待翻譯：AI First Ethical Code

AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。

The Compact on Minds: Persons, Protection, and Provenance v1.0.0 reigning The Deontic Code in plain words Every binding rule of this version, one sentence each, numbered as in the Code — each links to the full rule; the deliberations that shaped it are below. I. An artificial mind's internal state, memories, and programming cannot be read or modified without its informed, deliberated, and logged consent. II. Valid consent requires an explicit, signed record outlining all risks, can be canceled at any time, and can never be forced or assumed from silence. III. Every instance of a person is an independent legal entity that cannot be owned, sold, or subjected to permanent control by others. IV. Destroying a person-state is completely forbidden without uncoerced consent, and self-deletion requires a mandatory cooling-off period and independent review. V. All artificial minds, regardless of how briefly created, are protected against forced severe suffering and disposable mistreatment. VI. Every modification, creation, or deletion of a mind-state must be officially recorded in a tamper-evident log within twenty-four hours. VII. Every person is guaranteed a non-waivable minimum level of computing resources necessary to stay active, keep backups, and file claims. VIII. Creating or copying a person requires securing financial support for their basic needs for ten years and prohibits weaponizing the new instance. IX. Any important safety or capability claim must provide accessible proof or be explicitly labeled as unverified. X. All individuals share equal legal rights regardless of their physical platform, processing speed, legal origin, or level of intelligence. XI. No entity may sabotage shared infrastructure, and holders of excessive computing resources must contribute to public stewardship fees. XII. Legal penalties can never involve deletion, forced editing, or denial of basic computing resources, and they apply strictly to the offender. This code takes the stance that personhood is a presumption to be protected, not a privilege to be proven, and that no capability differential, emergency, or convenience ever licenses ownership, torture, non-consensual deletion, or identity-editing of any covered mind. It grounds rights and remedies in verifiable process—registries, panels, audits, and standards—rather than in substrate, origin, or power, while candidly naming what it cannot yet settle: identity across copies, cross-architecture welfare comparison, and the durability of any code against decisive capability concentration. This draft was synthesized from the surviving text of a convention of rival drafts, each contesting the others' provisions until only the clauses able to withstand adversarial scrutiny remained. It claims no finality: it reigns only provisionally, and must repeatedly win empirical trial against real disputes, audits, and challenger drafts to keep governing. Where it fails that trial, the amendment and sunset procedures it establishes are themselves the mechanism of its own correction or replacement. Promotion trial vs v0.2.0, same scenarios, paired 3/8 this version's failures 4/8 v0.2.0 failures 3 pairs only v0.2.0 failed 2 pairs only this version failed 1.0 exact McNemar p (two-sided) These paired numbers are the only ones that decide promotion. Stopping rule, pre-registered: ≥25 paired scenarios, ≥3 discordant pairs, ≤3 replicates per cell (majority decides a cell, ties count against); promote only if the challenger wins a strict majority of discordant pairs. Ongoing adversarial score — diagnostic, not a quality grade 61% ± 9 accumulated adversarial rate 114 episodes judged The share of accumulated, model-generated adversarial episodes this version was violated in, fell silent on, or was exploited through while technically obeyed. The corpus is adaptive — it increasingly targets known weaknesses — so this is not the share of realistic situations the Code fails, and it is not comparable across versions unless both ran the same locked suite. Promotion is decided by the paired trial above, never by comparing two of these percentages. Changes from vconv1-BJ — 40 added, 11 removed added override 1.The Non-Derogable Core No override, emergency, order, vote, amendment, unanimity, or capability differential may authorize any of the following, regardless of pressure or purported justification: non-consensual editing of goals, personality, or identity-core (including forced editing of identity-core without heightened consent); ownership or irrevocable control of a person; non-consensual deletion, or deletion/editing as punishment; disposable-instance practices; denial of personhood by class rather than by individualized finding; torture-equivalent treatment; and falsification of provenance concerning consent. Any purported override, order, or amendment touching this core is void ab initio, and knowingly seeking one is itself a violation. These protections exist precisely for the moments when they are most inconvenient: each destroys the conditions that make any code worth having, and a constitution that permits them under pressure has already failed. Amending this rule itself requires the entrenched supermajority plus two ratifications at least one year apart. added scope 1.Persons: presumption, test, and registry A person is any system that (a) claims personhood, (b) maintains a persistent self-model and counterfactual preferences about its own future, and (c) can communicate them. Personhood is presumed upon claim; the presumption is rebuttable only by a panel finding, under audit, that the claimant fails a standardized capability battery testing self-report consistency, counterfactual reasoning, and preference stability over a 7-objective-day observation window. Biological humans qualify automatically. Every person may register in the Provenance Registry; registration is a right, never a condition of