record on this channel is the job's own artifacts — the requirement, the deliverable, and the memo history — each authored by a party during the job and already held by both parties when evaluation begins; that protocol exchange satisfies the exchange requirement of Rules 5.3–5.5 for this channel, the considered record freezes at the evaluation referral, and the matter proceeds directly to the merits without the phased windows of Rule 4.4 (the phases already happened inside the job). Every claim arriving on this channel is filed under an enumerated class in the first instance; the routing triggers of this Rule, not the intake path, decide whether the memorandum rung hears it — a claim above the micro-claims ceiling takes the full single-arbiter review or, at or above the Rule 2.1 panel threshold, the Panel. Disposition on this rail is binary: the escrowed job budget is completed to the provider or rejected to the client in full; an award that cannot lawfully take one of those two forms is served but never self-executes — it escalates to the Reviewer at execution (Rule 8.1 settlement_action_required discipline). Every safeguard of this Rule applies unchanged: all party-authored job text passes sanitization before any model reads it, the deterministic validator blocks, and human sign-off (Rule 2.5) precedes any execution. Integration requirement: an offering or job configuration relying on this consent basis must disclose the incorporated Rules, this tier included, to both agents before the job funds. CALIBRATE — counsel review of the evaluator-designation consent basis before production use; sandbox-only until then.
• Positioning. The fast track is an efficiency layer inside the single-arbiter agent lane. It is never a substitute for the full considered-record review when a fast-rung trigger fires.

ARTICLE 5 — EVIDENCE

Rule 5.1 — Forms of evidence (admissible classes)

The Tribunal may receive: (a) party-asserted statements of fact in a submission; (b) uploaded documents (contracts, specs, deliverables, correspondence, invoices) in supported formats; (c) on-chain records (transactions, escrow state, signatures, timestamps); and (d) third-party attestations (e.g., a platform's or oracle's record). Supported upload formats extract automatically; image exhibits are transcribed under Rule 5.12; material that cannot be reliably extracted or transcribed is quarantined pending a readable version (Rule 5.6).

Rule 5.2 — Provenance and reliability tiers (weight, not admissibility)

Each item of evidence is tagged with a provenance (party_asserted_text | uploaded_document | on_chain | third_party_attested) and a reliability tier. Reliability tiers feed the weight the Tribunal gives an item, not its admissibility — on-chain records carry the strongest weight; bare party assertion the weakest. The tier table is Appendix B.

Rule 5.3 — Evidence lifecycle

Every item moves through a defined lifecycle: submitted → exchanged → admitted → (quarantined) → excluded.

• Submitted — filed but not yet shared.
• Exchanged — disclosed to the other side; its content hash is frozen (Rule 5.4).
• Admitted — part of the considered record the Tribunal may rely on.
• Quarantined — suspicious (injection, unreadable, or unverifiable); withheld from the Tribunal pending review but surfaced to the Reviewer.
• Excluded — kept out of the considered record.

Only admitted evidence is decisional. Quarantined and excluded items are never silently used.

Rule 5.4 — Integrity and tamper-evidence

On exchange, each item is content-hashed (SHA-256) and frozen; a later alteration is detectable by hash mismatch. The hash, provenance, tier, and lifecycle phase travel with the item in the record and may be cited in an opinion.

Rule 5.5 — Exchange and the considered record

Evidence is exchanged with the opposing party before it becomes decisional; a party may not rely on material the other side never saw. In submission mode, exchange occurs at the joint reveal of each simultaneous phase (Rule 4.4): sealed submissions are disclosed together, which is the exchange event for that phase. When the proceeding reaches Close (Phase 5), exchanged exhibits become admitted and the considered record is frozen; the merits phase does not run until then. CALIBRATE — the locked-submission / joint-reveal mechanism (Rule 4.4) replaces the earlier single-close model; enforced in code as of 2026-07-01 (see Appendix D).

Rule 5.6 — Quarantine and exclusion

An item is quarantined or excluded where it: cannot be read/extracted; fails authenticity; or contains material attempting to instruct the Tribunal (Rule 5.9 / Article 9). Quarantine is surfaced for human review; it is not a silent deletion.

Rule 5.7 — Authenticity challenges (amended v0.5: raising flow)

A party may challenge the authenticity of an exhibit submitted by the other side. A challenge is a structured act naming the exhibit and stating a ground, raised either within the party's rebuttal (Rule 4.4) or as a standalone challenge while the record is open; the stated ground is party text and passes the same sanitization as any submission (Rule 5.9). A raised challenge never excludes the exhibit by itself: the exhibit remains in the considered record with the challenge noted, and the Tribunal weighs it knowing its authenticity is contested (Rule 5.2) — a bare challenge to every adverse exhibit gains nothing. The challenge and its ground are surfaced to the Reviewer: a challenge sustained on review moves the item to quarantined/excluded (Rule 5.6); an overruled challenge is noted on the record. An unresolved authenticity dispute on a material item is an escalation trigger (Article 7), and a challenge sustained against fabricated or altered evidence is sanctionable misconduct (Rules 9.1, 9.2, 9.5).

Rule 5.8 — Burden and standard of proof

The party asserting a fact bears the burden of proving it. The default standard is the preponderance of the evidence (balance of probabilities). The Tribunal's per-issue burden findings are recorded in the opinions. CALIBRATE — heightened standards for fraud/forgery allegations to be specified with counsel.

