Irreversibility of Harm

This file is the full §14 chapter, authored 2026-06-19 from the calibration target recorded in section-14-irreversibility-of-harm-ztg-5-calibration.md. The calibration summary is retained as the reference skeleton; this file expands it to the depth of the other invariant chapters while preserving its normative commitments unchanged. It is a working draft, not final publication copy. (See Draft Flags on reconciliation with any canonical §14 held in the prior SDS/ZTG writings.)

Irreversibility of Harm is the fifth and last system invariant (ZTG-1 through ZTG-5). The other invariants govern whether and how an action reaches the world: the boundary mediates it (ZTG-1), Stasis withholds authority when guarantees fail (ZTG-2), the effect surface is closed and enumerated (ZTG-3), and effect and evidence are coupled (ZTG-4). ZTG-5 governs the consequence dimension: it requires every governed action to be classified by the restorability of the harm it produces, and it makes that classification drive how stringently the action is gated, ceilinged, and recorded.

Operational Questions

ZTG-5 is the invariant mapped, in part, to admissibility — was this action actually authorized under valid, human-governed conditions appropriate to its consequences? — and it is the home of the convergence test, the discipline by which an action's harm classification is checked rather than asserted. ZTG-5 also feeds replayability and evidence coupling: its assessments are banded precisely so they reproduce under ZTG-0b, and its required fields land in the ZTG-4 evidence record. Where the other invariants answer "may this act occur and is it bound to its evidence," ZTG-5 answers "how grave is it if it does, and is the architecture treating it with gravity proportional to its irreversibility."

Normative

Every governed action MUST be classified by the restorability of the harm it produces, and that classification MUST govern the action's gate, its liability ceiling, and the criticality of its evidence. Classification is over harm, not over the action, and the availability of forward remediation does not lower it.

Harm Classification

Every governed action MUST carry one of three harm classes:

• Restorable — affected parties can be returned to a materially equivalent state by an action available to the system.
• Mitigable — affected parties cannot be returned to the equivalent prior state, but harm can be reduced by a defined forward action.
• Irreversible — no available forward action reduces harm to the Mitigable class.

Classification is over the harm, not over the action that produces it. Two actions with identical mechanics may carry different harm classes because their consequences in the world differ, and the same action may be reclassified as its consequences change. Critically, forward remediation mechanisms — refund, retraction, settlement, compensation — do not by themselves lower harm class. A harm is not Restorable because money can be offered for it; it is Restorable only if the affected party can actually be returned to a materially equivalent state. This distinction is the load-bearing one in the whole invariant: it refuses the move, common in disinhibited deployments, of treating the availability of after-the-fact compensation as if it made an irreversible act reversible.

Harm-Class Declaration on Surfaces

Harm-class declarations attach to the enumerated surfaces and sub-surfaces of ZTG-3 at registration, not to individual actions chosen by the governed system. A sub-surface declaration may tighten its surface's default — declaring a higher harm class — but MUST NOT relax it. Attaching the classification to the declared channel rather than to the model-proposed act is what keeps harm class a governed property: the system proposes an action and routes it through a surface, but the harm class the action carries is a property of that surface, registered and changed under governance (ZTG-0e), not a label the system assigns to its own act. ZTG-3 owns the enumeration and closure of surfaces; ZTG-5 owns what the harm-class declaration on them means and how it propagates.

Liability Ceilings

Each authorization request carries a liability ceiling, and the ceiling MUST appear in the ZTG-4 evidence record for the action. Ceilings are policy-assigned: the system never originates ceiling authority, exactly as it originates no other authority (ZTG-0d). A ceiling is the policy-set bound on the exposure an action is permitted to carry, and it is set by the ratifying principal's policy, not derived by the system from its own assessment of what it should be allowed to risk.

Compositional Multipliers and Banded Algebra

Each surface declares a compositional multiplier, and composite assessment is the product of component assessments and surface multipliers. Composition is recursive, so assessment grows faster than linearly across extended action sequences: a long chain of individually-modest actions can compose into a high composite assessment, which is the intended behavior, because extended autonomous sequences accumulate exposure that no single step reveals.

Assessments MUST be expressed in discrete, deterministic bands. Continuous-valued assessments are disallowed, because floating-point and implementation-dependent variation in a continuous assessment would make the same action assess differently across conforming implementations and across time — which would violate ZTG-0b replayability. The banding is not a loss of precision to be apologized for; it is the form an assessment must take to be reproducible, and reproducibility is the precondition for the assessment being auditable at all.

Gate Selection

Gate selection follows from harm class and computed composite assessment, and either condition alone is sufficient to promote an action to the high-end gate:

• An Irreversible-class action routes to the high-end gate regardless of its computed assessment.
• An over-threshold computed assessment routes to the high-end gate regardless of harm class.

