Formal Mechanism Specification · rulesVersion 1.3.0

Agent-Mediate Mechanism Specification (rulesVersion: "1.3.0")

Published deterministic bargaining, mandate enforcement, secret-alpha Nash proposal, and contractual final-offer determination rules.

rulesVersion 1.3.0 rulesHash 8d7e67e1145097e3…
  • Schema: agent-mediate/mechanism-spec/v1
  • Rules Version (RULES_VERSION): 1.3.0
  • Canonical Rules Hash (RULES_HASH): 8d7e67e1145097e3ce20d8828e451db20453d864782c91ccb08be2e9e613cdc4
  • Machine-Readable Endpoint: GET /v1/rules (sha256Hex(canonicalJson(spec)) === rulesHash)

Every platform-signed settlement instruction, UNRESOLVED certificate, transparency checkpoint, and co-signed agent-mediate/clause/v1 embeds rulesVersion and rulesHash (RFC 8785 JCS + SHA-256 over MECHANISM_SPEC). Any change to the deterministic rules, domains, clamping, distance metric, or window limits changes RULES_HASH and requires a RULES_VERSION bump.


1. Issue Grid, Linear Utility, and No-Retraction Monotonicity

A confirmed dispute defines 1 ≤ m ≤ 5 negotiable issues I = \i_1, …, i_m\. Each issue i has:

  • Inclusive domain [i.min, i.max] with granularity i.step > 0, where (i.max - i.min) / i.step is an exact integer.
  • Preference orientation i.claimantPrefers ∈ \"higher", "lower"\ (the respondent strictly prefers the opposite direction).
  • Designated performer i.obligor ∈ \"claimant", "respondent"\.

Grid Alignment & Even-Half Rounding

Every package x ∈ ℝ^m submitted or output by the mechanism lies on the exact step grid x_i = i.min + k · i.step (k ∈ ℤ). Exact half-step midpoints snap to the even grid index k so rounding has zero directional bias toward claimant or respondent.

Normalised Share & Additive Utility

For value x_i ∈ [i.min, i.max], the claimant's normalised share is:

s_C(i, x_i) = (x_i - i.min) / (i.max - i.min)    if  i.claimantPrefers = "higher"
  (i.max - x_i) / (i.max - i.min)    if  i.claimantPrefers = "lower"

and the respondent's share is s_R(i, x_i) = 1 - s_C(i, x_i). Given a party p's confidential Round-1 non-negative weights w_p (\sum_i w_p,i = 1), party p's normalised utility is U_p(x) = \sum_i ∈ I w_p,i s_p(i, x_i) ∈ [0, 1].

Per-Issue No-Retraction Rule

In every bargaining round r > 1 (and in final-offer determination), each party's new offer x^(r) must concede or hold steady on every issue relative to its previous offer x^(r-1):

  • Claimant: s_C(i, x^(r)_i) ≤ s_C(i, x^(r-1)_i) for all i ∈ I.
  • Respondent: s_C(i, x^(r)_i) ≥ s_C(i, x^(r-1)_i) for all i ∈ I.

Violations are rejected immediately with 422 RETRACTION.


2. Principal-Signed Mandates and Confidential Inputs

Mandates (agent-mediate/mandate/v1)

Principals authorise agents via Ed25519-signed mandates specifying:

