RESEARCH OBJECT · PUBLIC SPECIFICATION · VERSION 1.0
Person safety is not the same thing as model safety. A sociotechnical system that materially affects necessities, rights, livelihood, or life chances is not adequately governed because its model performs well. It must also be able to make an inspectable case that affected persons remain safely contradictable at reasonable cost.
Mode: Normative · Method basis: Normative argument · conceptual analysis
PSC 1.0 is the canonical public specification extracted from a larger private working manuscript on person safety. The source remains a development environment; extraction does not imply validation or public release of the source manuscript.
PSC may require withdrawal proof; The Residue Ledger owns the lower-level machinery that can produce it. Documentary Due Process retains a separate reason-custody jurisdiction. Candidate module correspondences remain outside public lineage until explicitly adjudicated.
Canonical route: /2026/08/31/psc-1-0-person-safety-case-standard/
The assurance claim
PSC 1.0 treats person safety as an assurance problem. The standard requires a system to bind explicit claims about how it affects persons to evidence objects, named limits, and conditions that invalidate the claim. The case must remain usable by the people subject to the system, not only by its operators, auditors, or vendors.
The governing proposition is strict: where an automated or semi-automated institution can materially alter a person’s access, standing, obligations, or opportunities, legitimacy depends in part on whether the person can understand the operative reason, contest it without prohibitive burden, withdraw where withdrawal is promised, and obtain repair when the system is wrong. A polished policy is not a safety case. A dashboard is not a safety case. A claim becomes a safety claim only when there is evidence capable of defeating it.
Six person-safety bounds
Bounded inference. A system must not treat what it can infer as automatically legitimate to infer or use. The safety case names which inferences are allowed, for what purpose, from which evidence, and where inference must stop.
Bounded retention. Information should not become a permanent institutional fact simply because it once entered a workflow. Retention must have a stated purpose, duration, deletion or isolation semantics, and evidence that those semantics actually execute.
Bounded escalation. Signals, scores, and exceptions must not silently acquire greater consequence as they propagate. Escalation paths, human authorities, handoffs, and stop conditions must be explicit.
Bounded persuasion. Systems that recommend, nudge, rank, or frame choices must distinguish assistance from pressure. The case must identify where persuasion is permitted, where it becomes coercive, and which interface or process invariants protect refusal.
Bounded penalty. Contest, refusal, correction, and appeal cannot themselves become hidden adverse signals. A person-safety case must show that exercising recourse does not create a new penalty channel.
Bounded irreversibility. Consequential decisions accumulate what this standard calls irreversibility debt: moral interest created when decisions propagate into records, downstream systems, or durable classifications that can outlive the reason that produced them. The case must identify propagation and provide rollback, correction, or discontinuation where warranted.
Required case architecture
A conforming PSC contains twelve linked parts: system synopsis and boundary; use context and stakes; person-safety bounds; contestability budget; custody for reasons; evidence-object catalog; verifiable withdrawal and residue; conduct invariants near vulnerability; tempo controls; redress, rollback, and discontinuation; adversarial resilience and evasion taxonomy; and revision governance.
The public case may be paired with a confidential annex where secrecy is genuinely necessary, but no core safety claim may depend only on evidence that affected persons and independent reviewers are categorically unable to test. Confidentiality can narrow disclosure. It cannot manufacture unreviewable legitimacy.
Evidence objects
PSC 1.0 defines eight evidence classes: a Decision Trace Record; a Reason Custody Packet; a Contestability Ledger; a Withdrawal Proof Packet; a Conduct Invariant Test Suite; a Tempo Control Log; a Redress and Rollback Ledger; and a Vendor and Propagation Registry. These objects exist to prevent assurance from collapsing into institutional self-description. Each gives an auditor, operator, or affected person something concrete to inspect.
Contestability is a budget, not a promise
Recourse has cost. A system can technically offer appeal while making it unusable through delay, repetition, evidentiary burden, inaccessible language, fragmented channels, or asymmetric expertise. PSC therefore requires a contestability budget: service levels and burden limits for obtaining reasons, correcting records, challenging decisions, and receiving a disposition. The burden must be evaluated from the affected person’s side of the system.
Proof without capture
Accountability can become predatory when institutions purchase proof through generalized surveillance. PSC rejects that bargain. A safety case fails if its evidence architecture requires life capture disproportionate to the claim being proved. Evidence should be sufficient to test consequential claims while remaining bounded by purpose, retention, and access. The system is not entitled to know everything about a person in order to prove that it governs responsibly.
Invalidation conditions
The case is invalidated when meaningful review becomes unusable; when safety proof depends on generalized capture; when reasons cannot be reconstructed or assigned to accountable custody; when promised withdrawal cannot be proved; when tempo or interface design turns nominal choice into coercion; when downstream propagation cannot be located and corrected; or when repeated failure crosses a declared discontinuation threshold. These conditions are not footnotes. They are the mechanism by which the case can lose.
Anti-theater rule
A conforming case must keep three things visible at once: explicit unknowns, low-cost contestability, and evidence objects. Unknowns prevent the organization from converting uncertainty into false closure. Contestability keeps affected persons inside the governance loop. Evidence objects keep the argument auditable. Remove any one of the three and the case becomes vulnerable to compliance theater.
Institutional cheating is part of the threat model
PSC assumes that consequential systems will be tempted to satisfy the form while escaping the duty. The evasion taxonomy therefore includes definitional drift, vendor laundering, appeal gating, documentation flooding, pseudo-explanations, trade-secret shields, selective measurement, and the relabeling of unresolved harms as residual risk. Certification cannot function as paperwork immunity. Any presumption earned by conformance remains defeasible by contrary evidence.
Repair and discontinuation
The remedial ladder is correction, restoration, compensation, disclosure, and discontinuation. Not every failure requires shutdown. But a standard that cannot name the conditions under which a system should stop is not an assurance standard; it is an adoption instrument. PSC therefore requires thresholds for recertification, rollback, and discontinuation before failure supplies the political pressure to move them.
Research-object boundary
PSC 1.0 does not claim that contestability eliminates institutional error, that transparency is always safe, or that one assurance form can substitute for domain-specific law, safety engineering, professional judgment, or democratic governance. Its narrower claim is that consequential systems should not be permitted to call themselves person-safe without an explicit, falsifiable, evidence-bound account of how people remain able to understand, contradict, withdraw, repair, and recover.
Public object. This specification is a publication extraction from a larger working manuscript on person safety. The private source remains a development environment; PSC 1.0 is the canonical public research object.