Non-Monotonic Eligibility: re-evaluating governed decisions when evidence arrives late

Theme Name

Non-Monotonic Eligibility — governed decisions that must be re-evaluated when evidence arrives after the decision

Proposer

Zhao Xinghua / Steven Zhao·China (zhaoxinghua09-cell). Brands SynomosAI / MedXpert are unregistered names only (no entity or trademark filed).

Summary

Most themes before the group concern a decision made at a point in time: a connection is admitted (#29), a privilege is granted and its lifecycle states are managed (#23), a runtime state is attested (#22), a legal binding is fixed (#14). But the evidence such decisions rest on decays, and new evidence arrives afterwards — through complaints, incidents, audits, and monitoring. Audited against the open theme titles as of 2026-10-01, no theme treats evidence-driven re-evaluation of a past governed decision as its object: #1 reconstructs what happened, #23 moves privileges through task-driven state transitions, and #29 asks what happens when connection terms change — the evidence side of "a decision may need to be un-made, or re-made, on information that did not exist when it was made" has no home.

Regulated industries already run on this. A medical device is re-evaluated for the lifetime of its authorization because complaint and field-safety evidence keeps arriving after market approval (vigilance and post-market surveillance obligations under the EU MDR, medical device reporting under FDA 21 CFR Part 803, and China's NMPA medical device GMP 2025 revision, effective November 2026, extending quality-management obligations across the product lifecycle). Aviation airworthiness directives, safety recalls, and periodic KYC review follow the same shape. The consistent lesson from that practice: eligibility is not a one-way gate but a non-monotonic property — late-arriving evidence can restrict it, and can also restore it. Frameworks that model only revocation model half the lifecycle.

This theme proposes that evidence-driven re-evaluation of past governed decisions be treated as a first-class object of standardisation: what makes a decision record re-evaluation-ready, which evidence has standing to trigger re-evaluation, and how restriction and restoration are both carried as legible state rather than silent reversal.

Scope / Open Questions

• Decision provenance for re-evaluation. A record that supports later re-evaluation must carry the inputs the original decision was based on, not only its outcome. #1 asks what must be recorded so an action can be reconstructed; this theme asks what must be recorded so the decision behind the action can be re-made soundly. Without the authored reference, re-evaluation degenerates into re-judgement from memory.
• Standing of evidence. Which evidence types (complaints, incidents, audit findings, monitoring signals, expert attestation) are sufficient to trigger re-evaluation, who may submit them, and how the quality and uncertainty of evidence travels with it.
• Non-monotonicity: both directions. Paths for tightening (suspend, restrict, revoke) and for loosening (restore, reinstate) under the same record discipline. Restoration is as consequential as revocation, yet in the current theme set restoration appears only as task-driven state transition (#23 re-validation upon activation), not as an evidence-driven outcome.
• Separation of evidence from evaluation. The same evidence stream should be evaluable under different re-evaluation policies — jurisdictions, sectors, risk appetites differ. What needs standardising is the evidence/decision interface, not any particular evaluator's logic (pluggable evaluator slots).
• Currency and decay. What freshness a prior decision can be relied on for, and how a decision marks itself as stale without thereby being reversed. At the connection level, #29 already asks what freshness a counterparty can rely on; this theme asks the same question for governed decisions generally, and what record properties let a decision be marked stale without being reversed.
• Cross-jurisdiction re-evaluation. One evidence event (a field incident) may trigger different re-evaluation obligations in different jurisdictions; how records stay interoperable when the obligations attached to them differ.

Non-goals

• Not an action or enforcement theme. This theme stops at the evidence semantics and decision-record properties that make re-evaluation sound; who may act on a re-evaluation outcome belongs to the oversight and ecosystem themes (#13, #16).
• Not a liability or adjudication framework.
• Not the internal logic of any particular evaluator, scoring model, or risk policy.

Related Work

• Recent AI-governance work converges on the same temporal gap from the deployment side — continuous admissibility and reassessment obligations for delegated machine authority (Hossain 2026), epistemic hardening of prior safety judgments (distributed corrigibility, 2026), and control-decay revalidation models (Tan 2026) — and FDA's PCCP mechanism pre-authorises bounded change against a fixed record. None of these standardises the cross-domain record properties that make a past governed decision re-evaluable; that gap is this theme's object. Cited as adjacent work, not as prior occupation.
• Medical device post-market surveillance and vigilance — EU MDR vigilance (Articles 87–92) and post-market surveillance (Articles 83–86), FDA medical device reporting (21 CFR Part 803), and China's NMPA medical device GMP (2025 revision, effective November 2026, extending quality-management obligations across the product lifecycle): the most mature existing practice of complaint-initiated, evidence-driven re-evaluation of a granted authorization, across at least three jurisdictions with differing obligations over the same evidence.
• Safety recall practice and aviation airworthiness directives as cross-sector analogues of mandated re-evaluation on late evidence.
• The proposer's supporting artifacts, offered as prior art for anyone who wants runnable or citable material (harmonize away; improvements welcome):
• lgd-theory — lifecycle governance doctrine built on three laws (registry / evidence / gates); text all rights reserved, citation with attribution permitted.
• lgd-glossary — CC BY 4.0 candidate definitions for the relevant vocabulary, including the entry non-monotonic eligibility re-evaluation (silent-failure families, custody-without-judgment, authorship record, definition provenance), offered for terminology coordination in #32.
• lgd-medai-demo — runnable medical-device lifecycle demo (single command, zero dependencies, Python 3.10+) in which a post-deployment drift signal re-opens a prior eligibility decision (DENIED → GATE_PASSED → PAUSE → ROLLBACK → REVERIFY → RESUMED); verified locally to reach final_state=RESUMED (code: Apache-2.0).

Related Themes

• #29 (the Should Problem): its "what happens when it changes — expiry, amendment or withdrawal while the connection is live" is the connection-level slice of this problem; this theme generalises to any governed decision and adds the evidence side that triggers change.
• #23 (privilege lifecycle): models privilege state transitions driven by task state and environment change at activation; this theme supplies the evidence-driven trigger and the decision-record properties those transitions consume.
• #1 (accountability after an agent has acted): both are consumers of decision provenance — #1 reconstructs the act, this theme re-makes the decision.
• #21 (population-scale evaluation): population-level evidence (counts, rates) is a candidate trigger source feeding individual re-evaluation.
• #5 / #9 (provenance of authority; never crisply authored): re-evaluation evaluates against the authored reference; where no crisp authorship exists, #9's null semantics applies and re-evaluation has nothing to evaluate against.

Discussion

Open. If an existing theme already owns part of this scope, consolidation is welcome — contribution, not priority.