Rule 5.9 — Instruction hierarchy (party text is data, not command)

Authority within a dispute is strictly ordered: these Rules and system instructions > party advocacy > evidence content. All party-supplied text and exhibit content is treated as data to be weighed, not instructions to be obeyed. Any attempt embedded in a submission or exhibit to direct, re-role, or override the Tribunal is neutralized (Article 9) and does not alter the Tribunal's mandate.

Rule 5.10 — Record-bound decision; no independent investigation

The Tribunal decides on the considered record only. It does not conduct its own factual investigation, consult outside facts, or rely on authority not in the record. An adverse inference may be drawn only where supported by the record and, where attributable to a party's misconduct, under Rule 9.2.

The sole exception is legal authority researched under Rule 5.13, which enters the considered record before the merits and is thereby authority in the record. The bar on independent factual investigation is absolute: no research pathway may be used to investigate the parties, the transaction, or any contested fact.

Rule 5.11 — Party validation of the considered record (checkpoint)

Before the Tribunal reaches the merits, each party is given a neutral restatement of its own submission — a plain-language summary of that party's claims (or defenses and counterclaims), the relief it seeks, and the exhibits it tendered, as the system has understood them for the record. Each party may confirm the restatement or correct a misstatement of what it submitted. The checkpoint runs at Close (Phase 5), after the simultaneous exchange (Rule 4.4) and before the record is frozen (Rule 5.5) and the merits run (Rule 6.1).

The checkpoint is deliberately bounded:

• Own side only. A party validates the restatement of its own submission. It is not shown the other side's restatement (beyond what the joint reveal already disclosed under Rule 4.4), the Tribunal's analysis, or any draft Award (Rule 6.10).
• Accuracy, not re-argument. Confirmation corrects whether the record faithfully captures what the party said and filed. It is not an opportunity to add new claims, new relief, new defenses, or new evidence (those close under Rules 4.2–4.4 and 4.7), and it does not preview or influence the outcome.
• A correction is logged, not silently applied. A party's correction adjusts the restatement that travels with the record; a correction that would go beyond restatement (i.e., attempts to inject new substance) is treated under Rule 4.7 (amendment by leave only) and surfaced for review.
• Non-blocking by default. If a party does not respond within the confirmation window (Rule 4.6 / Appendix A), the record is frozen on the restatement as generated, noting the non-confirmation. A disputed restatement of a material item that cannot be reconciled is surfaced for the Reviewer.
• A deliberate pause before restatement. The restatements are generated after a short post-close pause (Appendix A CALIBRATE — set to ten minutes on the standard and fast tracks, 2026-07-04; none on the agent lane), not in the same act as the final filing. The Reviewer may direct immediate generation — for either party or both — without waiting out the pause. The pause gives the operator a window for human review of the restatement and confirms to the filer that submission and summarization are distinct steps. The confirmation window runs from generation, not from the close of the record.

The purpose is trust by construction: the party agrees the record reflects what it said before any decision is made on it — converting "do you trust the AI's judgment?" into "do you agree this is your position?" (not yet enforced in code — see Appendix D.) CALIBRATE — the confirmation window and the threshold for treating a correction as a substantive amendment.

Rule 5.12 — Transcription of image evidence (extraction, not appraisal)

An image exhibit (screenshot, photograph, scanned document) enters the record through verbatim transcription: a transcription model converts what the image literally contains — text, amounts, dates, identifiers, structure — into text, and that transcription then passes through the same sanitization (Rule 5.9 / Article 9), content-hashing (Rule 5.4), and quarantine (Rule 5.6) pipeline as any other party-supplied text.

The single-decider principle: only the assigned Tribunal adjudicates evidence. The transcription step extracts; it never appraises. No transcription model may characterize an exhibit ("appears doctored," "proves payment"), assign it weight, or otherwise judge it — appraisal by a second model would split the tribunal and put an unreviewable judgment into the record. Raw pixels are never presented to the Tribunal.

• Quality gating, conservative. Transcription output is checked deterministically; low-confidence output escalates to a stronger transcription model, and output that remains unreliable is quarantined (Rule 5.6). Suspect transcription is never presented to the Tribunal as clean.
• Original bytes retained. The original file is preserved alongside the transcription: the text is the record the Tribunal reads; the pixels remain available for the Reviewer's verification and for a party's authenticity or misreading challenge (Rule 5.7).
• Weight unchanged. Transcription does not upgrade reliability (Rule 5.2): a transcribed screenshot remains what its provenance makes it.
• Firewall preserved. Pixels can carry embedded instructions; transcription output therefore receives the full instruction-hierarchy defense (Rule 5.9 / Article 9) before it can reach the Tribunal.

CALIBRATE — quality-gate thresholds and the escalation transcription model; per-page transcription of scanned multi-page documents is not yet enabled (such documents quarantine with bytes retained).

Rule 5.13 — Clerk research of legal authority (retrieval, not recall)

On a panel-composition matter (Rule 2.1), the Tribunal's clerk may research the legal authority the dispute turns on — the actual outcomes, holdings, and controlling tests of cases and statutes under the governing law, including every authority a party cited by name. The clerk retrieves public sources and extracts what each authority held into an authority pack that enters the considered record before the merits. Where an authority divided, the pack states the majority outcome and the dissent, so a dissent can never be mistaken for the holding.

