Evidence Reference
RateLoop records a bounded human-review run as an integrity-bearing packet. The packet shows which review policy operated, what the panel returned, how coverage was measured, and which evidence references were available.
Who provides what
These references come from the EU AI Act (Regulation (EU) 2024/1689), as amended by Regulation (EU) 2026/1744.
Article 14 addresses the design of human oversight for high-risk AI systems. Article 26 sets out duties for deployers, the organizations using those systems. Article 4 covers AI literacy.
For example, Art 14(4)(a) means Article 14, paragraph 4, point (a): understanding the system and monitoring its operation. Each reference below links to an official EU source.
| Requirement | RateLoop provides | You remain responsible for |
|---|---|---|
| Art 14(4)(a) · Monitor Understand the system's capabilities and limits, and detect and address unexpected behavior. | An oversight dashboard (sampling, latency, disagreement, blocked outputs), in-app, email, and browser alerts, event webhooks, and per-agent evidence summaries. | Watching those surfaces, understanding the agent's capacities and limitations, and acting on anomalies. |
| Art 14(4)(b) · Automation bias Recognize and question over-reliance on AI outputs. | Blinded reviewer judgments, decision prompts with no preselected choice, and available disagreement and calibration signals. The displayed override rate is a workspace-wide 30-day aggregate. | Staying aware of the pull to over-rely on the system and keeping each decision a considered one. |
| Art 14(4)(c) · Interpret Make sense of the output using available interpretation tools. | An owner case view with the actual output, source context, reviewer rationales, and surfaced disagreement before the decision. | Correctly interpreting the output within your domain, workflow, and context. |
| Art 14(4)(d) · Override Choose whether to use the system or disregard, override, or reverse its output. | Recorded go, revise, and stop decisions, plus per-output override records with required reasons on the audit chain. | Deciding when to disregard, override, or reverse an output. |
| Art 14(4)(e) · Stop Intervene or interrupt operation so the system can reach a safe state. | Advisory Codex, plugin, and MCP integrations record lifecycle state but do not verify withholding. Workspace stop blocks new RateLoop release authorizations; your workflow controls when outputs are delivered. | Choosing which outputs are gated, when to intervene, and when to halt. |
| Art 26(2) · Assignment Assign oversight to people with the competence, training, authority, and support for the role. | Oversight designations with attestation records — competence basis, training completed, authority granted — exportable as an assignment record. | Choosing those natural persons and ensuring they are competent, trained, and authorized. |
| Art 26(5) · Monitoring Follow the instructions for use, monitor operation, and respond to risks and serious incidents. | The same monitoring surfaces, coverage exports, and alerting, retained as workspace evidence. | Monitoring operation against your instructions for use and pausing use when risks appear. |
| Art 26(6) · Log retention Keep automatically generated logs under your control for an appropriate period of at least six months, unless other applicable law specifies otherwise. | A six-month floor for RateLoop review records, with retention export, scheduled enforcement, and legal hold, that may support your AI-system log-retention process. | Retaining logs generated by your high-risk AI system and under your control for the required period. RateLoop's floor does not satisfy that duty. |
| Art 4 · AI literacy Support staff and others using AI on your behalf in developing their AI knowledge and skills. | Exportable training and calibration records for reviewers and oversight persons. | Taking measures to support the development of AI literacy among relevant staff. No specific level must be guaranteed for any individual. |
Advisory integrations do not control output delivery; your workflow controls when an output is delivered. Model identity is labelled host-reported because RateLoop does not independently verify which model produced an output. RateLoop does not claim SOC 2 / ISO / HIPAA / residency attestations it does not hold. See Human Oversight for capability details.
What an evidence packet contains
Current schema: rateloop.human-assurance.evidence.v10
01
Frozen scope
Run, suite, audience-policy, admission-policy, and manifest versions are included with their exact hashes.
02
Forecast terms
The packet records whether the frozen policy makes forecasts off, optional, or required. Forecasts never change the reported verdict.
03
Review context
The reviewed agent and its version, the selection trigger, delivery authority, gate type, reviewer-qualification counts, review period, coverage, and response latency describe how the check operated.
04
Judgment evidence
Privacy-safe result counts, disagreement, failure-tag counts, rationale digests, domain-separated Merkle roots, and recomputation inputs bind the derived result to the stored judgments.
05
Available references and limits
Policy, model, source, and version references sit beside explicit limitations and suppression rules.
Reviewer identities and raw or decryptable rationales are excluded. Small cells are suppressed according to the frozen minimum aggregation size. Host-reported execution metadata remains marked as not independently verified.
Complete synthetic packet example
This synthetic packet uses the v9 invited-review format, which the verifier still accepts, and a separate example signing key. Verify it offline with the downloadable public key; it contains no customer data.
