Skip to content

Instantly share code, notes, and snippets.

@XavierGeerinck
Created May 9, 2026 18:57
Show Gist options
  • Select an option

  • Save XavierGeerinck/2c7365a867e6ff81ebcdd2e4a033f212 to your computer and use it in GitHub Desktop.

Select an option

Save XavierGeerinck/2c7365a867e6ff81ebcdd2e4a033f212 to your computer and use it in GitHub Desktop.
Vanta AIMS Improvement Policy upload test (will delete after verification)

AIMS Improvement Policy

Status: Draft — log structure defined; initial entries seeded from closed/in-progress EG-### gaps Frameworks: ISO 42001 (Clause 10, A.10.4), ISO 27001:2022 (Clause 10), NIST AI RMF (GOVERN-1.5) Review cycle: Quarterly Owner: Compliance lead (log maintenance + quarterly review) · Platform engineering (gap closure evidence) Vanta controls: iso42001 CPL-68, CPL-60 · nistAiRmf GOV-3100, GOV-3107 Vanta test: approved-aims-improvement-policy-exists (NEEDS_ATTENTION)

Purpose

This log closes the ISO 42001 PDCA loop (Plan → Do → Check → Act) by recording every improvement to Scrydon's AI systems — whether driven by incident review, gap closure, customer feedback, monitoring trends, or regulatory change. It is the primary evidence artifact for:

  • Vanta test approved-aims-improvement-policy-exists (currently NEEDS_ATTENTION)
  • ISO 42001 Clause 10 (Improvement) — requires documented continual improvement actions
  • NIST AI RMF GOVERN-1.5 — AI risk governance includes mechanisms for organizational learning

Improvements recorded here are the output of the PDCA cycle. Inputs come from the source channels below.

Scope

Applies to all AI systems in the AI System Inventory and the AI governance controls documented in the compliance framework.


Improvement Source Channels

Source Cadence Input to Log
Incident review Per incident Post-incident corrective actions
Customer feedback Ongoing; reviewed quarterly Feature and safety improvement requests
Internal engineering review Per PR / per sprint Engineering gap closures (EG-### IDs)
AIUC-1 quarterly delta Quarterly New AIUC-1 requirements from changelog
Post-Deployment Monitoring trends Weekly (automated) + quarterly (manual review) Drift events, threshold breaches
External security / pen-test findings Annually + per engagement Security remediation actions
Compliance framework updates Per framework publication Framework alignment gaps

Improvement Record Template

Each entry in the log follows this structure:

Field Description
ID Unique improvement ID (format: AI-IMP-YYYY-NNN)
Source Source channel (e.g., EG-019 gap closure, monitoring trend, customer feedback)
AI System(s) Affected system(s) from inventory
Description What was improved and why
Hypothesis Expected outcome after improvement
Metric How improvement will be measured
Owner Engineering or compliance owner
Target Date Planned completion date
Status Open / In Progress / Closed
Evidence PR number, ADR, test results, Vanta evidence link
Closed Date Date closed (when Status = Closed)

Improvement Log

Closed / Substantially Complete


AI-IMP-2026-001 — DB TLS Enforcement

Field Value
Source EG-002 (Sensitive columns in plain text — DB TLS gap)
AI System(s) All (auth DB, agentic DB, analytics DB, ontology DB)
Description Enforced TLS on all Postgres connections in production via shared resolveSslMode() helper; in-cluster Postgres now terminates TLS via self-signed cert
Hypothesis Eliminates risk of credential/token interception over plaintext DB connections
Metric ssl=on confirmed in Postgres server logs; NODE_ENV=productionssl: 'require' behavior verified
Owner Platform engineering
Target Date 2026-04-16
Status Closed
Evidence packages/common/src/node/db-tls.ts; helm/scrydon/templates/infra/db.yaml; helm/scrydon/templates/infra/db-tls-secret.yaml
Closed Date 2026-04-16

AI-IMP-2026-002 — Default-On Inline DLP (EG-019)

Field Value
Source EG-019 (No mandatory inline LLM content DLP — critical)
AI System(s) Workflow Engine, Copilot, Agents, Embedding Pipeline
Description Implemented withDlp capability-layer decorator — default-on PII detection / content scanning / classifier signals on every LLM/STT/TTS/embedding call. Fail-closed on missing scanContext. Streaming scanned per flush window. 13 dlp.* audit events with strict Zod schemas.
Hypothesis Every LLM call is DLP-scanned; PII leakage via unguarded egress is eliminated as a default failure mode
Metric 100% of LLM calls wrapped by withDlp; 0 calls that bypass decorator (verified in PR review); audit events emitted per scan
Owner Platform engineering
Target Date 2026-04-21
Status Substantially closed (residuals: Guardrails v2 migration + EG-012 prompt-injection validator)
Evidence packages/sdk-authoring/src/integrations/dlp/with-dlp.ts; ADR 2026-04-17; PR #767
Closed Date 2026-04-21