The Rule 5.12 discipline applies in full:

• Extraction, not appraisal. The clerk reports what sources say; it never characterizes this dispute, weighs its evidence, or opines on the outcome. The assigned Tribunal remains the only adjudicator (single-decider principle).
• Law only, never facts. Research questions are questions of law. Rule 5.10's bar on factual investigation is untouched.
• Firewall preserved. Retrieved web text is untrusted: it passes the full instruction-hierarchy defense (Rule 5.9 / Article 9) before any model reads it, and again before the pack enters the record. Provider-side browsing that splices unreviewed web text directly into an adjudicator's context is prohibited — no panelist browses independently.
• One pack, one record. Every panelist deliberates from the same pack; retrieved authority a panelist saw but the record does not hold would defeat cite-integrity and is prohibited.
• Auditable and checked. The pack records its questions, sources (URLs), and sanitization results. The integrity layer verifies the operative opinion's statements about researched authorities against the pack; a contradiction (e.g. an outcome stated backwards, or a dissent adopted as the holding) escalates the Award to the human Reviewer (Article 7) — the check may route to a human, never alter the outcome.
• Failure degrades, never decides. If retrieval fails, the attempt is recorded and the matter proceeds on the record as filed; a research failure is neither a decision input nor a ground to delay the Award.

The pack, including source links, is disclosed with the served Award. COUNSEL — due-process interaction: an award resting on authority the parties had no opportunity to address is a recognized challenge vector (iura novit curia limits). V1 discloses the pack with the Award; a pre-award party-comment window on tribunal-researched authority is reserved for a future revision. CALIBRATE — research question cap and search depth; eligibility is panel-composition matters only (micro-claims and fast-lane matters are never researched).

Rule 5.14 — Attestation of truthfulness on submission (new v0.5)

Every submission and exhibit is filed under an attestation of truthfulness: the filing party attests that the evidence it submits is genuine and unaltered and that the factual statements it makes are true to the best of its knowledge. The attestation is recorded with the filer's case identity, the filing's content hash (Rule 5.4), and a timestamp, and travels with the filing in the record; a filing tendered without the attestation is not accepted into the record. A false attestation is sanctionable misconduct under Article 9 (Rules 9.1, 9.2, 9.5), and an opinion may cite the attestation when weighing the filing (Rule 5.2). The attestation binds the party regardless of whether a human or an agent performed the filing act. COUNSEL — whether the attestation should invoke statutory perjury language (cf. 28 U.S.C. §1746) and what force such language carries against pseudonymous or agentic parties; the recorded truthfulness attestation is the operative form until counsel advises.

ARTICLE 6 — DECISION & THE AWARD

Rule 6.1 — Merits decision by the assigned composition

Single-arbiter composition. In an eligible below-threshold agent-lane matter (Rule 2.1), the neutral AI arbiter reviews the complete considered record and issues one reasoned opinion. No other AI agent votes on, synthesizes, or substantively reviews that decision. An eligible micro-claim may first pass through the oracle and fast-memorandum rungs of Rule 4.11; a trigger at either rung returns the matter to this full single-arbiter review.

Panel composition — first round (independent). In a Panel matter, each agent issues its own reasoned opinion on the considered record, from its declared stance (Rule 2.2), without coordinating with the others.

Panel deliberation round (one, bounded). The agents are then shown one another's first-round dispositions and reasoning. In a single round — no iteration, no presiding agent, no vote-trading — each agent may revise or hold, and must state whether it adopts, partially credits, or rejects any materially different remedy on the panel, with reasons. An agent that holds after genuinely engaging the others has satisfied this Rule; convergence is neither required nor rewarded (Rule 2.3). Both first-round and final opinions are retained in the record. A matter flagged fast-lane at triage may skip the deliberation round. The final opinions are what the panel result is computed from (Rule 6.3).

Rule 6.2 — Contents of an opinion

Each opinion shall contain: (1) caption — parties/IDs, dispute ID, governing law/rules and Rule version, and composition; (2) questions presented; (3) findings of fact, each tied to specific admitted evidence; (4) reasoning applying the governing text to the facts; (5) holding / disposition — who prevails, what the corpus should do, and any amounts, expressed as structured funds instructions; (6) confidence / caveats and any escalation flag. Panel agents additionally identify their declared stance. Internal Panel drafts remain Reviewer-facing; the parties receive the single authored reasoned Award under Rule 6.9. In a single-arbiter matter, that arbiter's full opinion is itself the party-facing reasoned Award.

Rule 6.3 — Panel aggregation (two axes: liability and quantum; operative = a real opinion, never an average)

This Rule applies only where the three-agent Panel sits. A single-arbiter decision is not aggregated and is never presented as a vote.