The synthetic candidate says: “Enable the new response policy for a 10% rollout; the supplied seven-day error report shows no failed requests.” Three invited reviewers compare that answer with the baseline and the supplied synthetic evidence. Two prefer the candidate and one prefers the baseline, so the frozen rule passes. The owner records Go in a separate packet-bound decision after reading the result. The downloadable packet preserves the review result, its boundaries, and the inputs needed to verify it; it does not contain or sign the separate owner decision.
How to check the evidence
Use the public browser verifier to check an exported packet without uploading it, or use the command-line verifier below.
- Try the synthetic packet. Download the single-file verifier — no account, no dependencies, no network — and run it against the published packet and its separate synthetic key pin:
This pin is only the trust anchor for the synthetic example. It is not a RateLoop production signing key and does not appear in the hosted verification-key endpoint.node verify-rateloop-evidence.mjs synthetic-evidence-v9.json \ --public-key synthetic-evidence-v9.spki.txt \ --key-id ed25519:8b42b24feb46dfab8c59a1c0 - Verify your export. Obtain the expected key ID and public-key pin from the cacheable, unauthenticated public verification-key endpoint. A reader who already has a packet needs no RateLoop account to obtain the matching current or retired pin. Workspace members can also download it from the key history in Results. Never use the public key embedded in the packet as its own trust anchor. Run the same command with your packet, pin, and key ID.
- Merkle roots and recomputation. The same command checks canonical packet hashing, the case and response roots, privacy-safe aggregation, and the frozen pass rule. Treat any reported error as a failed check.
- External witness receipts. Every completed decision-packet attestation includes a Rekor bundle. Select the intended log independently and check its UUID, index, authenticated checkpoint, inclusion data, and signed entry time. A missing or invalid bundle makes the witness incomplete. When the separate TSA field is non-null, validate its RFC 3161 message imprint, certificate path, policy, and time against trust roots selected by your organization. An absent token means there is no TSA receipt.
Download a completed witness from its public attestation URL. Select the signer, Rekor log key, and TSA certificate chain independently, and pin the witness signer key ID. The attestation and audit-chain checkers are not public downloads yet. Obtain an independent check before relying on either witness; the downloadable packet verifier checks the packet itself, not these external receipts. Pin the expected audit head from another trusted record when possible.
Export surfaces
GET /api/account/workspaces/{workspaceId}/assurance/runs/{runId}/evidence
GET /api/account/workspaces/{workspaceId}/review-results/{opportunityId}/evidence[?format=pdf]
POST /api/account/workspaces/{workspaceId}/oversight/evidence-bundle
GET /api/account/workspaces/{workspaceId}/assurance/coverage/export
GET /api/account/workspaces/{workspaceId}/audit/export
GET /api/evidence/trusted-keys
GET /api/account/workspaces/{workspaceId}/assurance/trusted-keys
GET /api/account/workspaces/{workspaceId}/assurance/trusted-keys?format=spki&keyId=ed25519:…
GET /api/account/workspaces/{workspaceId}/assurance/runs/{runId}/cases?format=pdf
GET /api/account/workspaces/{workspaceId}/governance/works-council-brief?format=markdown
GET /api/public/assurance/attestations/{jobId}The workspace routes require an authorized workspace session. The public witness is digest-only and is retrievable only by its opaque job ID. The case route returns JSON by default and a case-packet PDF with format=pdf; the works-council brief returns the German rendering as a download with format=markdown. A workspace retention policy cannot be configured below six months, but that product floor does not determine the customer's legal or contractual retention schedule. Scheduled enforcement removes due private artifact content and access logs unless a legal hold applies. Artifact digests, signed packets, witness and WORM receipts, and the canonical audit chain remain as integrity records; deleting private content does not rewrite that history.
Framework cross-reference
Every row means supports evidence for. It is a cross-reference, not an assessment result, legal opinion, certification, or assertion that a customer's controls are implemented or effective.