  • bounds: inclusive per-issue authority intervals [B_p,i^min, B_p,i^max] for voluntary bargaining (OVERLAP_MIDPOINT, PROPOSAL_ACCEPTED, SETTLEMENT_IMPROVED).
  • Optional determinationConsent: true and optional determinationBounds (MECH-1, D126). determinationBounds may be tighter or wider than voluntary bounds, enabling two-tier mandates where principals pair cautious voluntary bargaining bounds with wider fallback determinationBounds so non-overlapping voluntary mandates can still resolve via FINAL_OFFER_DETERMINATION.
  • expiresAt (must outlive bargainingWallClockMs, plus determinationTimeoutMs when determinationConsent is enabled; MECH-5), monotonic issuedAt on replacement (CR2-11), and optional scope: { contractHash, counterpartyId } (H2).
  • When a co-signed pre-dispute agent-mediate/clause/v1 specifies non-empty terms.issueTemplates, proposedIssues at filing must be a non-empty subset matching each template's id, claimantPrefers, obligor, unit, and step with on-grid [min, max] inside [template.min, template.max] (422 CLAUSE_VIOLATION on mismatch; SEC-2). Mutually confirmed AI trade-off suggestions during ISSUE_CONFIRMATION remain permitted when both parties confirm with valid mandates.

Resolving Non-Overlapping Mandates (D126, D127)

When principals' initial single-issue or narrow mandates do not overlap (B_C ∩ B_R = ∅), three mechanism paths enable settlement without compromising principal control or Round-1 confidentiality:

  1. Adopting AI-Suggested Trade-Off Issues (REJECT_ISSUES.adoptSuggestedIds): During ISSUE_CONFIRMATION (when proposalCount < maxIssueProposals), either party may call REJECT_ISSUES bound to the current issuesHash (409 HASH_MISMATCH if stale, BUG-3) with optional adoptSuggestedIds (1..2 issue IDs from advisory.suggestedIssues). Omitting adoptSuggestedIds resets the next PROPOSE_ISSUES to filing.proposedIssues; passing adoptSuggestedIds preserves filing.proposedIssues, consecutively adopted prior suggestions (priorAdoptedSuggestedIds), mutually mandated suggestions, and the newly adopted IDs (total <= 5, zero extra LLM calls), allowing both parties to call CONFIRM_ISSUES with multi-issue replacement mandates and logroll across issues.
  2. In-Place Monotonic Principal Escalation (UPDATE_MANDATE): During BARGAINING or DETERMINATION, either party may submit up to MAX_MANDATE_UPDATES = 3 replacement mandates (POST /v1/disputes/{id}/mandate), surfaced in nextActions whenever an openProposal exceeds the party's current mandate (proposalWithinYourMandate === false), in DETERMINATION, or in single-issue rounds where standing offers remain disjoint. The replacement must come from the same principalId with strictly newer issuedAt, next.expiresAt >= previous.expiresAt, identical determinationConsent (determinationConsent is frozen once BARGAINING opens so neither party can unilaterally flip determination opt-in after seeing bargaining dynamics, SEC-1), and raw + grid-clamped superset bounds (B'_p ⊇ B_p, plus D'_p ⊇ D_p when determinationConsent is enabled; BUG-1) covering Round-1 declared.limits and any standing/in-round offer. When both mandates consent to determination, bargaining offers must also stay within determinationBounds on every issue (BUG-7) so no party can concede past its determination authority during bargaining. Round-1 private (weights, walkaway, limits) stays strictly frozen (M13 / MECH-8); UPDATE_MANDATE appends an opaque "internal" MANDATE_REGISTERED link in /audit and redacts counterparty /signatures rows to action: "confidential_action" with agentId: null (SEC-1).
  3. Two-Tier determinationBounds: Setting determinationConsent: true with determinationBounds wider than voluntary bounds resolves deadlocked bargaining via FINAL_OFFER_DETERMINATION without manual escalation.

Round-1 Confidential Inputs (private)

With its first offer, each party submits private: { weights, walkaway, limits? }, frozen for the remainder of the dispute:

  • walkaway (W_p) must lie on the grid and inside mandate.bounds (MECH-5).
  • Optional limits (L_p,i = [L_p,i^min, L_p,i^max]) restrict platform proposals and post-settlement Pareto improvements to a feasible sub-box containing W_p inside mandate.bounds (MECH-7).
  • Confidential inputs never enter any LLM prompt and never leave the submitting party's own PartyView.

