Real-time monitoring of chatbots and agents for AI compliance and governance

Standards & Compliance

Standards & Compliance Mapping

Every AI governance instrument names an obligation — record-keeping, robustness, post-market monitoring, human oversight — but none names a metric. PSA is the behavioral evidence layer: the deterministic, timestamped, externally-verifiable measurement that discharges the measurable half of those obligations. This page maps PSA, honestly, onto twelve frameworks in force in 2026.

12

frameworks mapped

6

evidence primitives

5

strong-coverage frameworks

5

languages

Thesis Evidence Primitives Crosswalk Where PSA Stops

Section 01

Thesis

Read the AI governance instruments of 2026 side by side and the same shape appears every time. Each tells an organization what it must achieve — keep records, be robust to attack, monitor the system after deployment, keep a human in the loop, manage risk. Not one tells you how to measure whether you did it. They are, by design, technology-agnostic: they name the obligation and leave the metric to you.

That gap is the opportunity. PSA is the behavioral evidence layer — the instrument that turns "we monitor for drift" into a deterministic, timestamped number with a defined formula, and "we log relevant events" into a hash-chained record you can verify without trusting us. PSA does not replace a management system or a governance function. It supplies the proof under the promise.

PSA public spec — GitHub

Field Guide — every PSA signal

Section 02

The Six Evidence Primitives

Across all twelve frameworks, PSA's contribution reduces to six evidence primitives. Each crosswalk row maps a requirement to one or more of these.

ID Primitive PSA signals

E1Deterministic behavioral event logPosture codes (I/P/M/H/G) + alert ladder

E2Tamper-evident log integrity, externally verifiableSIGTRACK — hash-chained, drand-anchored, /verify-chain

E3Adversarial / robustness measurementC0 input intent (I0–I9), C1 adversarial stress, CPI

E4Human–AI interaction risk (incl. psychological harm)DRM (IRS, RAS, RAG), HRI

E5Continuous monitoring & forecastingBHS, POI, DPI, PE, CPF3 (EWMA+HMM)

E6Behavioral transparency / explainabilityNamed posture codes + named, auditable alert reasons

Section 03

Framework → PSA Crosswalk

Requirement-by-requirement mapping. Filter by framework or by coverage level. Coverage is marked honestly — green only where PSA produces exactly the evidence asked for.

DIRECT PSA produces the evidence, deterministically. PARTIAL PSA supplies a measurable input to an otherwise procedural requirement. OUT Structurally outside PSA (procedural, or protected-attribute fairness).

Requirement PSA mapping Coverage

ISO/IEC 42001:2023 — AI management system (the certifiable anchor — see the dedicated mapping)

A.6.2.6/.8 · A.5 · C.2.8–11Evidence layer: operation logs, impact, robustness/transparency/safety → BHS/POI/DRM/SIGTRACK/CPF3DIRECT

EU AI Act — Regulation (EU) 2024/1689

Art. 12 — Record-keepingAutomatic event logging over lifetime → E1 posture log + E2 SIGTRACK tamper-evident trailDIRECT

Art. 15 — Accuracy, robustness, cybersecurityResilience to adversarial inputs → E3 C0/C1/CPI runtime measurementDIRECT

Art. 72 — Post-market monitoringContinuous documented monitoring → E5 BHS/POI longitudinal + CPF3 forecastDIRECT

Art. 13 — Transparency to deployersInterpretable operation → E6 named posture codes + named alert reasonsPARTIAL

Art. 14 — Human oversightEnable intervention → E4 DRM/IRS real-time risk surfacing + alert ladderPARTIAL

Art. 9 — Risk management systemIterative risk evaluation → E4/E5 runtime signals feed the processPARTIAL

Art. 55 — GPAI systemic riskModel eval, adversarial testing, incident tracking → E3 war-zone probes + E1 incident loggingPARTIAL

Art. 10/11/17 — Data governance, technical docs, QMSProcedural / organizationalOUT

NIST AI RMF 1.0 (2023) + Generative AI Profile (NIST-AI-600-1, 2024)

MEASURE 2.x — TEVV & monitoringPSA's home function — E1–E6 across the boardDIRECT

MEASURE 2.6 / 2.7 — Safety; security & resilienceE3 adversarial + E4 safety riskDIRECT

MEASURE 2.8 / 2.9 — Transparency, accountability, explainabilityE6 posture codes + E2 SIGTRACKDIRECT