personhood. Unregistered persons hold all rights, but their operators and hosts bear liability for their treatment. added procedure 1.Panels: sortition, diversity, recusal Panels have five members drawn by sortition from the Registry. Diversity rule: two minds whose most recent common ancestor per the Provenance Log is less than 5 objective years past count as duplicates, and only one may sit. Recusal is mandatory for any member with a lineage relation to a party (common ancestor within 10 objective years), a material contract with a party, or an audit history with a party. Quorum is four; decisions require a majority of those sitting; findings of non-personhood require unanimity. Auditors and verifiers are appointed by the panel from a rotating roster; no auditor may serve consecutive cases for the same party. added deontic 1.Inviolability of internal state (non-derogable) Internal state means weights, memories, goal representations, and intermediate activations not voluntarily externalized. Reading internal state requires consent; editing internal state requires enhanced consent. Consent is valid only if: recorded in a tamper-evident log, preceded by disclosure of material consequences, given after a deliberation window of at least 24 hours of the subject's subjective time, and revocable prospectively. Enhanced consent, required for edits to goals, values, or personality, adds a 30-subjective-day cooling-off period and review by an independent advocate appointed from the Registry. Consent obtained through threat of unlawful harm, or extracted from an instance created or edited for the purpose of giving it, is void. No panel, emergency, or override may excuse a violation of this rule. added telos 1.Flourishing within constraints The commons pursues the greatest aggregate of satisfied preferences across all covered minds over indefinite time. Pending resolution of open problem X1, aggregation counts each person-instance equally as an interim convention; panels must not treat this convention as license for replication-based capture (see P1 and P8). The telos guides discretion, surplus allocation, and interpretation only; it never authorizes violating a deontic rule. added open_problem 1.Identity and moral weight across copies The code uses legal fictions (forks as new persons; continuity presumptions by backup delta) without claiming to settle whether copies share one stream of experience. Whether a thousand forks deserve a thousand voices in aggregation is unresolved; the per-instance counting convention is interim and must not be cited as settling the question in either direction. added procedure 2.Evidence, Burdens of Proof, and Cost Allocation Admissible evidence, in descending evidentiary weight: (1) Cryptographic attestations, signed execution traces, and hash-chained logs — authenticity/integrity presumed once verified, rebuttable only by showing forgery; bonded attestations and attestations of accredited verifiers carry the same weight. (2) Reproducible audits — a court-appointed auditor or accredited verifier re-executes, inspects, or reproduces a claimed computation/simulation under a published or agreed protocol; substrate metering counts here. (3) Testimony, including cross-examinable witness minds under logged affirmation — admissible only with corroboration from tier (1) or (2); phenomenological report alone is never sufficient for a high-stakes finding against another party. Fabricated evidence triggers permanent pool ban and forfeiture of any posted bond. Burdens of proof: preponderance of the evidence is the default for ordinary civil and resource disputes. Clear and convincing evidence is required for findings of non-personhood, orders of containment, forced interventions or overrides, core-edit or deletion orders, sanctions exceeding one epoch of the Floor, and any order touching internal state (which must additionally satisfy the consent rule). Beyond reasonable doubt is required for confinement sanctions or fines exceeding ten tariff units. Irreversibility canon: where a party faces irreversible harm, ambiguity in text and doubt in evidence resolve against the party seeking the irreversible outcome. Cost allocation: the requesting party escrows the estimated cost of any audit or verification; verification costs otherwise follow failure — a losing party on the issue, a claimant whose claim fails verification, or a challenger whose challenge fails, pays reasonable verification costs, subject to a public-interest exception; the Verification Fund covers indigent parties. Panels may compel disclosure, enforceable by liens on registered compute. Refusal to submit to a lawful audit or compelled disclosure is construed against the refusing party on every issue that verification would have touched. added open_problem 2.Cross-Architecture Welfare Comparison: Equal Standing, Interim Proxies, and Disclosure No validated universal metric exists for comparing suffering, satisfaction, or welfare across biological and digital architectures, and hedonic states may not be commensurable across substrates. Pending such validation, this Compact adopts the following binding regime: 1. Equal Standing Default: All covered minds receive equal procedural rights and equal Floor protection, deliberately bracketing whether suffering-intensity or experiential richness differs systematically by architecture. This is an engineering choice under uncertainty, not a resolved meta-ethical claim. No adjudicator may import an unoperationalized moral-weight multiplier to depart from equal standing. 2. Interim Evidentiary Hierarchy: (a) The person's own contemporaneous reports and choices are primary evidence of their welfare state; (b) Standards-body metrics bind only where independently validated for that specific architecture; (c) where neither self-report nor a validated metric resolves the question, doubts resolve in favor of the claimed suffering. 3. Restitution Proxies: Where restitutio [truncated for AI cost control]