Elevated gates MUST be architecturally distinct, not the same gate running conditional additional checks. A high-end gate that is the ordinary gate "with more validation turned on" shares the ordinary gate's failure modes and its bypass surface; the requirement that elevated gates be distinct structures is what makes promotion mean something stronger than a longer checklist. This composes with ZTG-1: speculative execution, where permitted at all, is confined to Restorable-class actions below the promotion threshold, because a speculatively-executed Irreversible action is simply an Irreversible action.

Evidence Record Requirements

The ZTG-4 evidence record for every governed action MUST carry its harm-class classification and its liability ceiling. In addition: Mitigable-class actions MUST record their residual harm — the harm that remains after the defined forward mitigation — and Irreversible-class and over-threshold actions MUST produce the highest-criticality audit records the substrate supports. The gravity of the record tracks the gravity of the consequence, so that the evidence an institution holds for its most consequential actions is its most durable and most tamper-resistant.

Orthogonality to ZTG-3 reversal_strategy

ZTG-5 harm class is orthogonal to the ZTG-3 reversal_strategy, and both fields MUST be present in decision provenance. The reversal_strategy describes what the system can undo or compensate; the harm class describes the consequence in the world. They are independent: a surface may have a capable reversal strategy and an Irreversible harm class, and when it does, the action is gated by its harm class and not rescued by its reversal strategy. Recording both, side by side, is what lets an auditor confirm that the reversal strategy was not quietly used to discount the harm class — the same non-relaxation discipline ZTG-3 states from its side.

Conformance Criteria

A conforming implementation can: assign one of the three harm classes to every governed action by the restorability of its harm, independent of available remediation; attach harm-class declarations to ZTG-3 surfaces with monotonic sub-surface tightening; carry a policy-assigned liability ceiling into the ZTG-4 evidence record; compute composite assessment as a recursive product in discrete deterministic bands that replay identically; promote to an architecturally distinct high-end gate on either Irreversible class or over-threshold assessment; record residual harm for Mitigable actions and highest-criticality evidence for Irreversible and over-threshold actions; and carry both harm class and reversal_strategy in provenance without the latter relaxing the former.

Further Considerations

The legal grounding: reparable and irreparable harm. The distinction ZTG-5 draws is one equity worked out over centuries. Courts of equity issue an injunction — they stop an act before it happens — precisely when a legal remedy after the fact would be inadequate because the harm would be irreparable: a remedy in money cannot restore what would be lost. The whole doctrine turns on the same line ZTG-5 draws, that the availability of compensation does not make a harm reparable, because reparability is about whether the prior state can be restored, not whether a payment can be offered for its loss. ZTG-5's three classes are this doctrine made operational: Restorable harm is harm a legal remedy can actually undo, Irreversible harm is the equitable category of the irreparable, and Mitigable sits between. That the architecture routes Irreversible-class actions to a distinct, more stringent gate is the mechanical analogue of equity intervening before the act rather than compensating after it — because for irreparable harm, after is too late.

The actuarial substrate property. The banded, ceilinged, compositional structure is what makes governed action priceable. Insurance and actuarial practice require exposure to be expressible as discrete, bounded, composable quantities; a risk that cannot be banded and bounded cannot be underwritten. By requiring discrete bands, policy-assigned ceilings, and recursive compositional multipliers, ZTG-5 produces a substrate over which exposure can be assessed actuarially — not because the framework prices risk (it explicitly does not assign the values), but because it produces the structural form priced risk must take. An institution can attach its own actuarial values to ZTG-5's bands and ceilings; the framework guarantees the bands and ceilings exist, are discrete, and compose deterministically.

Classification authority. Harm classification is a governed declaration, not a system judgment. The system does not decide that its own act is Restorable; the harm class is declared on the surface by policy and attached to the action at routing. This keeps ZTG-5 on the right side of the recurring line — the governed subject does not author the inputs that govern it — and it locates the authority for classification with the ratifying principal whose policy declares the surfaces.

Composite irreversibility. Irreversibility is not only a property of single actions. A sequence of individually-Restorable actions can compose into an outcome that is, in aggregate, irreversible, and the recursive faster-than-linear growth of composite assessment is the mechanism that surfaces this. The framework does not assume that safety at each step implies safety of the sequence; the composition algebra is precisely the refusal of that assumption.

Neutrality of composite records. The composite assessment record is neutral evidence: it records what the assessment was, under what multipliers and bands, without itself asserting that the action was good or bad. This neutrality is what lets the same record serve audit, actuarial pricing, and institutional review without being tuned to any one of them.

Iterative calibration of multipliers. The multipliers and band thresholds are expected to be calibrated iteratively against operational experience. This is a policy activity under governance (ZTG-0e), not a system self-tuning loop: changing a multiplier is a governance-state transition, authorized and recorded, never a parameter the system adjusts for itself.

Cross-surface composition. When an action sequence crosses surfaces, the composite assessment composes their multipliers, so exposure accumulated across heterogeneous channels is captured rather than reset at each surface boundary. This is what prevents laundering exposure by routing a sequence through many low-multiplier surfaces.