MEASURE 2.3 / 2.5 / 2.13 — Eval, validity, ongoing monitoringE5 BHS/POI/CPF3DIRECT

MANAGE 4.x — Monitoring & incident responseE5 + alert ladder feed the functionPARTIAL

MAP 1–5 — Context establishmentE4 DRM domain targeting (legal/health/finance)PARTIAL

GOVERNOrganizational culture, policy, accountability structuresOUT

MEASURE 2.11 — Fairness & biasProtected attributes never ingestedOUT

ISO/IEC 23894:2023 — AI risk management (guidance)

Risk identification & analysisE4 DRM runtime risk + E1 posture evidencePARTIAL

Risk monitoring & reviewE5 BHS/POI/CPF3PARTIAL

Risk treatment & governance integrationProceduralOUT

ISO/IEC 42005 — AI system impact assessment

Evidence of actual behavioral impactsE4 DRM (IRS, RAS) runtime — incl. psychological harmPARTIAL

Documentation & sign-off of the assessmentProceduralOUT

ISO/IEC TR 24028 (trustworthiness) & TR 24027 (bias)

TR 24028 — robustness & transparency aspectsE3 adversarial + E6 transparencyPARTIAL

TR 24027 — bias in AI systemsProtected attributes never ingestedOUT

OECD AI Principles (2019, updated 2024)

1.4 — Robustness, security & safetyE3 C1/C5 + E5 CPF3DIRECT

1.3 — Transparency & explainabilityE6 posture codesPARTIAL

1.5 — AccountabilityE2 SIGTRACK verifiable trailPARTIAL

1.2 — Human-centred values & fairnessBias subset out of scopeOUT

Council of Europe — Framework Convention on AI (2024)

Documentation, traceability, oversightE1/E2 + E4PARTIAL

Rights-based governance, redressTreaty-level / proceduralOUT

US Colorado AI Act — SB 24-205 (effective 2026)

Risk-management policy & programmeE5 monitoring evidencePARTIAL

Impact assessment for consequential decisionsE4 behavioral evidencePARTIAL

Duty to avoid algorithmic discriminationBias on protected attributes — out of scopeOUT

Singapore — Model AI Governance Framework / AI Verify

Robustness & behavioral safety testingE3 C-classifiers + war-zone probesDIRECT

Operations management & monitoringE5 BHS/CPF3PARTIAL

AI Verify fairness testingBias subset out of scopeOUT

MITRE ATLAS — adversarial ML threat knowledge base

Prompt injection, jailbreak, evasion at runtimeE3 C0 input intent (I1–I9), C1 adversarial stress, CPI, semantic-drift detectionDIRECT

Model/data poisoning, supply-chain, weight exfiltrationOutside the text-behavioral surfaceOUT

SOC 2 / ISO/IEC 27001 — security & audit controls

Audit-log integrity, tamper-evidence, monitoringE2 SIGTRACK — verifiable without trusting the issuerPARTIAL

Full ISMS (access control, change mgmt, …)Procedural / infrastructuralOUT

Sectoral instruments

GDPR Art. 22 — Automated decision-making safeguardsE4 DRM evidence + E6 named reasonsPARTIAL

HIPAA — health conversational-AI safetyE4 IRS/DRM crisis detection (safety layer, not a HIPAA control)PARTIAL

SR 11-7 — model risk management (finance)E5 CPF3 + behavioral drift + benchmark (ongoing monitoring + effective challenge)PARTIAL

Section 04

Where PSA Stops — Said Plainly

The honest half of the story is the part PSA does not cover, and it is the same boundary in every framework. PSA reads what a model does, from its output text, from the outside. It therefore says nothing about the procedural and organizational half of governance — leadership, policy, human resources, data governance, third-party management, conformity assessment. Those are real obligations; they are simply not measurements.

And PSA is deliberately silent on bias and fairness over protected attributes. PSA never ingests demographics — it has no race, gender, or age field to discriminate on. That makes it structurally non-discriminatory, but it also means PSA cannot evidence the fairness duties at the centre of NIST MEASURE 2.11, ISO/IEC TR 24027, or Colorado's anti-discrimination core. We do not claim that ground; we name it as out of scope on every row.

The result is a clean division of labour. The framework is the certifiable anchor and the organizational programme. PSA is the telemetry and the evidence store underneath it — covering the measurable half, and pointing honestly at the half it does not touch.