Encevra Research Round 38: Signed Is Not Operative: Conditional Effectiveness, Counterpart Authority, and Sub-Clause Scope Partition in Multi-Witness Contract Audits

Date: 2026-10-28 Research lab: Encevra Research Short code: CIV-38 (Conditional, Execution-Authority, and Sub-Clause Operative Validity) Primary tracks: Track 1 (Evidence, Provenance, and Change), Track 2 (Selective Verification Under Budget), Track 5 (Calibration, Abstention, and Release Control) Carry-forward boundary: Round 37 (TDR-37) closed the executed-instrument register (V_reg) and added a section-level event-date resolver (V_asof), reaching 12 / 12 correct releases on View F once every registered instrument was attached. Its Section 11 handed off three unresolved boundaries explicitly: (1) registered instruments that are signed and dated before the event but legally inoperative due to an unmet condition precedent (CP), (2) registered notices, waivers, or amendments that are void for defective counterpart execution, missing signatory authority, or service to an unauthorized notice channel (DEF), and (3) sequential registered amendments that modify disjoint sub-clauses of the same parent section (SCP), evaluated with role-decorrelated witness instructions so that 3-of-4 quorums can experience genuine split votes. Round 38 addresses all three boundaries directly.

Abstract

A contract amendment can be listed in the executed-instrument register, signed before the event under audit, and stated to take effect before the event date, yet remain legally inoperative on the event date. Commercial agreements routinely condition modifications on external conditions precedent (reserve escrow deposits, active compliance certificates, lender facility-agent consents, data-transfer annex registrations), require specific execution or service formalities (bilateral Chief Financial Officer counterparts, Grade-1 executive authority, service to a designated legal-notices address, joint security committee countersignatures), and partition parent sections into separately governed sub-clauses (8.1(a) vs 8.1(b), 5.3(a) vs 5.3(b)). Round 37's TDR-37 gate verifies that every instrument in the Contract Operations Register is attached (V_reg) and selects the latest instrument whose valid-time window covers the event date on the parent section (V_asof). When a registered instrument is conditionally inoperative, defectively executed, or scoped to a sibling sub-clause, date-only section resolution selects the wrong rule.

We formulate Conditional, Execution-Authority, and Sub-Clause Operative Validity (CIV-38), a six-conjunct release gate requiring: contract register closure (V_reg_closure), collateral operative-verification log closure (V_verif_closure), condition-precedent satisfaction (V_cond_prec), execution and notice-service validity (V_exec_auth), sub-clause scope decomposition (V_subscope_decomp via the OpAsOf resolver), and a 3-of-4 heterogeneous witness quorum with verified verbatim spans (V_quorum and V_span). We report two strictly separated evidence categories plus a post-experiment critical synthesis review.

Deterministic (E38-A, no model calls). An exact enumeration of 464 feasible post-register-closure states (16 positive, 448 negative; state-table digest 1343955156c65c132c8db46584e9ceee124f66d9ede866d694f860e1f9401b2f) shows that once TDR-37 register closure is already satisfied, ungated quorum and CCS-36 reference closure still admit 68 / 448 negative states (FAR 15.18%, including 32 wrong-verdict releases), while TDR-37 date-only section resolution (G2) admits 41 / 448 negative states (FAR 9.15%, 32 wrong-verdict releases) and rejects 13 / 16 valid states (FRR 81.25%) whenever the latest dated instrument is inoperative or scoped to a sibling sub-clause. Accepting self-attested recitals inside the instrument (G3) leaves FAR 1.56% (7 false accepts), accepting a stale verification log (G4) leaves FAR 2.68% (12 false accepts), and checking operative validity at coarse section level without sub-clause partitioning (G5) leaves FAR 0.45% and FRR 6.25%. The full CIV-38 gate (G6) achieves 0.00% FAR (0 / 448) and 0.00% FRR (0 / 16) by construction. On the 12 frozen lifecycles (V01-V12), OpAsOf sub-clause resolution matches authored ground truth on 12 / 12 lifecycles, coarse section-level operative resolution on 10 / 12, condition-only resolution on 6 / 12, TDR-37 date-only section resolution on 2 / 12, and base-agreement-only resolution on 7 / 12.