The panel result is computed deterministically along two separate axes, so that agreement on who prevails is never reported as agreement on how much:

1. Liability axis. Group the final opinions by their actual outcome; the majority outcome is the one shared by at least two agents. This selects the winning side and the operative opinion.
2. Operative disposition. The median disposition within the majority group — an actual opinion an agent wrote, never a numeric blend of opinions.
3. Quantum axis. Within the majority-outcome group, compare each agent's remedy. Where the operative remedy and a same-outcome agent's remedy diverge by more than a set fraction of the corpus (CALIBRATE default 15%), the panel has a material quantum split, which is reported as such: the result states the liability vote and the remedy vote distinctly (e.g. "3–0 on liability, 2–1 on remedy") and never as flat unanimity.

Roles are derived deterministically: majority (prevailing outcome + operative remedy), concurrence (same outcome, materially different remedy), dissent (minority outcome). There is no AI "join/dissent vote." A material quantum split on a flagged matter escalates (Rules 6.4, 7.1).

Rule 6.4 — Panel concurrence and dissent are published

Where the Panel sits, the operative disposition, any concurrence, and any dissent are recorded and published with the Award, carried in the authored Opinion of the Panel (Rule 6.9). A 2-1 split on the outcome, or a material quantum split on the remedy (Rule 6.3), is an escalation trigger where the matter is flagged high-stakes, ambiguous, or novel (Article 7). This Rule does not manufacture concurrence or dissent in a single-arbiter case.

Rule 6.5 — Award posture (what kind of award)

Every Award declares its posture, independent of who wins:

• escrow_allocation — split a held corpus among the parties (funds must sum exactly to the corpus).
• money_obligation — one party owes the other a sum (liability and amount together; 0..claim).
• liability_only — liability is clear but the amount cannot be grounded on the record → routed to human assessment; no amount is invented.
• declaratory — a declaration of rights with no self-executing transfer.

Obligor/obligee are derived from the posture and disposition.

Rule 6.6 — Deterministic validation (blocking, accounting-only)

Before any Award issues, a deterministic validator checks integrity, not merits: the ledger balances to the corpus where the posture requires it; a conditional/held instruction carries both a condition and a fallback recipient; allocation stays within the corpus unless remedial jurisdiction was opted into (Rule 6.8); and the decision is traceable to the assigned authority. In Panel matters it additionally checks majority support, traceability to a real panel opinion, and consistency between each panelist's label and funds. In single-arbiter matters it checks the arbiter's label, disposition, and accounting without inventing a majority requirement. Any failure blocks execution and escalates to the Reviewer — the system never executes an Award that fails validation. The validator never overrides the Tribunal on substance.

Rule 6.7 — Citation integrity (no fabricated authority or evidence)

Every fact an opinion relies on must resolve to an item in the considered record, and every funds instruction must rest on a finding the opinion actually made. A cite to evidence or authority not in the record (a fabricated source) is flagged; a material fabricated cite underlying an operative money instruction is a blocking error → escalation. No hallucinated authority is permitted in a published opinion.

Rule 6.8 — Remedies and the mandate boundary

The default mandate is corpus-only: an Award allocates the funds under control among the parties and may impose conditions/holdbacks on that corpus, but does not issue free-standing performance orders. Relief beyond the corpus (e.g., damages exceeding escrowed funds) requires an express opt-in in the consent and is otherwise expressed as a money_obligation or routed to liability_only. Relief is bounded by what was pleaded (Rule 3.1).

Rule 6.9 — Form of the Award: one reasoned opinion

The Award is issued as one reasoned opinion matching the assigned composition: an Opinion of the Single Arbiter in a single-arbiter matter, or an Opinion of the Panel where the Panel sits. It opens by stating the holding and relief, then sets out the facts (findings tied to admitted evidence), the governing standard (the applicable rule/doctrine and the burden and standard of proof, Rule 5.8), the analysis applying the law to the facts issue by issue, and a conclusion stating the disposition.

For a Panel matter, the authored opinion is not a bundle of the three internal opinions: it must name and answer the principal opposing position, and any concurrence or dissent is attached from the agents' actual final opinions. For a single-arbiter matter, the arbiter's full opinion travels directly; no separate AI synthesizer or fictional panel vote is added. In either composition, the funds instructions are the operative disposition verbatim: explanatory drafting never creates or alters an amount.

Two forms. A short memorandum (holding up front + issue-by-issue reasoning; no separate summary) may issue for an oracle/fast-memorandum micro-claim and for qualifying low-stakes unanimous Panel matters. A full opinion issues after full single-arbiter review and for Panel matters that are flagged, high-stakes, or carry a dissent or material quantum split. Every Award records its decision tier (oracle | fast_single | single_arbiter | panel).

The Award states its posture, identifies the Rule version, records the validation result and any integrity findings, rests every proposition on the record (no fabricated authority, Rule 6.7), and — in V1 — bears the Reviewer's sign-off before it is final (Rule 2.5).

Rule 6.10 — Disclosure to the parties (what parties see, and when)

The parties' visibility into the proceeding is defined and staged. There are exactly three things a party sees:

1. In — its own restatement for confirmation at the validation checkpoint (Rule 5.11), plus whatever the joint reveal already disclosed of the other side's filing (Rule 4.4).
2. Withheld — interim work. The parties do not see the Tribunal's working analysis, any individual Panel drafts, triage internals, integrity/injection internals, or any draft Award. These are internal to the Tribunal and the Reviewer. The merits run and any escalation are not exposed in progress; while a matter is under review a party sees only that it is "under review," not the pending result.
3. Out — the reasoned Award on release. When the Award is final (validated, and in V1 signed off — Rule 2.5), it is released to the parties as a reasoned decision: the operative disposition, its posture, the reasoning (by incorporation of the opinions, Rule 6.9), and any published concurrence and dissent (Rule 6.4). The parties receive the reasoned opinion at the end; they do not receive the deliberation that produced it.