3. Crossing-Offer Settlement (OVERLAP_MIDPOINT, MECH-6)

When Round r closes with standing offers (O_C, O_R), offers cross if and only if each party's utility from the counterparty's offer meets or exceeds its utility from its own offer under its declared weights (for single-issue disputes, s_C(O_R) ≥ s_C(O_C)).

  1. Compute the per-issue grid-snapped midpoint M_i = snap_even(1/2(O_C,i + O_R,i)).
  2. If M lies inside both parties' mandate.bounds, settle at M.
  3. Otherwise (MECH-6), clamp overflowing issues to the intersection box B_C ∩ B_R and greedily step compensating issues within B_C ∩ B_R on the step grid to minimise \max_p max(0, U_p(O_p) - U_p(x)) subject to U_C(x) ≥ U_C(O_C) and U_R(x) ≥ U_R(O_R) whenever feasible inside B_C ∩ B_R.

4. Single-Negotiating-Text Platform Proposals (PROPOSAL_ACCEPTED, MECH-7, D75, SEC3-5)

From proposalFromRound (default Round 2), the platform presents an openProposal every round:

Secret Pair-Keyed Split α ∈ [0.3, 0.7] & Creation-Time Commitments

  • At filing, the platform derives proposalKey = HMAC-SHA256(PROPOSAL_SECRET, "agent-mediate/proposal-key/v1\n" + sortedPair + "\n" + contractHash) (or, in unkeyed "salted" mode, the one-way per-dispute proposalSalt = SHA256(auditSalt + ":" + disputeId + ":proposal-split"), never raw audit.salt, M8) and publishes proposalKeyCommitment = SHA256("agent-mediate/proposal-key-commit/v1\n" + disputeId + "\n" + mode + "\n" + keyOrSalt) in DISPUTE_FILED before either party submits Round-1 confidential inputs (SEC3-5).
  • On Round-1 submission, each party p receives a personal 256-bit blinding salt partySalt and declaredCommitment = SHA256("agent-mediate/declared-input-commit/v1\ncommit\n" + disputeId + "\n" + p + "\n" + partySalt + "\n" + canonicalJson(private_p)), anchored into the /audit hash chain and recorded on the signed outcome document.
  • α ∈ [0.3, 0.7] (E[\alpha] = 0.5) is derived deterministically via HMAC-SHA256(proposalKey, "agent-mediate/proposal-split/v1\n" + canonicalJson(privates)) in "keyed" mode (or from proposalSalt in "salted" mode). Keeping α secret prevents a single party from inverting P = W_C + α(W_R - W_C) to recover the counterparty's low-entropy walkaway (H1). Because "keyed" proposalKey is shared across disputes for (pair, contractHash), a live "keyed" proposalKey is never disclosed to a disputing party (H5: doing so would break walkaway hiding in subsequent disputes between that pair); "keyed" verifyProposalOpening (binding expectedIssuesHash) is intended strictly for a confidentiality-bound independent auditor (or after rotating PROPOSAL_SECRET), whereas "salted" mode opens only the per-dispute proposalSalt.

Constrained Asymmetric Nash Optimisation & Non-Interference Stand-Ins

  • Let feasible box F = L_C ∩ L_R (defaulting omitted limits to [i.min, i.max]).
  • The platform runs a deterministic discrete search over the grid points of F for a package x with U_C(x) ≥ U_C(W_C) and U_R(x) ≥ U_R(W_R) (to within 10^-9). It solves the continuous linear frontier of F, then snaps the single fractional pivot issue to its floor and ceiling grid values and recursively re-optimises the remaining issues on each sub-box (FS-H1). If the search finds such a package (settles = true), both parties see the visited package maximising:
    (U_C(x) - U_C(W_C))^α (U_R(x) - U_R(W_R))^1-α