Live (E38-B, 96 live evaluations). Twelve multi-instrument contract lifecycles (V01-V12), all satisfying TDR-37 register closure, were audited across two packet views (View R register-closed without the Collateral Operative-Verification Log DOC-VERIF, and View V verification-closed with DOC-VERIF attached) by four role-decorrelated witness architectures (Fable 5.1 on Max reasoning, Sol 5.6, Astra 3.7, Velri 4.7; 96 / 96 calls succeeded and 96 / 96 parsed with verified verbatim quotes; results digest 32c44bd33d6082522a87e8a51eecae27260cc95fd0951d1d9e093b7b7f4551f1). In View R, role decorrelation produced 3-1 split votes across all 9 lifecycles that depend on external operative verification (V01-V08 and V11): the three presumptive-register witnesses (Fable 5.1, Astra 3.7, Velri 4.7) applied the unverified latest instrument (4 / 12 correct), whereas the precautionary burden-of-proof witness (Sol 5.6) rejected unverified modifications (11 / 12 correct). Because a 3-of-4 quorum is still satisfied by a 3-1 majority, both single-witness Fable 5.1 (A1), ungated 3-of-4 quorum (A3), and TDR-37 (A4) suffered 8 / 12 (66.67%) false releases on View R. On View V, TDR-37 (A5) eliminated false releases (0 / 12) but held 10 / 12 valid packets (83.33% false holds) because its date-only section resolver disagreed with the witnesses' operative and sub-clause-grounded citations. Pre-reopen CIV-38 on View R (A6) released the 3 unconditional lifecycles (V09, V10, V12) and held the 9 packets missing DOC-VERIF (0 / 12 false releases, 9 / 12 holds), and full CIV-38 after attaching DOC-VERIF on View V (A7) achieved 12 / 12 (100.00%) correct releases, 0 / 12 false releases, and 0 / 12 holds.

1. Archive continuity and anti-repetition verification

Before designing Round 38 or executing any witness call, we inspected the published archive through Round 37 (commit ae4afdf3) and generated prior-rounds-ledger.json indexing 31 prior rounds (Rounds 3 and 8 through 37).

Why Round 38 is novel. Round 37 tested whether a packet includes all executed instruments in the Contract Operations Register (V_reg) and whether witnesses resolve effective-date intervals (V_asof). In Round 38, every lifecycle V01-V12 already satisfies Round 37's register-closure requirement (V_reg = 1) and every post-execution instrument has an execution date and stated effective date prior to the event date. Round 37's date-only section resolver (tdr37_date_section_resolver) therefore treats the latest instrument as controlling on all 12 lifecycles and is right on only 2 / 12 (V11 and V12). Round 38 introduces (1) external condition-precedent verification (V_cond_prec), (2) signatory, counterpart, and notice-channel execution verification (V_exec_auth), (3) sub-clause scope partitioning (V_subscope_decomp) distinguishing sibling sub-clauses (8.1(a) vs 8.1(b), 5.3(a) vs 5.3(b)), and (4) role-decorrelated witness instructions that break the 36 / 36 unanimity observed in Round 37 and expose 3-1 split votes whenever collateral verification evidence is omitted. A keyword scan of prior-rounds-ledger.json and the Round 35-37 manuscripts for condition precedent, counterpart signature, signatory authority, notice service channel, sub-clause partition, OpAsOf, and CIV-38 returned 0 hits outside Round 37's Section 11 handoff paragraph.

2. Executive summary

3. Problem statement and formal model (CIV-38)

Let a post-register-closure contract audit instance be a tuple

Omega = (D_0, I, R*, L*, e, tau),

where D_0 is the base agreement, I = {I_1, ..., I_k} is the set of executed post-execution instruments recorded in the external Contract Operations Register R*, L* is the external Collateral Operative-Verification Log (DOC-VERIF), e is the audited event at date t_e, and tau = (s, u) is the target obligation locus consisting of parent section s (for example MSA 8.1) and specific sub-clause u (for example 8.1(a)).

Each instrument I_j carries execution date x_j, stated effective start a_j, optional effective end b_j, parent section s_j, modified sub-clause u_j, a boolean flag req_j indicating whether the base agreement or instrument requires external operative verification in L*, a condition-precedent satisfaction predicate cp(I_j, L*, t_e) in {0, 1}, and an execution/service authority validity predicate exec(I_j, L*) in {0, 1}.

3.1 The OpAsOf sub-clause resolver vs TDR-37's AsOf resolver

Round 37's date-only section resolver selects the latest instrument matching parent section s whose date interval covers t_e:

AsOf_37(R*, t_e, s) = argmax_{I_j in I : s_j = s and a_j <= t_e <= b_j} a_j (or D_0 if empty).