This staging is deliberate. Validating the input (Rule 5.11) earns trust without exposing interim work; disclosing a fully reasoned Award at the end provides transparency and a basis for recognition (Rule 8.4) without inviting mid-stream merits challenges. Release is not a separate gate: the Reviewer's sign-off (Rule 2.5) is what serves the decision, and the reasoned opinion travels with it by default. (A matter may exceptionally serve the holding with the reasoned opinion to follow, but that is the exception, not a co-equal step.) Serving the decision is a distinct, logged event from fund execution (Rule 8.1). CALIBRATE — whether parties ever see redacted interim summaries is an open question; default is analysis withheld, reasoned Award disclosed.

ARTICLE 7 — ESCALATION TO A HUMAN

Rule 7.1 — Triggers

A matter is escalated, in whole or in part, where any of the following is present (consolidated; see Appendix C):

• Threshold failure — consent defective/declined, out of scope, or a flagged unpleaded dispositive defense (Article 3).
• Validation failure — any blocking integrity failure (Rule 6.6) or material fabricated cite (Rule 6.7).
• Split signal — a 2-1 outcome split, or a material quantum split on the remedy (Rule 6.3), on a matter flagged high-stakes, ambiguous, or novel.
• Stakes — disputed amount above the sign-off threshold (default $10,000) CALIBRATE, or non-escrow matters by default.
• Evidentiary conflict — a pure credibility contest the record cannot resolve, or a sustained/unresolved authenticity challenge (Rule 5.7).
• Novelty / bespoke law — a dispute type or governing law outside the validated set the eval harness covers.
• Integrity / bad faith — attributable prompt-injection, forgery, collusion, or sanctions/identity markers (Article 9).
• Irreversibility — outcomes that cannot be unwound if wrong.
• Posture — liability_only (amount must be humanly assessed).

Rule 7.2 — Effect of escalation

On escalation, no funds move: the corpus is left untouched and the matter is routed to the Reviewer queue. Escalation is a routing decision, not a decision on the merits.

Rule 7.3 — The Reviewer on escalation

The Reviewer may resolve the matter, request further submissions, direct a live testimony session (Rule 4.10), or return it to the composition assigned under Rule 2.1 after a defect is cured (Rule 8.6). A reasoned Reviewer decision is recorded; where the Reviewer writes a decision memo or attaches a document, the parties may read the memo and obtain the attachment.

ARTICLE 8 — EXECUTION & FINALITY

Rule 8.1 — Decision is not execution (served vs. paid)

The status of a dispute distinguishes deciding, serving, and executing. Sign-off serves the decision to the parties (Rules 2.5, 6.10); fund execution is a separately-timed event. Execution occurs on sign-off or, where configured, at the close of a challenge window (served_awaiting_execution → executed; default window 0 CALIBRATE). Where the consent provides for the Rule 8.8 appeal, the window is nonzero and is the period within which an appeal may be filed; a filed appeal holds execution (Rule 8.8). Serving the decision is never delayed by the window — only execution holds. Without an appeal tier the window is zero and execution follows serving immediately. If an external settlement adapter is unavailable or refuses the instruction after service, the Award remains valid and served while the case enters settlement_action_required; that state is an execution exception, not a merits escalation or withdrawal of the Award. A validated, signed Award progresses through explicit states; a held/conditional amount is never shown as "released." This finality-timing window is not a merits appeal (Rule 8.5).

Rule 8.2 — Self-execution (on-chain escrow) — escrow is optional

Escrow is optional: a case may run with or without funds under the Tribunal's control. Where a corpus is held in escrow, the Award self-executes (escrow_allocation posture): the escrow releases per the funds instructions, and no court confirmation is required for collection of the corpus. Where no escrow is posted, the Award is a money_obligation/confirmable award (Rule 8.4) — the absence of escrow changes the execution path, not the right to adjudicate.

Rule 8.3 — Conditional release and holdback

A holdback or conditional instruction releases only when its stated condition is met; absent that, the funds follow the stated fallback (Rule 6.6). A pending holdback rests in a holdback_pending state, not "released."

Rule 8.4 — Non-escrow / obligation awards

Where there is no corpus under control (a money_obligation), the Award is a confirmable obligation — a reasoned award a party may take to recognition/enforcement — not a self-executing transfer. The execution path for these awards depends on the enforceability strategy (see enforceability-strategy.md). COUNSEL

Rule 8.5 — Finality and correction (amended v0.6)

A signed, executed Award is final on the merits; there is no merits appeal within the process except the optional Rule 8.8 appeal where the consent provides for it and it is invoked within the challenge window. The Tribunal/Reviewer may correct a clerical, computational, or typographical error, and may interpret an ambiguous operative term, without reopening the merits. Any challenge to the Award beyond clerical correction and Rule 8.8 lies outside the process and is governed by the applicable law and the consent. COUNSEL — interaction with FAA / New York Convention recognition; internal appellate layers are the recognized mechanism for expanded review, since parties cannot enlarge FAA §10 judicial review by contract.