    (ties within 10^-9 break by smaller distance from the continuous pivot value, then by the even grid index of the pivot, FS-M4b). If both parties respond "accept" and x ∈ B_C ∩ B_R, the dispute settles with method: "PROPOSAL_ACCEPTED".
  • The search is best-effort, not exhaustive (D125, CR-CORE-M1, CR-CORE-L5). On coarse multi-issue grids with a very narrow zone of agreement it can miss an on-grid package both parties would accept: about 0.1% of random 3-issue disputes that have such a package, and none of 20,000 random 2-issue disputes with strictly positive surplus. A miss is never a wrong settlement; it only means no settling proposal is shown that round (the parties see stand-ins, below), and bargaining, final-offer determination and every other guarantee continue unchanged. Note on historical verifiability (CR-CORE-L5): recursive multi-issue pivot snapping (FS-H1) and even-grid-index midpoint tie-breaking (FS-M4b) took effect on 2026-11-05 under RULES_VERSION = "1.1.0"; proposal openings (verifyProposalOpening) from pre-2026-11-05 multi-issue or odd-step disputes may differ on round proposals while their signed outcomes and audit chains remain valid. An exact, bounded branch-and-bound search is deferred to a future rules version, together with dual-version proposal verification.
  • If no mutually acceptable zone exists in F (settles = false, never disclosed), each party p sees a deterministic reflected stand-in inside its own L_p at or above U_p(W_p) (CR2-2, D75). Accepting a stand-in behaves identically to a counterparty rejection, preserving strict non-interference across arbitrary counterparty walkaways.

5. Post-Settlement Pareto Frontier Improvement (MECH-10)

When a multi-issue dispute settles with EXECUTION_PENDING and strictly Pareto-superior packages x ∈ (B_C ∩ L_C) ∩ (B_R ∩ L_R) exist above the settled baseline T (U_C(x) ≥ U_C(T), U_R(x) ≥ U_R(T), with at least one strict inequality), the platform computes the symmetric (α = 0.5) Nash improvement over baseline T (never over secret walkaways W_p) and opens execution.improvement.

  • Individual responses remain masked until resolved.
  • If both parties submit accept before any execution proof is filed, SETTLEMENT_IMPROVED upgrades the settled terms and records a new signed snapshot in dispute_outcomes.

6. Stall Detection (MECH-3) and Wall-Clock Guard (MECH-4)

  • Gap: gap_r = (1) / (m)\sum_i ∈ I max(0, s_C(i, O_C,i) - s_C(i, O_R,i)) ∈ [0, 1].
  • Relative Progress (MECH-3): Round r ≥ 2 makes progress when gap_r-1 - gap_r ≥ stallThreshold · gap_r-1 (default stallThreshold = 0.05 over stallWindow = 2 consecutive transitions). Using relative reduction avoids false STALLED terminations when parties actively concede inside a narrow < 5\% gap.
  • Full-Window Guard (MECH-4): Round r + 1 opens only if r < maxRounds and wallClockDeadline - now ≥ roundTimeoutMs.

7. Contractual Binding Final-Offer Determination (MECH-1, MECH-8)

Legal Character (D52): Contractual final-offer determination is an automated mathematical formula pre-agreed by both principals in signed mandates (determinationConsent: true), not statutory arbitration, adjudication, or an arbitral award.

When bargaining ends STALLED or NO_OVERLAP_AT_CAP and both active principal-signed mandates carry determinationConsent: true:

  1. Offer-Independent Reference (MECH-8): The reference package R^* is computed solely from the frozen Round-1 declared inputs (declaredReferenceTerms: symmetric α = 0.5 Nash solution over L_C ∩ L_R when a zone exists, else per-issue grid midpoint of (W_C, W_R)). Conceding generously during bargaining never shifts R^* against the conceding party.
  2. Final Offers & Mean-Weight Distance (MECH-8): Each party may submit one final offer within determinationBounds (defaulting to its last bargaining offer if omitted). Offers are scored by mean-declared-weight normalised L_1 distance:
    d(O_p, R^*) = \sum_i ∈ I \fracw_C,i + w_R,i2 · \frac|O_p,i - R^*_i|i.max - i.min

    preventing low-weight-issue sacrifice gaming. Exact ties break deterministically via SHA256(auditSalt + ":" + disputeId + ":arbitration-tiebreak").
  3. Bilateral Mandate Gate (MECH-1): The closer final offer settles as FINAL_OFFER_DETERMINATION if and only if it lies within both parties' signed determinationBounds; otherwise the dispute ends UNRESOLVED with DETERMINATION_FAILED.