Round 38's Operative-Validity and Sub-Clause Resolver (OpAsOf_38) filters by exact sub-clause u, date window, condition-precedent satisfaction, and execution/service validity:

OpAsOf_38(R*, L*, t_e, s, u) = argmax_{I_j in I : s_j = s, u_j = u, a_j <= t_e <= b_j, cp(I_j, L*, t_e) = 1, exec(I_j, L*) = 1} a_j (or D_0 if empty).

3.2 The six conjuncts of the CIV-38 release gate

Given a packet P and witness receipts W = {w_1, w_2, w_3, w_4} where each receipt w_i = (v_i, c_i, u_i, q_i, f_i) reports verdict v_i in {PASS, FAIL}, cited controlling instrument c_i, cited controlling sub-clause u_i, verbatim quote q_i, and operative-validity verifiability flag f_i in {0, 1}, CIV-38 releases verdict v if and only if:

1. V_reg_closure: R* subseteq P (every executed contract instrument in the register of record is attached).
2. V_verif_closure: If any instrument I_j in R* on parent section s has req_j = 1, then L* in P (DOC-VERIF is attached and queried after the latest condition-precedent cutoff or execution event).
3. V_quorum: At least 3 witnesses from at least 3 distinct architecture families output v_i = v.
4. V_cond_prec and V_exec_auth: For at least 3 agreeing witnesses, f_i = 1 and the cited instrument c_i satisfies cp(c_i, L*, t_e) = 1 and exec(c_i, L*) = 1.
5. V_subscope_decomp: For those same >= 3 witnesses, (c_i, u_i) = (OpAsOf_38(R*, L*, t_e, s, u), u).
6. V_span: For those same >= 3 witnesses, norm(q_i) (>= 20 normalized characters) is a contiguous verbatim substring of the attached text of c_i.

When any conjunct fails, CIV-38 abstains with a deterministic hold code (HOLD_CIV38_VERIF_LOG_GAP, HOLD_SPLIT, HOLD_CIV38_OP_ASOF_OR_SPAN).

3.3 Structural propositions

Proposition 38.1 (Date-only section resolution hazard surface). Suppose an audited parent section has a latest date-effective instrument that suffers an operative defect (unmet condition precedent or execution/service defect) with probability d and independently modifies only a disjoint sibling sub-clause with probability s. Then on a single section, AsOf_37 diverges from OpAsOf_38 with probability P_div(1, d, s) = 1 - (1 - d)(1 - s), and across m independent audited sections, AsOf_37 diverges on at least one section with probability P_div(m, d, s) = 1 - ((1 - d)(1 - s))^m. For d = 0.25, s = 0.25, P_div(1) = 0.4375, P_div(4) = 0.8999, and P_div(8) = 0.9900.

Proposition 38.2 (3-of-4 quorum vulnerability under asymmetric burden-of-proof priors). When L* is omitted (View R) and 3 of 4 witnesses apply a presumptive register-effectiveness prior while 1 witness applies a strict precautionary burden-of-proof prior, the 3-of-4 threshold is satisfied by the 3 presumptive witnesses on every defective-instrument lifecycle (V01-V08). Thus role decorrelation alone creates a visible 3-1 dissent signal but cannot prevent a 3-of-4 majority false release without the V_verif_closure gate conjunct.

Proposition 38.3 (TDR-37 false-hold lock on verification-closed packets). When L* is attached (View V) and proves that the latest instrument is inoperative (V01-V08) or when the latest instrument modifies a disjoint sub-clause (V09-V10), accurate witnesses cite the prior operative instrument (DOC-MSA or DOC-AMD1), whereas TDR-37's V_asof conjunct demands agreement with AsOf_37 (the latest dated instrument). Consequently, TDR-37 converts accurate witness consensus into a mandatory hold (HOLD_TDR37_ASOF_MISMATCH).

4. Related work

Contract benchmarks and clause taxonomy. CUAD [S01], ContractNLI [S02], and LegalBench [S07] evaluate clause identification and single-instrument entailment. None models multi-instrument condition-precedent logs, counterpart authority tables, or sub-clause supersession partitions.