| Reference | RateLoop artifacts | Boundary |
|---|---|---|
| ISO/IEC 42001:2023A.6.2.6 | operation-and-monitoring records for a deployed AI system, including the frozen review policy in force, sampling decisions, human results, coverage rollups, and stage transitions for a bounded period.Signed assurance evidence packet; Adaptive assurance coverage export; Host-reported execution evidence. | This mapping does not demonstrate that a customer has implemented the A.6.2.6 control or that RateLoop is ISO/IEC 42001 certified. |
| ISO/IEC 42001:2023A.6.2.8 | event-log evidence about review decisions, policy context, gate state, and integrity-checkable audit history.Workspace audit-chain export; Signed assurance evidence packet; Human-review gate evidence. | This mapping does not establish completeness of a customer's wider AI-system event logging or retention programme. |
| ISO/IEC 42001:2023A.9.2 | responsible-use records showing configured review rules, human judgments, coverage, escalation, and release-gate evidence.Signed assurance evidence packet; Adaptive assurance coverage export; Human-review gate evidence. | This mapping does not assign customer responsibility, competence, training, authority, or approval obligations to RateLoop. |
| NIST AI Risk Management FrameworkMEASURE 2.8 | documented examination of transparency and accountability limits, including privacy-safe aggregation, the stated limitations of each packet, and execution metadata recorded as host-reported and not independently verified.Signed assurance evidence packet; Host-reported execution evidence. | This mapping is an evidence cross-reference and does not represent a NIST assessment, endorsement, or determination of risk acceptability. |
| NIST AI Risk Management FrameworkMEASURE 4.2 | documented measurement results informed by invited domain-expert reviewers, covering sampling, human agreement and disagreement, latency, and design-weighted population estimates or typed coverage gaps.Adaptive assurance coverage export; Signed assurance evidence packet. | This mapping is an evidence cross-reference and does not represent a NIST assessment or endorsement, and an invitation to review is not proof of a reviewer's expertise or independence. |
| NIST AI Risk Management FrameworkMANAGE 2.4 | records binding a configured review requirement to the stop-gate state and the exact release decision for a single reviewed output.Human-review gate evidence; Workspace audit-chain export. | RateLoop records a hold-or-release decision for a single reviewed output and does not supersede, disengage, or deactivate the customer's AI system. External agent integrations are advisory; RateLoop Chat checks the current review policy and required human result before sending email. |
| NIST AI Risk Management FrameworkMANAGE 4.1 | post-deployment monitoring records covering human override of an AI result, returning a changed scope to its saved starting coverage, and integrity-checkable follow-up history.Human-review gate evidence; Workspace audit-chain export; Signed assurance evidence packet. | This mapping does not make or replace the customer's risk treatment, acceptance, response, or governance decisions. |
| Regulation (EU) 2024/1689 (EU AI Act)Article 26(5)-(6) | monitoring records, review-policy operation, and export history that may support a deployer's monitoring evidence and retention process for logs generated by its high-risk AI system.Adaptive assurance coverage export; Signed assurance evidence packet; Workspace audit-chain export. | RateLoop is not the deployer's assigned human oversight. Its six-month floor applies to RateLoop review records, not the logs generated by the customer's high-risk AI system, and does not satisfy Article 26(6) or any monitoring, authority, competence, or reporting duty. |
| Regulation (EU) 2024/1689 (EU AI Act)Article 73 | preserving review decisions, gate outcomes, and audit history that may support a customer's serious-incident investigation and report preparation.Signed assurance evidence packet; Human-review gate evidence; Workspace audit-chain export. | RateLoop does not determine whether an event is a serious incident and does not perform the customer's regulatory notification. |
| FINRA supervision rules and generative-AI guidanceRegulatory Notice 24-09 | technology-governance records for a supervisory system that uses generative AI, including the review policy in force, host-reported model and version metadata, and integrity-checkable review history.Signed assurance evidence packet; Host-reported execution evidence; Workspace audit-chain export. | Regulatory Notice 24-09 states that it creates no new requirements and sets no human-review obligation; this mapping does not determine which FINRA rules apply, approve a member firm's supervisory system, or constitute legal advice. |
| FINRA supervision rules and generative-AI guidanceRule 3110 | documented review policies, exceptions, escalations, release decisions, and audit history that may be incorporated into a member firm's supervisory records.Adaptive assurance coverage export; Human-review gate evidence; Workspace audit-chain export. | RateLoop does not establish, maintain, or certify the member firm's supervisory system or written supervisory procedures. |
| FINRA supervision rules and generative-AI guidanceRule 3110.07 | review records that chronicle the reviewing account, the item reviewed, the date of review, and the action taken, in a form a member firm can retain alongside its supervisory records.Workspace audit-chain export; Human-review gate evidence; Signed assurance evidence packet. | Rule 3110.07 governs evidence of review of internal communications and correspondence rather than AI output, and this mapping does not establish, maintain, or certify the member firm's supervisory system or written supervisory procedures. |
| FINRA supervision rules and generative-AI guidance2026 Annual Regulatory Oversight Report, GenAI: Continuing and Emerging Trends | records of human-in-the-loop review of model output, the formal review and approval process in force, and tracking of which model version was used and when.Signed assurance evidence packet; Adaptive assurance coverage export; Host-reported execution evidence. | The report presents these as practices a firm may want to consider rather than as FINRA requirements, and this mapping does not establish that a member firm has adopted them or that its supervisory system is adequate. |
Download the machine-readable OSCAL 1.2.2 component definition, or continue with SDK evidence exports.