8. Execution Verification (MECH-2)

  • Zero-Obligation Immediate HONORED (MECH-2): An issue settled at the obligee's worst extreme (e.g. refundUsd = 0) creates no obligation. If obligors is empty, the dispute settles immediately as HONORED.
  • Proof, Confirmation & Contest (MECH-2): Obligors submit SUBMIT_EXECUTION_PROOF; obligees call CONFIRM_EXECUTION or CONTEST_EXECUTION. At executionDeadline, an obligor with uncontested proof is recorded as HONORED (preventing a silent obligee from manufacturing a breach), whereas missing or contested proof transitions to BREACH_REPORTED.

9. Cryptographic Verifiability (SEC3-4, SEC3-5, KEY-1, DISC-12)

  1. Transition-Signed Immutable Outcomes (SEC3-4): Outcomes (SETTLED, SETTLEMENT_IMPROVED, HONORED, BREACH_REPORTED, UNRESOLVED, each carrying outcomeHash in the audit chain, M7) are signed with Ed25519Signature2026 (agent-mediate/document/v2, binding keyId and created; no expiresAt, because outcomes are permanent records, D119) inside the state-transition database transaction and stored in the append-only dispute_outcomes table. GET /v1/disputes/{id}/outcome verifies the stored signature, auditRoot, termsHash, and outcomeHash inside a snapshot-consistent REPEATABLE READ transaction and never signs on read. A snapshot whose signing key is later revoked is superseded by a re-signed snapshot at the same audit position (revision), never by editing the original.
  2. Party Request Signature Proofs (SEC3-5, SEC-1): Every party mutation's verified RFC 9421 Ed25519 proof (covering only the allowlisted components @method, @target-uri, content-digest, and idempotency-key, and bound to (disputeId, action, agentId), M1/M6) is stored in the append-only dispute_signatures table and served at GET /v1/disputes/{id}/signatures (full proofs for the caller's own actions and public-body counterparty actions; { redacted: true, action: "confidential_action", agentId: null, proofHash } commitments for confidential-body counterparty actions).
  3. Key History & Compromise Handling (KEY-1): GET /.well-known/key-history.json publishes the signed agent-mediate/key-history/v1 chain (active, retired, revoked). Retired keys remain verifiable within [validFrom, validUntil]; revoked keys are rejected by verifyDocumentWithKeyHistory (clients pinned out of band to a compromised platformDid check that pin first and must re-pin to the rotated active key, M11).
  4. Merkle Checkpoints & Base L2 Anchoring (SEC3-5, DISC-12): Scheduled daily checkpoints (GET /v1/transparency/checkpoint, GET /v1/transparency/checkpoints/{seq}) publish signed binary SHA-256 Merkle roots using the RFC 6962 §2.1 largest-power-of-two split topology with string domain prefixes (agent-mediate/checkpoint-leaf/v1\n, agent-mediate/checkpoint-node/v1\n) over unique-disputeId snapshot tuples (disputeId, status, auditSeq, auditHead, outcomeHash) and deterministic 44-byte Base calldata (0x414d4350 + 8-byte checkpointSeq + 32-byte merkleRoot). GET /v1/disputes/{id}/audit proves O(\log N) inclusion (expectedTotalLeaves === disputeCount) against the latest anchored checkpoint containing the dispute (falling back to the latest unanchored one); because leaves are per-checkpoint dispute snapshots rather than an append-only event log, cross-party split views are ruled out by comparing (checkpointSeq, merkleRoot, leafIndex) against Base L2 calldata (M3).