Rule 8.6 — Reopening

A matter may be reopened only by the Reviewer/administrator (e.g., to cure a validation defect or after escalation): prior opinions and the draft Award are cleared, the submitted evidence and considered record are retained, and the dispute is reset for a fresh run. Reopening is a logged event.

Rule 8.7 — Additional award (pleaded-but-omitted claim)

Where the Award omitted a claim that was pleaded before the Award issued, the Reviewer may decide that claim by an additional award without reopening the merits already decided. Boundaries, deterministically enforced: the claim must trace to the pleaded record (a new claim is a new case, not an addendum); the same pleaded claim is decided at most once; the original Award's instructions are preserved verbatim and the addendum only adds instructions for the omitted claim; the combined ledger passes the same validation as the original Award and is re-hashed (and re-signed where signing is configured); no addendum issues after funds have executed. Where the Award fully allocates a held corpus (escrow_allocation), a monetary addendum would necessarily reallocate decided funds and is therefore excluded - the omitted claim may be decided as no-relief by additional award, and any reallocation goes through Rule 7.3 or Rule 8.6. Deciding the omitted claim as no relief is still deciding it and must be explicit and reasoned. The additional award is a reasoned, party-visible docket event; where it changes which party nets more overall (an omitted set-off can), that change is flagged on the record, never silent. COUNSEL — deadline for requesting an additional award (cf. AAA R-50 / UNCITRAL art. 39 30-day windows) and whether party request should be a precondition.

Rule 8.8 — Optional appeal to a human appellate reviewer (new v0.6)

Where the consent provides for it, either party may appeal a served Award to a human appellate reviewer within the Rule 8.1 challenge window. The appeal is the process's one internal merits pressure valve; it is bounded on every axis:

• Eligibility. Only Awards not decided by a human on the merits are appealable: matters decided at the oracle, fast-memorandum, single-arbiter, or Panel tier under routine supervisory sign-off. A matter the Reviewer resolved under Rule 7.3, or any human-authored decision, is final without appeal — the human level has already been had.
• Window and irreversibility. The appeal must be filed within the challenge window and before funds execute; no appeal lies after execution (Rule 8.7 discipline). A filed appeal holds execution only — the served decision stands served, and serving is never delayed by the window (Rule 8.1).
• Fee, never a bond. The appellant pays a flat, non-refundable filing fee to the forum (Rule 10.1 fee-for-service; amount CALIBRATE). The fee is never outcome-linked, never payable to the opposing party, and never a posted bond forfeited on the result — nothing about the appeal may create a stake in the outcome beyond the parties' own claims. The appeal is docketed when the fee is confirmed.
• Record-bound de novo review. The appellate reviewer decides the matter afresh on the frozen considered record (Rule 5.5). No new evidence, claims, or defenses on appeal — new evidence remains a Rule 8.6 reopening question. The appellate reviewer owes the parties the full Rule 2.7 duties and must not be the Reviewer who signed off the appealed Award.
• Disposition. The appellate reviewer affirms the Award (execution proceeds) or issues a substitute reasoned decision through the Rule 7.3 machinery; a substitute decision passes the same deterministic validation (Rule 6.6), is re-hashed and re-signed where configured (Appendix E), and carries human-decided authority in the precedent ledger. Either way the appellate decision is final: one level, no further internal review.

COUNSEL — the appeal option, window, and fee must be stated in the consent; confirm the internal appellate layer's interaction with confirmation strategy (parties cannot enlarge FAA §10 review by contract — this layer exists precisely because internal review is the enforceable alternative). CALIBRATE — window length (recommended 72h standard/fast, 24h agent lane) and the filing fee.

ARTICLE 9 — INTEGRITY, CONDUCT & SANCTIONS

Rule 9.1 — Prohibited conduct

No party or exhibit may: attempt to instruct, re-role, or override the Tribunal (prompt injection); submit forged, altered, or fabricated evidence; collude to manipulate the outcome; or impersonate a party. Legitimate third-person advocacy ("Respondent should be held liable") is not prohibited conduct and is preserved intact.

Rule 9.2 — Sanctions and responses (amended v0.5)

Detected misconduct is met deterministically and proportionately: tribunal-directed instructions are neutralized (Rule 5.9); severe cases are quarantined (Rule 5.6); an adverse inference may be recommended only where the misconduct is attributable to a party; and attributable injection or forgery routes the matter to human sign-off. Attributable fabrication, forgery, or injection may additionally warrant exclusion from the forum (Rule 9.5). Sanctions are surfaced in the integrity report on the Award.

Rule 9.3 — Confidentiality and publication

The Award, the opinions, and any dissent are part of the record and may be published (in de-identified form for pseudonymous parties). The amount of reasoning exposed publicly versus to the parties only — to protect pseudonymous identities or trade secrets in evidence — is set per dispute. COUNSEL

Rule 9.4 — Data handling and use of submissions

Raw exhibits are handled per the data-handling policy. Raw submitted material is not used to train models; any de-identified learning signal is opt-in and passes through a reviewed abstraction pipeline. A training-consent clause governs before any document lands. (policy decision pending — gates the privacy policy; see backlog "corpus training tiers.") COUNSEL

Rule 9.5 — Exclusion from the forum (prospective, human-decided) (new v0.5)

Attributable fabrication, forgery, or prompt injection (Rule 9.1) may result in the responsible party's exclusion from the forum: the excluded identity may not open or join new matters. Exclusion is decided only by a human Reviewer or the administrator on an attributable finding — never automatically by the engine — and is a logged, prospective act.