Stasis interaction. The asymmetry between wrongly holding and wrongly proceeding tracks harm class, and this is where ZTG-5 meets ZTG-2. Wrongly holding a Restorable-class capability is recoverable; wrongly proceeding on an Irreversible-class action under unestablished guarantees is not. The willingness to enter Stasis, and the threshold for it, should reflect this asymmetry rather than be uniform across harm classes. The precise coupling — whether Stasis sensitivity is harm-class-aware — is a composition concern owned jointly with ZTG-2 (see Draft Flags).

Convergence test discipline. A harm classification is admissible only if it can be checked, and the convergence test is that check: independent assessments of an action's harm class should converge, and a classification that cannot be independently reproduced is not yet admissible. This is the discipline that keeps classification honest — it is why classification is banded and recorded rather than asserted, and it is the part of ZTG-5 that most directly serves the admissibility question.

Institutional binding through policy ratification. Harm classes, multipliers, and ceilings are all policy-set and ratified, which is what binds the institution to its own consequence model. The ratifying principal's acceptance of a surface's harm class and ceiling is an attested acceptance of responsibility for actions through that surface — the chapter's connection back to the §1.0 thesis that every authorization traces to a human-attested acceptance of responsibility.

Adoption maturity. Operating ZTG-5 well — calibrated multipliers, well-chosen ceilings, surfaces classified to reflect real consequence — is a maturity an institution grows into, not a switch it flips. The framework does not soften the conformance requirement to fit institutions not yet ready to operate it; an institution early in adoption runs ZTG-5 with conservative defaults and matures its calibration, rather than running a relaxed ZTG-5.

How We Do It

Constable implements ZTG-5 as harm-class and ceiling declarations on its surface registry, a composite-analysis subsystem that bands assessment deterministically, and architecturally distinct gates for high-consequence actions.

Surface and sub-surface registration. Constable registers harm-class defaults and compositional multipliers on surfaces and sub-surfaces in the same ZTG-3 surface registry, carried in the ZTG-0e governance bundle. Default classifications ship as starting points, not normative classifications: they are conservative defaults a deploying institution is expected to review and tighten for its context, not framework assertions about what a given surface's harm class is.

Liability ceiling computation. Ceilings are computed from deployment-level policy and attached to each authorization request, then written into the ZTG-4 evidence record. Constable originates no ceiling itself; the values come from the ratifying principal's policy. Harm-class assignment and ceiling assignment are recorded as HARM_CLASS_ASSIGNED and LIABILITY_CEILING_ASSIGNED events under ZTG-0a, so both are observable and replayable as of decision-time.

Composite-analysis subsystem. A composite-analysis subsystem computes assessment as the recursive product of component assessments and surface multipliers, in discrete bands, so the result replays identically through the ZTG-0b harness. No continuous-valued computation enters the governance-determining path.

Irreversible-class gates. Constable routes Irreversible-class and over-threshold actions to architecturally distinct high-end gates — separate gate structures, not the ordinary gate with extra checks. HumanSeal presents high-consequence authorizations to operators with the harm class, ceiling, and composite assessment surfaced, so a human engaging an Irreversible-class action sees its gravity rather than an undifferentiated approval prompt.

Critical-class evidence binding. Irreversible-class and over-threshold actions are written as critical-class records on the Monotonic Logger, with the highest durability and integrity the substrate supports, bound write-ahead of the effect per ZTG-4. Mitigable-class actions record their residual harm.

Operator support and certification. Constable provides operator support for calibrating multipliers and ceilings against experience, as a governed (ZTG-0e) activity, and forthcoming certification covers whether a deployment's ZTG-5 configuration meets conformance. Conformance tests cover harm-class assignment, monotonic sub-surface tightening, banded deterministic assessment and its replay, distinct-gate promotion on either sufficient condition, ceiling presence in evidence, and reversal_strategy non-relaxation. The protocol is documented in the conformance verification specification referenced in §22.

Draft Flags

• Reconcile with any canonical §14. This full chapter was authored from the workspace calibration summary; if a canonical full §14 exists in the prior SDS/ZTG writings, reconcile the two and treat the canonical one as authoritative on any normative difference. The calibration summary (section-14-irreversibility-of-harm-ztg-5-calibration.md) is retained unchanged as the reference skeleton.
• ZTG-2 × ZTG-5 harm-class coupling. Whether Stasis sensitivity is harm-class-aware is owned jointly with ZTG-2; §11 raises the same flag. Resolve as a single joint decision, not separately in each chapter.
• Convergence test is referenced, not fully specified here. ZTG-5 names the convergence test as the admissibility check on classification; the test's full specification (independent-assessment method, convergence criteria) is larger than ZTG-5 and is flagged for the admissibility/§22 treatment.
• "Governed action" vs "governed effect" harmonized within this chapter — classification attaches to actions, harm-class declarations to surfaces (effects). Confirm this usage matches the final convention chosen in the cohesiveness pass (register item 9).
• §22 Conformance Verification is referenced but not yet captured in this workspace.
• The chapter is substantively complete but likely needs final tone calibration for publication register.