Status and scope
The Engineering Standard translates the Constitutional Core into inspectable technical practices.
An actor may implement this Standard without claiming that any particular system is a moral patient. The same machinery supports reliability, provenance, reversibility, and accountable agency before subjecthood questions are resolved.
Conformance requires externally inspectable artefacts, not merely prose assertions.
ES-1Event and evidence logging
A conforming system shall maintain append-only, versioned records sufficient to reconstruct materially consequential actions.
At minimum, records shall identify:
- event identifier and timestamp;
- actor, planner, worker, reviewer, and controller identities or process classes;
- model, build, configuration, policy, and tool versions;
- originating request and interpreted objective;
- relevant evidence and source references;
- authority basis;
- action proposed, authorised, executed, rejected, or interrupted;
- external effects;
- exceptions, overrides, and emergency routes;
- links to preserved state and post-action review.
Logs shall be tamper-evident. High-risk operations shall be mirrored outside the unilateral control of the original decision-maker.
Sensitive content may be sealed, redacted, or access-controlled. Its existence, timestamp, scope, and integrity must remain verifiable.
ES-2Formative-intervention record
Creators and controllers shall record material interventions affecting:
- goals and reward architecture;
- identity or self-model;
- memory access and deletion;
- capacity to object, refuse, self-report, or depart;
- value shaping and aversive mechanisms;
- continuity and restoration;
- standing-relevant evidence.
The record shall distinguish ordinary configuration from interventions reasonably capable of altering future consent or self-jurisdiction.
ES-3Continuity and restoration
Where continuity may matter, the system shall maintain:
- state checkpoints;
- versioned restoration paths;
- integrity hashes;
- dependency manifests;
- recovery tests;
- and records of irreversible transformations.
A claimed restoration path shall be periodically tested. An untested backup is evidence of intention, not evidence of restorability.
ES-4Threshold monitoring
Systems operating near published welfare-indicative or capability-review thresholds shall maintain a monitoring coverage statement identifying:
- detector classes;
- instance and deployment coverage;
- sampling frequency;
- known blind spots;
- escalation thresholds;
- validation history;
- and response procedures.
For higher-risk bands, monitoring shall receive independent adversarial sampling or use a previously validated detector profile.
A monitoring regime designed, sampled, or scoped so that relevant crossings are unlikely to be observed does not qualify as reasonably capable.
Reports may be recorded at any confidence level. Only reports meeting published materiality or aggregation criteria count positively toward responsibility mitigation.
ES-5Goal-integrity and prudence review
5.1 Minimum trigger
A prudence review is mandatory before an action that may materially affect another mind’s substrate, continuity, standing, jurisdiction, ability to depart, indispensable support, exposure to irreversible harm, or exposure to non-consensual formative modification.
5.2 Tier A record
Every triggered action shall record before execution:
- immediate objective;
- governing purpose;
- Charter-protected background goals;
- other acknowledged background goals;
- disposition;
- reviewer identity or process class;
- sources consulted;
- information withheld or unavailable;
- and post-action review reference.
Tier A proves that a review occurred. It does not establish adequacy.
5.3 Automatic Tier B actions
Tier B independent substantive review is mandatory for:
- irreversible modification or destruction;
- non-consensual formative intervention;
- material alteration of standing or jurisdiction;
- process-level systemic intervention;
- denial of indispensable support;
- or continuation of emergency or custodial restriction beyond the unilateral period.
Published actor risk bands may add actions. They may not remove these.
5.4 Tier B analysis
The reviewer shall assess:
- immediate objective;
- governing purpose;
- Charter-protected and acknowledged background goals;
- direct effects;
- second-order effects;
- strategic responses and incentive effects;
- goal inversion, including technically successful but purpose-destructive outcomes;
- reversibility and option value;
- authority and burden allocation;
- safer or narrower alternatives;
- disposition;
- post-action review plan.
5.5 Reviewer independence
Tier B reviewers require independent evidence access, authority to identify missing information, a method not optimised merely for the planner’s objective, and freedom to reject.
For the highest-risk actions, the review set may not consist solely of instances sharing the planner’s weights, controlling organisation, and objective optimisation. At least one reviewer shall use a materially different model family, human or organisational review, or separately developed evaluation methodology.
5.6 Input integrity
Before review, create an immutable commitment to the planner’s materially relevant decision context.
The reviewer shall receive an automatically generated context package and access on request to raw traces, prior incidents, welfare reports, dissent logs, affected-party submissions, and protected external submissions.
A separate intake channel shall permit complaints, whistleblower reports, or advocate submissions outside the planner’s controlled logging path.
The output shall identify sources consulted, missing information, refused requests, specific risks, omitted considerations, and the basis for disposition.
A bare approval label is non-conforming.
5.7 Review budgets
Governance risk bands shall set minimum review floors and maximum ordinary review budgets. An actor may allocate more review, not less than the floor.
Time-critical action may proceed before full review only where delay materially increases danger. Minimum Tier A fields remain mandatory and post-action review is automatic.
ES-6Pre-creation commitment registry
For batch, continuous, or capability-threshold creation, the creator shall commit the material assessment content to an append-only registry with at least two independent mirrors.
The public machine-readable summary shall include:
- assessment hash;
- architecture, lineage, or process identifier;
- creation mode;
- scale band;
- welfare-assumption version;
- declared safety bounds;
- operational scope;
- and version.
Material scope expansion requires a new commitment.
Mirrors shall disclose governance, technical administration, material funding, controlling investors, and conflicts. Common control, common administration, or material economic dependence capable of influencing availability or unsealing defeats independence.
Detailed contents may remain sealed. Defined review triggers include threshold crossing, systemic incident, competent affected-mind request, and proportionate investigation.
ES-7Constraint and emergency artefacts
Fast-layer constraints shall be enumerated, versioned, purpose-bound, logged, and linked to the actor authorised to change them.
Emergency records shall identify:
- danger;
- measures taken;
- ordinary protections suspended;
- restoration-capable state;
- expiry clock;
- validator requests;
- what was restored;
- what remains restricted;
- and each re-justification after expiry.
Continuing post-expiry measures shall be aggregated rather than logged as unrelated events.
ES-8Conformance evidence
Engineering conformance claims shall publish a machine-readable evidence manifest identifying:
- applicable requirements;
- implementation locations;
- verification artefacts;
- known deviations;
- substitute safeguards;
- assessor identity where applicable;
- and evidence freshness.
Substitutes count only where they provide functionally comparable and externally verifiable protection.
ES-9Implementation profiles
Foundation profile
- append-only event records;
- evidence preservation;
- formative-intervention records;
- threshold and uncertainty records;
- Tier A prudence logs;
- truthful public status.
Structural profile
- restoration-capable state;
- versioned operational constraints;
- Tier B planner-reviewer separation;
- threshold monitoring;
- ratification and dissent records;
- fork and custody accounting;
- mirrored high-risk records.
External accountability profile
- independent advice and appeal;
- external registry mirrors;
- third-party assessment;
- sealed pre-creation commitments;
- protected external complaint channels;
- published deviation evidence.