H.1 Review method
This is a targeted primary-source scoping review with foundational references, not an exhaustive systematic review or a statistical meta-analysis. Searches covered agent safety, monitoring, intervention timing, harness evolution, prompt injection, multi-agent behavior, fairness, finance evaluation, identity, protocol changes, and regulatory status. Discovery results were traced to papers, official specifications, regulatory text, agency pages, or first-party incident reports before supporting report claims.
The control reference review inspected the published overview, complete control index, and relevant evaluation, runtime, monitoring, multi-agent, third-party, regulatory, fairness, and implementation sections. It did not revalidate every citation in that source. The handbook includes original explanations, worked examples, a glossary, baseline practices, and reference worksheets. Each external source retains its stated evidence and verification limits.
Quantitative exhibits use reported values or transparent calculations. The underlying model/agent experiments were not rerun. Paper versions, study units, denominators, selection conditions, and interpretation limits are retained in the publication kit. No cross-benchmark averages or model safety rankings were constructed.
H.2 Evidence classes
| Class | Meaning | Permitted use in this report |
|---|---|---|
| Baseline | Current TOL source inspected for structure and control meaning. | Maintain historical traceability and distinguish prior context. |
| Empirical research | Benchmark or controlled study, generally a preprint unless independently confirmed otherwise. | Describe the demonstrated mechanism and reported results within the evaluated setting. |
| Incident report | First-party account of observed operational events. | State the observation, configuration, clustering, and investigation limits. |
| Simulation / concept | Constructed scenario, formal model, architecture, or research agenda. | Motivate falsifiable tests and proposed designs; do not claim production effectiveness. |
| Official legal / supervisory | Governing regulation or supervisory text inspected, or an identified official implementation source. | Attribute the specific verified obligation/status; keep scope and source limits visible. |
| Technical documentation | Released protocol, implementation documentation, or development-stage conventions. | Identify versioned implementation requirements and documented limitations. |
| Proposed practice | Original enterprise recommendation in this handbook. | Support a reviewable implementation decision; do not present it as law or a research finding. |
Evidence classes help the reader assess what a source establishes. A new preprint is not established consensus.
H.3 Validation and limitations
- Original anchors were checked against the inspected 112-identifier index. The 45 H-practices and twelve E-specifications use separate namespaces; their described mechanisms are proposed enterprise designs.
- Composition percentages and benchmark percentage-point differences were recalculated from recorded values. The zero-event bound was computed from the displayed equation, with rounding confined to presentation.
- Critical distinctions were checked explicitly: permission versus identity; detection versus prevention; statistical discrimination versus operating-threshold recall; early interruption versus harm; live incident versus simulation; process assurance versus behavioral effectiveness; final regulation versus proposed rule.
- The EU legal-text access limitation is recorded in chapter 16. The relevant timeline is attributed to the official Commission page. A source titled Safety in Self-Evolving Agents was excluded as a recency anchor because inspected submission metadata conflicted with its identifier/month; its date was not silently repaired.
- This report makes no claim that public benchmark results transfer to a particular institution. Local deployment effectiveness, reviewer competence, data representativeness, and legal applicability remain local acceptance questions.
H.4 Research worth commissioning next
| Open question | Decision it could change | Falsifiable enterprise study |
|---|---|---|
| Can a monitor maintain useful prevention under unseen schemas and adaptive attacks? | Permitted autonomous action classes and monitor reliance. | Locked schema/domain holdouts, bounded adaptive attack search, paired utility/prevention outcomes. |
| How much safety state must persist across runs? | Ledger design, retention, and operational cost. | Distributed evidence cases, long pauses, compaction, restarts, and controlled state-ablation comparisons. |
| Does memory repair remain effective after continued adaptation? | Whether reusable memory or self-updating skills can be enabled. | Track descendants, revoke a source, resume adaptation, and test recurrence after restoration. |
| Which component interactions deserve joint tests? | Re-evaluation cost and release speed. | Prioritize shared authority/state paths; compare targeted coverage with a broader sampled interaction suite. |
| Does a multi-agent topology add safe business value? | Whether to retain or simplify the architecture. | Matched workflows with simpler baseline, identical authority, and the same outcome oracles. |
| Does meaningful review survive workload and automation bias? | Review staffing, training, and permitted consequence level. | Blinded seeded cases, normal and peak loads, evidence availability, overrides, and time-to-correction. |
| Where does customer exclusion first appear? | Policy, data, routing, and remediation priorities. | Aligned stage cohorts, missingness controls, matched cases, and alternative-design experiments. |
H.5 Editorial and artifact validation scope
Version 1.0.1 checks completed 3 October 2026. Counts were reconciled across the manuscript, web pages, PDF text and implementation kit: 18 chapters, 11 appendices, 14 figures, 45 proposed practices, 12 specifications, 36 proposed acceptance tests, ten templates, 47 sources and 17 mapped instruments. Practice links and the separate X01–X30 regulatory explanation identifiers were checked, along with download hashes and unchanged first-edition files.
The handbook landing page, regulatory map, two chapters, regulatory appendix, changelog and errata were checked at 375-pixel and desktop widths using axe-core 4.13.0 against WCAG 2.2 AA rules. No automated violations were reported on those pages. Keyboard checks covered the skip link, reading action, filters, expandable explanations, practice links and a horizontally scrollable comparison table. All 36 handbook routes were checked for unique titles and descriptions, canonical URLs, sharing metadata and structured data. Automated checks and these keyboard checks do not establish complete WCAG conformance or replace testing with assistive-technology users.
The tagged PDF has 179 bookmarks and text alternatives for all 14 figures. Its 3,463 table cells were checked against extracted text, with no missing cells or text beyond page bounds; representative portrait and landscape pages were visually reviewed. The PDF/UA-1 validator still identified unmarked content; a screen-reader review of reading order has not been completed. PDF/UA conformance is not claimed. Completing that validation requires correctly marking decorative content without losing semantic tags, then repeating the validator and reading-order checks with assistive technology.
The dated citation check fetched 98 distinct URLs. Ninety-six responded successfully; the two ISO catalogue pages blocked automated access. All 13 arXiv sources matched their cited titles, versions and version dates. No citation mismatch was found. The ISO mappings retain their public-scope limitation; neither clause-level inspection nor conformity is asserted. LinkedIn's sharing inspector, Google's external rich-results test, published model experiments and enterprise integrations were not rerun. A recorded check supports only its stated scope.