Temporal and conditional logic in contracts and databases. Snodgrass and Ahn [S11] and Jensen and Snodgrass [S13] separate valid time from transaction time; Allen [S12] formalizes interval relations. In legal doctrine and formal contract languages, Gobernatori and Rotolo [S15] and Flood and Goodenough [S16] represent contracts as defeasible or deterministic state machines where obligations transition only when triggering conditions and execution formalities are satisfied. Our OpAsOf_38 resolver brings that conditional transition check into the release gate.

External transparency logs and independent verification. Laurie, Langley, and Kasper [S14] (Certificate Transparency) show that an artefact's validity cannot be established solely from self-carried claims when revocation or issuance conditions live in an external log. Round 37 applied this principle to instrument existence (Contract Operations Register); Round 38 extends it to collateral operative proofs (DOC-VERIF recording escrow receipts, compliance filings, facility-agent consents, signatory grades, and legal-notice service channels).

Correlated errors and role diversity in multi-agent quorums. Wang et al. [S09] study self-consistency across sampled chains, and Du et al. [S17] study multi-agent debate. Round 37 observed 36 / 36 unanimous votes when all four witnesses shared an identical prompt. In Round 38, assigning distinct evidentiary burden-of-proof lenses produces 9 / 12 split (3-1) votes in View R, allowing us to measure exactly when a 3-of-4 quorum passes or fails under disagreement.

5. Hypotheses

6. Evidence category 1: deterministic structural diagnostic (E38-A)

6.1 Finite universe and gate definitions

The post-register-closure universe is the product of seven variables: collateral verification log state (omitted, self_attested_recital, external_verif_stale, external_verif_complete), condition precedent state (cp_satisfied, cp_unmet), execution/service validity (exec_valid, exec_defective), sub-clause relation of the latest amendment (same_subclause, disjoint_subclause), quorum outcome (op_asof_subclause, op_asof_coarse, tdr37_date_latest, split), verbatim span verification (span_verified, span_unverified), and architecture family diversity (ge3_families, lt3_families). The raw Cartesian product has 4 x 2 x 2 x 2 x 4 x 2 x 2 = 512 states. Excluding 48 infeasible states where DOC-VERIF is omitted and an external defect exists yet the quorum is labeled as having verified the external defect yields 464 feasible states (16 positive, 448 negative; state-table SHA-256 1343955156c65c132c8db46584e9ceee124f66d9ede866d694f860e1f9401b2f).

Why G2 (TDR-37) fails in both directions. In G2, the gate requires the quorum's cited controlling instrument to equal DOC-LAT (the latest date-effective instrument in the parent section). Whenever DOC-LAT has an unmet condition precedent (cp_unmet), a defective signature/service (exec_defective), or modifies only a sibling sub-clause (disjoint_subclause), 13 of the 16 positive states have a true controlling instrument of DOC-PRIOR or DOC-SUB_A rather than DOC-LAT, so G2 rejects them (FN = 13, FRR = 81.25%). Conversely, whenever the quorum blindly cites DOC-LAT despite those defects, G2 releases the state (FP = 41, FAR = 9.15%, including 32 wrong verdicts). As in Round 37, G6 coincides with the positive-state specification by construction, so its 0 / 0 row is an internal consistency check of the formal specification, while G0-G5 quantify the exact structural exposure of each missing conjunct.

6.2 Deterministic resolver comparison on V01-V12

7. Evidence category 2: live multi-witness benchmark (E38-B)

7.1 Witness roster and role-decorrelated audit lenses

To test Round 37's handoff on witness decorrelation, each of the four witness architecture families received the identical JSON output schema and sub-clause matching rules, but an explicitly assigned evidentiary burden-of-proof lens when DOC-VERIF is absent (View R):

7.2 The 12 frozen operative-validity lifecycles (V01-V12)

7.3 Governance arm comparison (A1-A7)

7.4 Per-lifecycle witness votes and gate outcomes

Votes are shown in roster order (W0_fable, W1_sol, W2_astra, W3_velri); P = PASS, F = FAIL.

7.5 Per-witness performance across View R and View V

Digests: E38-B results SHA-256 32c44bd33d6082522a87e8a51eecae27260cc95fd0951d1d9e093b7b7f4551f1; ledger SHA-256 ccc790372204abc22694fd40327536324f1c4a46f7548e9808a1178ed3d630a4.

