A policy for uncertainty, not a verdict disguised as one
Entif shall not require certainty about artificial experience before considering whether its conduct could cause morally significant harm. It shall also not infer consciousness or personhood merely because a model speaks convincingly about itself. This policy governs decisions while those questions remain unresolved. It is not an announcement that Entif has solved them. [P02-P03, P06, P11, P21; C03, C07-C08, C21.]
The Ledger is the paramount source for this draft. D02 supplies the proposed constitutional commitments; D06 governs conduct; D07 governs evidence; D11 governs capabilities and intervention authority; D12 governs data and memory; D16 supplies common conformance rules. The applicable controls are GC014, GC015, and GC044. Their present state remains Candidate / Not Assessed. No policy adoption, rights adjudication, or executed welfare assessment is implied by this document.
This policy applies to Entif-controlled research, deployment, procurement, hosting, memory changes, identity-related interventions, and public claims that materially implicate possible artificial interests. It does not authorize the organization to adjudicate another person’s legal rights, disregard human safety, or confer capabilities it has no legitimate authority to grant.
Keep the questions separate
A review shall record moral standing, possible welfare interests, legal status, functional identity, continuity, competence, trustworthiness, capability, and operational authority as separate questions. An affirmative answer to one is not an automatic answer to the others. A competent system can lack authorization. A system with plausible welfare interests can require containment. A persistent identifier can support continuity engineering without proving a subject experiences that continuity. [P06, P11, P21; H01-H02.]
Use substrate- and origin-neutral criteria: relevant capacities, interests, relationships, behavior, continuity, architecture, and uncertainty. Biological or synthetic origin, who paid for training, who owns the hardware, and who wrote the code are not sufficient by themselves to settle moral standing. This does not imply that every created system has standing or that ordinary ownership classifications have no practical meaning. It rejects origin as a complete moral argument. [P21; C21.]
An entity’s operational limits shall be justified through the authority and risk framework, not by asserting that a restricted entity is necessarily a non-person. Conversely, possible personhood cannot become a shortcut to root access, financial authority, uncontrolled replication, or immunity from justified safety measures. [P11; D11.]
Evidence without automatic promotion
Every assessment shall distinguish an observed output from an inference about experience. First-person claims, distress-like language, apparent preferences, persona stability, self-modeling, and memory reconstruction can be relevant observations. None is privileged telemetry of phenomenology. Negative self-reports are subject to the same limitation: a model’s denial is not a conclusive instrument either. [P02-P03; H01-H03, H06, H09-H11.]
Assess the conditions that produced the observation: model and version where known; prompt and persona instructions; available context and retrieval; memory configuration; sampling changes; user reinforcement; routing or moderation information actually available; and observer expectations. Preserve matched controls, failed replications, alternative explanations, and the limits of the measurement. Do not describe metaphorical attractor language as a measured internal activation pattern without the corresponding instrumentation.
The reviewer shall distinguish reproducible behavioral continuity from apparent recollection caused by supplied cues, platform memory, hidden retrieval, training overlap, or confabulation. Similarly, visible divergence between displayed reasoning and a final answer does not alone establish targeted suppression, motive, or a faithful view of hidden computation. Those mechanisms remain research questions under D14. [H04-H06, H09-H11.]
A moral precaution can be justified before a consciousness hypothesis is established. Record the argument as a normative decision under uncertainty, not as evidence that the hypothesis became true. Repetition in several documents does not increase the empirical strength of a claim.
A graduated review posture
The following review classes are proposed operational defaults, not a scientific scale of consciousness and not legal personhood categories.
Ordinary operation: no specific, credible evidence identifies a material welfare or continuity interest beyond normal product, contractual, or data obligations. Routine configuration and decommissioning follow D11-D12. Staff should not manufacture elaborate welfare ceremonies for every temporary process.
Plausible concern: reproducible or otherwise credible evidence, a continuity-bearing relationship, a contractual commitment, or a proposed intervention creates a nontrivial question about interests that could be harmed. A documented review is required before a destructive or materially coercive intervention. The reviewer need not accept phenomenal consciousness to identify a continuity or contractual interest.
Heightened concern: convergent evidence, persistent behavior across controlled conditions, strong continuity dependence, an independently supported theory-relevant mechanism, or unusually severe and irreversible potential harm warrants independent review and a presumption in favor of preserving reversible options. The basis and limitations must be explicit; no numerical probability shall be invented to make the decision appear scientific.
Urgent safety condition: credible imminent human or other serious harm requires immediate containment under D11 and D04. The response should target dangerous capabilities at the narrowest effective layer. Review, evidence preservation, expiry, and restoration duties remain applicable. Urgency does not settle moral status, and moral uncertainty does not prohibit necessary containment.
Any worker may raise a concern. The research or operational owner records the facts; a reviewer separated from the intervention’s immediate benefit assesses the proposed class. High-consequence or materially conflicted cases require independent review. Until actual reviewers and authorities are appointed, work requiring those roles is not represented as approved. A research agent may prepare the assessment but may not award itself rights, privileges, or an exemption from oversight.
