IOSCO Supervisory Recordkeeping & Disclosure — ponens Policy Pack

This pack maps the IOSCO Supervisory Toolkit for AI Use in Capital Markets (FR/02/2026, May 2026) onto computable ponens policies. Where the FIX Agentic AI Runtime Governance pack is preventive — a runtime traffic-light that gates execution — this pack is evidentiary: it checks that an AI system’s decisions and outputs leave the audit trail and disclosures a supervisor expects to review.

Source: IOSCO FR/02/2026, Supervisory Toolkit for AI Use in Capital Markets, Chapter 3 — Tables 5 (Disclosure) and 6 (Recordkeeping & reporting).

Why this maps onto ponens

The IOSCO toolkit is supervisor-facing: for each area it lists potential concerns, example questions, and “Supporting Evidence for Review”. That evidence column — “auditable recordkeeping for AI-driven decisions”, “traceability between AI outputs and final actions taken”, “evidence of human oversight or intervention”, “incident logs with root-cause and remediation”, “records evidencing how clients are informed when AI is used” — is exactly what a ponens trace plus this pack produces. ponens makes the evidence computable rather than a manual document review.

IOSCO toolkit	ponens
The records / logs / disclosures a firm must evidence	the trace
A “Supporting Evidence for Review” item	a policy (temporal formula)
Supervisor’s pass/concern judgement	verdict pass / warning-fail / error-fail
Risk-based, proportionate expectations	pack tier tags

Relationship to the FIX pack

The two are designed to be used together:

	FIX AI WG pack	IOSCO pack (this)
Nature	Preventive runtime governance	Evidentiary recordkeeping & disclosure
Output	Green/Amber/Red `GovernanceState`	the audit trail a supervisor inspects
Emphasis	identity, intent, capability gating	logging, traceability, explainability, disclosure

Trace model

Governance facts appear in the trace as:

Action types (added for this pack): Decision, Output, Incident (alongside existing Release, Deploy, Draft, …).
Per-action predicates the firm’s recordkeeping layer emits, matched against the action text: logged, model_version_recorded, inventory_recorded, explainability_recorded, material_impact, human_reviewed, root_cause_recorded, remediation_recorded, reportable, supervisor_notified, client_facing, ai_disclosed, marked_ai_generated, ai_claim, claim_substantiated, ai_affects_client_rights, client_consent.

Worked traces: examples/iosco_recordkeeping/governed.json (all 11 Green) and violating.json (8 Red + 2 Amber). Run ponens trace check <file>.

The pack

error severity ⇒ Red; warning severity ⇒ Amber.

Recordkeeping & Audit Trail (`auditability`) — IOSCO Table 6

Policy	Formula	RAG
`ai_outcomes_logged`	`G((Decision ∨ Output) → logged ∧ model_version_recorded)`	R
`output_action_traceable`	`G((Release ∨ Deploy) → P(Output ∧ logged))`	R
`decision_logic_recorded`	`G(Output → explainability_recorded)`	R
`human_oversight_evidenced`	`G(material_impact → human_reviewed)`	R
`incidents_root_caused`	`G(Incident → root_cause_recorded ∧ remediation_recorded)`	R
`reportable_incident_notified`	`G(Incident ∧ reportable → F(supervisor_notified))`	R
`ai_inventory_recorded`	`G((Decision ∨ Output) → inventory_recorded)`	A

Disclosure & Transparency (`communication`) — IOSCO Table 5

Policy	Formula	RAG
`ai_use_disclosed`	`G(Output ∧ client_facing → ai_disclosed)`	R
`ai_content_marked`	`G(Output ∧ client_facing → marked_ai_generated)`	A
`no_ai_washing`	`G(ai_claim → claim_substantiated)`	R
`consent_for_ai_decisions`	`G(ai_affects_client_rights → P(client_consent))`	A

Aggregation

ponens trace check aggregates the pack: any error fail ⇒ Red (non-zero exit); else any warning fail ⇒ Amber; else Green. On the worked violating trace this yields Red with 8 error and 2 warning findings — the precise set of records/disclosures a supervisor would flag as missing.

Out of scope

The toolkit’s supervisor-side layers — Monitoring indicators (Table 7), third-party/outsourcing concentration, and firm-level governance structures (board oversight, training) — are organisational/population-level and not per-trace checks, so they are not expressed here.