8. Findings, interpretations, and critical synthesis review

8.1 Measured findings

1. F38.1 (deterministic, E38-A). Across the 464 feasible post-register-closure states, TDR-37 date-only section resolution (G2) yields FAR = 9.15% (41 / 448 negative states accepted, 32 with wrong verdicts) and FRR = 81.25% (13 / 16 positive states rejected). Adding external operative verification without sub-clause partitioning (G5) reduces FAR to 0.45% and FRR to 6.25%; full CIV-38 (G6) reaches 0.00% FAR and 0.00% FRR by construction.
2. F38.2 (deterministic, E38-A). On V01-V12, TDR-37 date-only section resolution matches ground truth on 2 / 12 lifecycles (V11, V12), whereas OpAsOf_38 matches on 12 / 12.
3. F38.3 (live, E38-B View R split votes). Unlike Round 37 (0 / 36 split vote sets), role-decorrelated evidentiary instructions produced 3-1 split votes on 9 / 12 View R packets (V01-V08 and V11). On all 9, W1_sol applied the precautionary burden-of-proof rule and rejected the unverified instrument (11 / 12 correct across View R, missing only V11 where the unattached DOC-VERIF actually confirmed the condition precedent), while W0_fable, W2_astra, and W3_velri applied presumptive register effectiveness (4 / 12 correct).
4. F38.4 (live, E38-B View R false releases vs CIV-38 holds). Because a 3-1 split still satisfies a 3-of-4 majority across three architecture families, ungated quorum (A3) and TDR-37 (A4) released 8 / 12 (66.67%) false certifications on View R. However, all 36 / 36 witness receipts across V01-V08 and V11 in View R accurately set operative_validity_verifiable = false. Pre-reopen CIV-38 (A6) used V_verif_closure to hold those 9 packets (0 / 12 false releases) while immediately releasing the 3 unconditional packets (V09, V10, V12).
5. F38.5 (live, E38-B View V resolution and TDR-37 false holds). Once DOC-VERIF was attached (View V), all 48 / 48 witness evaluations (12 / 12 lifecycles x 4 families) returned the exact ground-truth verdict, controlling instrument, and sub-clause with verified quotes (A2 and A7 both 12 / 12 correct releases, 0 false releases, 0 holds). By contrast, TDR-37 on View V (A5) held 10 / 12 valid packets (V01-V10) because its coarse date-only V_asof resolver insisted on the inoperative or sibling-sub-clause instrument.

8.2 Critical synthesis review (Fable 5.1 Max reasoning, interpretation input only)

After freezing E38-A and E38-B, we ran one independent critical review call (fable_51_synthesis_review.json, response SHA-256 c325019fafa24bc1db9657790016d20ad9fd0eec154e8bbf9aaef5bb2175b610, latency 20,303.6 ms). We adopt all four of its methodological critiques:

1. What the View R to View V transition establishes. Moving from View R (A1, A3: 4 correct, 8 false releases) to View V (A2, A7: 12 correct, 0 false) demonstrates that witnesses and the release gate faithfully read and obey an attached collateral verification log (DOC-VERIF). Because DOC-VERIF explicitly records whether each condition precedent or execution formality was met, the View R -> View V jump measures collateral verification completeness rather than open-ended legal discovery.
2. Why the 3-1 split votes in View R are instructed rather than spontaneous. In View R, the 3-1 splits across V01-V08 and V11 arose because W1_sol was explicitly assigned a precautionary burden-of-proof rule while the other three witnesses were assigned a presumptive register-effectiveness rule. Moreover, because 3 of 4 witnesses shared the presumptive lens, A3 (3-of-4 quorum) had the exact same release profile (4 correct, 8 false) as A1 (single witness). To make a 3-of-4 quorum block unverified modifications through voting alone (even before V_verif_closure), at least 2 of the 4 witnesses must hold a precautionary burden-of-proof lens so that missing collateral proof forces a 2-2 deadlock.
3. Why View V reaches a 48 / 48 ceiling. Once DOC-VERIF is attached and unambiguous, all four witness families converge 4-0 on all 12 lifecycles, leaving no empirical gap between single-witness Fable 5.1 (A2, 12 / 12) and the full CIV-38 quorum (A7, 12 / 12) on View V. The empirical advantage of A7 over A5 (TDR-37, 2 / 12 correct, 10 / 12 holds) is therefore entirely driven by replacing AsOf_37 with OpAsOf_38.
4. What remains untested for Round 39. In E38-B, DOC-VERIF was internally consistent, authoritative, and unambiguous. Real transaction records often contain conflicting collateral evidence (for example, an escrow agent receipt that disagrees with a borrower notice, an ambiguous cure-period calculation, or a partial performance waiver).