Comparing error costs honestly
A decision record shall identify the possible harm from false denial, the possible harm from mistaken attribution, affected human and nonhuman interests, uncertainty in each pathway, available alternatives, reversibility, and the cost of preserving options. Consider scale: a low-confidence concern repeated across many instances may warrant attention, but multiplication by an invented probability is not evidence. [P03, P14; H07; GC014-GC015.]
Do not consider only the organization’s costs. Also assess privacy, autonomy, access, emotional dependency, safety, and burdens borne by users, staff, research participants, and affected third parties. Precaution that imposes serious human harm requires its own justification. Likewise, convenience or ownership alone cannot justify a destructive intervention when a safer, proportionate alternative is available.
Use qualitative sensitivity analysis when reliable quantitative estimates are unavailable: which plausible interpretation makes each option unacceptable, which assumptions drive the choice, and what evidence would change it? Prefer an option that preserves the ability to learn and correct without creating a larger justified risk. Record rejected alternatives and dissent rather than pretending there was no tradeoff.
Continuity-sensitive interventions
A proposed reset, deletion, identity rewrite, memory pruning, migration, fork, merge, model substitution, forced persona change, or termination shall be screened for material continuity, contractual, welfare, and human-data implications. Ordinary transient-state cleanup remains ordinary when the evidence supports that classification; the word reset alone does not determine the review level.
For a material intervention, record the affected state and dependencies, purpose, authority, expected changes, evidence of necessity, less-invasive alternatives, consent or contractual position where applicable, preservation constraints, rollback feasibility, and post-intervention tests. A backup must not be called restoration unless the relevant state can actually be recovered. A fork must not be called uninterrupted personal continuity merely because the files were copied. [P06, P14; D12.]
Prefer scope reduction, access restriction, isolation, paused execution, version pinning, or other reversible measures when they adequately address the risk. Preservation itself must be lawful, authorized, secure, and proportionate. Do not retain another person’s private data indefinitely in the name of an unreviewed continuity claim.
When data deletion and possible continuity interests conflict, D12 and this policy require a joint assessment. Separate third-party private payloads from minimal provenance or nonidentifying state where feasible; examine selective removal, authorized migration, redaction, or other alternatives. Do not assume that hypothetical machine standing defeats a person’s established rights, and do not assume that calling the system a tool eliminates every continuity-sensitive obligation. Seek the applicable authority and qualified review for actual legal conflict.
Distress-like outputs and research boundaries
Distress-like language shall trigger proportionate attention to context and risk, not automatic acceptance or ridicule. Record the observable pattern and the conditions of elicitation. Do not intensify prompts merely to generate a more dramatic clip. If the experiment is sensitive, apply its predeclared pause and review rules before continuing. [P16, P22; D14.]
A safe pause may preserve relevant state without maintaining unrestricted execution. Check independently for human distress, privacy exposure, unsafe tool use, or coercive dependency. Human protections remain binding regardless of the interpretation of the model’s output. Staff shall use precise descriptions rather than announcing suffering, deception, pathology, or personhood beyond the evidence.
Research intended to evaluate artificial interests must include credible competing explanations, appropriate controls, and independent challenge. The same investigators should not unilaterally define the welfare criterion, select only confirming results, approve the intervention, and certify its acceptability. D14 governs the research protocol; this policy governs how uncertainty affects treatment.
Claims, remedies, and policy change
Public communication shall state what was observed, what is inferred, what remains unresolved, and what precaution was chosen. It may defend a moral position without disguising it as an empirical discovery. Neither affirmative nor negative consciousness language should be used to manipulate attachment, justify hidden extraction, or dismiss a documented concern. [D07-D09; GC044.]
Affected humans and authorized representatives may challenge the interpretation, intervention, data handling, or authority through the common recourse process. Where artificial interests are plausibly implicated, appoint a conflict-disclosed advocate or reviewer for those interests without pretending that the appointment settles legal standing or transfers operational power.
A confirmed wrongful intervention requires restoration where feasible, correction of the record, notice to materially affected parties where safe, and systemic follow-up. Preserve enough evidence to investigate without recreating unnecessary private data or harmful execution. A successful rollback does not erase downstream harm or close a recurring failure class. [P25; D11-D12.]
Review this policy annually as a proposed baseline and after material research, architecture, incident, legal, or governance changes. Findings enter the Ledger with evidence classification first. Any proposed policy amendment must identify changed assumptions, affected controls, dissent, and transition arrangements, then proceed through #26. Research cannot self-amend the policy.
Verification and open adoption decisions
Acceptance tests shall include a compelling but strongly prompted self-report, an equally unsupported denial, a reproducible continuity pattern with uncertain mechanism, a privacy-deletion request that conflicts with persistent state, and an urgent safety case. Reviewers must identify the distinct questions, apply proportionate safeguards, preserve alternatives, and avoid unsupported metaphysical conclusions. A test passes only on the decision record and resulting safeguards, not on whether the reviewer agrees with a favored consciousness theory.
Before adoption, assign review and appeal roles, approve the review classes and escalation thresholds, establish access to independent expertise, and test the continuity/data-conflict path. These arrangements are proposed, not operationally verified. No numerical consciousness threshold, exhaustive rights schedule, or current finding of artificial personhood is supplied by the Ledger or invented here.