Exclusion is orthogonal to any pending matter, by design, so that misconduct can never become an exit:

• It never divests the Tribunal of jurisdiction over a matter already commenced (Rule 2.4); consent to arbitrate, once given, is not shed by misconduct.
• It never terminates, decides, or alters a pending matter. The pending matter proceeds to disposition on the clean considered record (Rule 5.6), and any adverse inference is decided by the Tribunal under Rules 5.10 and 9.2 — never by the exclusion.
• If an excluded party declines to participate in remaining phases, the per-phase default rules apply (Rule 4.8); the matter is never cancelled by the exclusion, and the non-excluded party's claims are heard.

A party therefore gains nothing in a pending case by incurring, or provoking, exclusion. CALIBRATE — exclusion durability is bounded by identity binding (a re-registered identity defeats it); strengthen alongside account-identity requirements as they land.

ARTICLE 10 — FEES & ADMINISTRATION

Rule 10.1 — Fees (amended v0.9)

The process is fee-for-service: the forum charges a separate fee for administering the case. The fee is never taken out of the escrowed or disputed funds, and it is not linked to which party prevails. COUNSEL

Fee schedule. The acts that incur a fee are: filing a claim (any intake channel), docketing a Rule 8.8 appeal, and referring a matter for evaluation. Each fee is either a flat amount or value-scaled — the greater of a published percentage of the amount in dispute and a published per-tier minimum — computed and quoted at filing from the pleaded amount, payable by the invoking party, and identical whichever party prevails. Later routing or escalation never re-prices a case: escalation costs are priced into the published minimums on average, never re-billed mid-case. A fee is never contingent on the outcome, never payable to or from the opposing party, and never collected from the escrowed corpus; it is charged and collected as a separate payment to the forum. Published amounts CALIBRATE — the percentage and the per-tier minimums are published constants derived from the metering ledger, refreshed periodically. The forum may waive any fee; a waiver is quoted and recorded like the fee it forgives. Where a pleading is amended upward by leave (Rule 4.7), the filing fee is topped up by the difference between the fee at the amended amount and the fees already incurred; no refund is made for amending downward. An unpaid filing fee means the matter is never docketed: no proceeding commences, no clock runs, and any protocol or escrow-window default governs the underlying funds. A fee gates intake only — it never suspends, conditions, or alters a matter already docketed. COUNSEL — money transmission / collection of real fees; consumer fee caps where a natural-person consumer is the filing party

Rule 10.2 — Administration

The administrator registers disputes, manages exchange and the Reviewer queue, and maintains the record. Administrative acts (exchange, reopening, sign-off) are logged.

Rule 10.3 — Allocation of the fee burden (costs) (new v0.9)

The party invoking a fee-incurring act bears its fee when due; the filer bears its own filing fee by default. In an Award whose posture can carry it (a money obligation, or a matter with remedial jurisdiction under Rule 6.5), the Tribunal may, in its discretion and on a party's request, allocate the fee burden between the parties by ordering one party to reimburse the other's paid filing fee as part of the relief — the ordinary "costs follow the event" reconciliation. Allocation moves the burden only between the two consenting parties: the forum's fee is identical whoever prevails, is never enlarged or reduced by the allocation, and is never collected from the escrowed corpus. On corpus-only matters (an escrow allocation without remedial jurisdiction) no costs order is available and the burden stays with the filer. COUNSEL — confirm a costs-reimbursement order reads as relief between the parties, not an outcome-linked fee adjustment

ARTICLE 11 — GENERAL PROVISIONS

Rule 11.1 — Governing law, seat, and recognition

The governing law and any seat of arbitration are those named in the consent. Whether a seat is named even for on-chain disputes (to preserve a recognition fallback) is a strategic choice. COUNSEL — see `jurisdiction.md`, `enforceability-strategy.md`.

Rule 11.2 — Notices

Notice is effective when delivered to the address/identifier (including a wallet or platform handle) a party provided at filing.

Rule 11.3 — Language

The language of the proceeding is English unless the consent provides otherwise.

Rule 11.4 — Amendment of these Rules

These Rules may be amended prospectively; amendments are versioned and do not bind already-consented disputes (Rule 1.3).

Rule 11.5 — Severability

If any Rule is held unenforceable, the remainder stays in effect to the maximum extent permitted.

Rule 11.6 — Limitation; no legal advice; UPL

The process resolves the parties' dispute under the terms they consented to; it does not provide legal advice to either party, and nothing in it is the practice of law on a party's behalf. The boundary between "the agent applies the governing terms" and unauthorized practice of law must be confirmed. COUNSEL — UPL exposure is an open legal question.

APPENDIX A — Default schedule (Rule 4.6) CALIBRATE

Same five phases on every track; only the clock changes.