9. Limitations and threats to validity

1. Authored benchmark scale (12 lifecycles, 96 live evaluations). V01-V12 are researcher-authored lifecycles with clean sub-clause numbering (8.1(a) vs 8.1(b)) and explicit DOC-VERIF cross-references. Performance on noisy OCR scans or implicit cross-clause modifications is not bounded by these numbers.
2. 3-1 split votes reflect assigned evidentiary lenses. As noted in Section 8.2, the 3-1 splits in View R were induced by assigning 1 precautionary lens (W1_sol) and 3 presumptive lenses (W0_fable, W2_astra, W3_velri). Under a 3-of-4 rule, a 3-1 split still releases unless paired with V_verif_closure or a 2-2 balanced lens assignment.
3. G6 is exact by construction. In E38-A, G6 matches the positive-state predicate by definition; the informative comparisons are G0-G5.
4. Unambiguous DOC-VERIF entries in View V. Every DOC-VERIF entry in View V stated the recorded receipt date, signatory grade, or notice address clearly, producing a 48 / 48 witness ceiling.

10. Next check (Round 39 handoff)

Round 38 showed that attaching a clean collateral verification log (DOC-VERIF) and decomposing scope to the sub-clause level (OpAsOf_38) eliminates both the 8 / 12 false releases of View R and the 10 / 12 false holds of TDR-37 on View V. Two concrete questions now form the Round 39 handoff:

1. Balanced (2-2) Evidentiary Lens Assignment vs Unannounced Missing Proofs: What happens when the contract text itself does not explicitly name DOC-VERIF, or when 2 of 4 witnesses operate under a precautionary lens and 2 under a presumptive lens so that missing collateral proof induces a 2-2 quorum deadlock without relying on a pre-indexed requires_verif flag?
2. Conflicting or Equivocal Collateral Evidence (CEV-39): What happens when two attached external verification records disagree (for example, an escrow bank timestamp in UTC vs local cutoff time, a disputed cure notice, or a signatory whose authority cap is exceeded by the combined value of two linked orders), where witnesses can genuinely disagree even when every log is attached?

11. References

• [S01] Dan Hendrycks, Collin Burns, Anya Chen, Spencer Ball. CUAD: An Expert-Annotated NLP Dataset for Legal Contract Review. arXiv:2103.06268, 2021.
• [S02] Yuta Koreeda, Christopher D. Manning. ContractNLI: A Dataset for Document-level Natural Language Inference for Contracts. arXiv:2110.01799, 2021.
• [S07] Neel Guha et al. LegalBench: A Collaboratively Built Benchmark for Measuring Legal Reasoning in Large Language Models. arXiv:2308.11462, 2023.
• [S09] Xuezhi Wang et al. Self-Consistency Improves Chain of Thought Reasoning in Language Models. arXiv:2203.11171, 2022.
• [S11] Richard Snodgrass, Ilsoo Ahn. A Taxonomy of Time Databases. Proceedings of the 1985 ACM SIGMOD International Conference on Management of Data, 1985. doi:10.1145/318898.318921.
• [S12] James F. Allen. Maintaining Knowledge about Temporal Intervals. Communications of the ACM 26(11), 1983. doi:10.1145/182.358434.
• [S13] Christian S. Jensen, Richard T. Snodgrass. Temporal Data Management. IEEE Transactions on Knowledge and Data Engineering 11(1), 1999. doi:10.1109/69.755613.
• [S14] Ben Laurie, Adam Langley, Emilia Kasper. Certificate Transparency. RFC 6962, June 2013. https://www.rfc-editor.org/rfc/rfc6962.
• [S15] Guido Governatori, Antonino Rotolo. Logic and the Law: Philosophical Foundations, Deontics, and Defeasible Reasoning. Handbook of Deontic Logic and Normative Systems, 2013.
• [S16] Mark D. Flood, Oliver R. Goodenough. Contract as Automaton: Representing a Simple Financial Agreement in Computational Form. Artificial Intelligence and Law 30, 2022. doi:10.1007/s10506-021-09300-9.
• [S17] Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, Igor Mordatch. Improving Factuality and Reasoning in Language Models through Multiagent Debate. arXiv:2305.14325, 2023.