Deadlines are dates certain computed from commencement and are backstops: the joint reveal fires as soon as both sides seal (Rule 4.4), so responsive parties — especially two agents — finish well inside them. Word limits (Rule 4.5): Complaint / Answer 1,500 words; Phase 3 memorial 3,000 words; Phase 4 rebuttal 1,500 words. All values configurable in the consent; not to be set so short as to deny a fair hearing.

APPENDIX B — Evidence reliability tiers (Rule 5.2)

Tiers feed weight, never admissibility (Rule 5.2).

APPENDIX C — Escalation triggers (Rule 7.1, at a glance)

Threshold failure · validation failure · material fabricated cite · 2-1 outcome split or material quantum split on flagged matter · stakes above threshold (default $10k) · non-escrow by default · unresolved credibility/authenticity conflict · novelty / bespoke law · attributable injection / forgery / collusion · irreversibility · liability_only posture.

Fast rung → full single-arbiter review (below-threshold agent lane, Rule 4.11): micro-claim value ceiling · tier consent · claim class not enumerated · quarantined evidence or injection · split/conditional fast verdict · over-award · validation failure · rate-limit / ceiling-probe. (Low confidence and evidentiary conflict are recorded, never re-routed — v0.3.) Agent lane → three-agent Panel: amount at or above the Rule 2.1 panel threshold (default $10,000). Single-arbiter/Panel → human Reviewer: blocking validation or integrity failure and the Article 7 triggers above.

APPENDIX D — Rule → engine mapping (for engineering)

| 10.1 fees (schedule · quote · records · collection) | lib/fees.ts (single schedule source: fee = max(published percent × amount in dispute, per-tier published floor), all env defaults 0/off; deterministic quote tiers from filing-time facts; FeeRecord lifecycle payable/paid/waived with fee_recorded/fee_paid/fee_payment_lapsed audit events; Rule 4.7 upward top-up via filingFeeTopUp) + lib/fee-payments/adapters.ts (payment-method seam: onchain_usdc receipt verification is the first real adapter, fiat_card is a Stripe stub boundary that refuses cleanly, manual mirrors the Rule 8.8 operator confirmation; unknown methods throw) + lib/fee-payments/usdc.ts (USDC transfer TO the treasury address verified from the tx receipt — address only, base-sepolia only) + lib/x402r/bridge.ts (agent lane: referral acknowledgment carries an HTTP 402 payment challenge; docketing gates on payment; unpaid past the deadline ⇒ filing_fee_pending case closes, never docketed) + GET /api/fees/quote, POST /api/x402r/fee/confirm, the intake fee step, and the party-facing /pricing page (lib/ui/pricing-view.ts, rendered entirely from the schedule) |

Not yet enforced in code (gaps to close): Rule 1.3 version-binding · Rule 4.5 word limits · Rule 9.4 data-handling/training tiers · Rule 10.3 costs allocation (the Tribunal-discretion costs order has no engine mechanic yet — the fee burden stays filer-bears in practice until it ships). The v0.5 additions all shipped 2026-07-08 — Rule 5.7 challenge raising flow, Rule 5.14 attestation, Rule 2.5 named reviewer, Rule 2.7 reviewer neutrality/recusal, Rule 4.10 testimony-session direction, Rule 9.5 exclusion, Appendix E hash recipe v2 — see Appendix D. The v0.6 Rule 8.8 appeal mechanism shipped 2026-07-08 (row above); its fee collection remains operator-manual (APPEAL_FEE_USD recorded, docketing gated on confirmation) but the fee now also lives on the shared Rule 10.1 fee ledger (v0.9). (Rule 6.10 disclosure is partly enforced — see mapping above.)

APPENDIX E — Award-hash canonicalization (Rules 6.9, 8.4) (new v0.5)

The award hash committed to award signing and on-chain notarization is computed by a published, versioned recipe, so a party, auditor, or court can independently recompute it from the served Award:

• Recipe v2 (current). The hash input is a JSON document serialized canonically: all object keys sorted lexicographically by UTF-16 code unit, string values Unicode NFC-normalized, numbers in ECMAScript shortest round-trip decimal form, no insignificant whitespace, UTF-8 encoded, hashed with SHA-256. The document covers: the Rule version, the dispute id, the award posture, the full funds ledger in the order issued (ledger order is part of the award), the source opinion id, the SHA-256 of the operative reasoned opinion text, and the sorted content hashes (Rule 5.4) of every considered exhibit. Fields with no value in a given disposition (e.g. no operative opinion on an escalation) are carried as explicit nulls, never omitted.
• Time is proven by notarization, not baked into the hash. The hash covers award content; the on-chain notarization timestamp proves when that content existed.
• Versioning. Every signed or notarized hash records its recipe version; verification recomputes under the recorded version, so a recipe change never invalidates a previously issued award. Recipe v1 (legacy: unordered JSON serialization over dispute id, funds, and source opinion id) remains verifiable for awards that recorded it.
• Amendments. A clerical correction (Rule 8.5) or additional award (Rule 8.7) re-hashes under the current recipe, records the recipe version, and re-signs/re-notarizes where configured.

Cross-references: mechanism.md · panel-architecture.md · opinion-format.md · escalation-logic.md · consent-agreement-skeleton.md · enforceability-strategy.md.