AI-IMP-2026-003 — SIEM Streaming Infrastructure (EG-005)

Field Value
Source EG-005 (No SIEM streaming for on-premises monitoring)
AI System(s) All (audit log system)
Description Implemented HMAC-signed webhook subscriptions per org (at-least-once delivery + DLQ), NDJSON streaming export (date-range, paginated), per-org retention with nightly chunked DELETE. Dapr cron components in production Helm.
Hypothesis Customers can forward audit events to their SIEM in near-real-time with cryptographic authenticity
Metric Webhook subscription delivers events within 10 min of emission; DLQ depth = 0 under normal load
Owner Platform engineering
Target Date 2026-05-08
Status Substantially closed (customer runbook polish pending)
Evidence packages/better-auth-audit-log/src/routes/audit-log-subscriptions.ts; helm/scrydon/templates/auth/dapr-cron-audit-log.yaml; security/siem-forwarding
Closed Date 2026-05-08

AI-IMP-2026-004 — Classification Enforcement Foundation (EG-020)

Field Value
Source EG-020 (Bell–LaPadula classification enforcement)
AI System(s) RAG / Knowledge Base, Agents, Embedding Pipeline
Description Shipped classification schema (migrations 0024–0026 + 0047–0048), BLP aggregation helper, clearance_egress rule, user_clearance table with four-eyes overrides, classification library, DLP extensions, retrieval clearance filter, and ClassificationPanel admin UI.
Hypothesis Documents above user clearance are invisible at retrieval; tool dispatch to over-clearance endpoints is blocked; spillage events are detected and attributed
Metric Integration test: cross-clearance query returns 0 rows; security.clearance_egress_blocked emitted on violation
Owner Platform engineering
Target Date 2026-05-06
Status Foundation closed (integration follow-up: wire loadEffectiveUserClearance into all search() callers; per-document classification UI; chunk-level propagation)
Evidence packages/db-agentic/schema.ts (migrations); packages/better-auth-integrations/src/classification/; ADR 2026-04-24
Closed Date 2026-05-06

In Progress


AI-IMP-2026-005 — AI Quality Metric Persistence (EG-008)

Field Value
Source EG-008 (AI performance monitoring not persisted)
AI System(s) Workflow Engine, Guardrails, Copilot
Description Add guardrails_results table to agentic DB; add OTel span attributes for hallucination score and PII entity counts; create AI governance Grafana dashboard
Hypothesis Hallucination rates, PII trigger rates, and guardrails effectiveness become measurable as time-series trends
Metric Hallucination score captured in DB for > 95% of Guardrails-enabled workflow runs; Grafana dashboard shows 7-day rolling hallucination rate
Owner Platform engineering
Target Date Q3 2026
Status Open
Evidence EG-008; apps/agentic/lib/core/telemetry.ts (existing OTel instrumentation)
Closed Date

AI-IMP-2026-006 — Prompt Injection Detection (EG-012)

Field Value
Source EG-012 (No prompt injection detection — Vanta risk R-19)
AI System(s) Workflow Engine, Copilot, Agents
Description Implement pattern-based prompt-injection validator in Guardrails engine; add to adversarial CI test suite; sanitize user-provided systemPrompt in workflow blocks
Hypothesis Common injection patterns (role-play override, instruction injection, indirect RAG injection) are detected and blocked at the withDlp boundary
Metric > 95% detection rate on curated injection corpus; 0 jailbreaks on test set
Owner Platform engineering
Target Date Q3 2026
Status Open
Evidence EG-012; packages/sdk-authoring/src/integrations/templates/scrydon/guardrails/validators/
Closed Date

AI-IMP-2026-007 — PII Test Coverage Expansion (EG-015)

Field Value
Source EG-015 (PII detection test coverage gaps)
AI System(s) withDlp, Guardrails
Description Add test cases for all 27 entity types (currently ~6 tested); create benchmark dataset per region (US, UK, EU, APAC); measure and document precision/recall per entity type
Hypothesis PII detection accuracy is measurable and consistently ≥ 90% recall across all entity types
Metric 27/27 entity types have ≥ 3 test cases; precision ≥ 95%, recall ≥ 90% per entity (where detection method allows)
Owner Platform engineering
Target Date Q3 2026
Status Open
Evidence EG-015; packages/sdk-authoring/src/integrations/templates/scrydon/guardrails/__tests__/validators.test.ts:84-266
Closed Date

AI-IMP-2026-008 — Agent Autonomy Framework (EG-016)

Field Value
Source EG-016 (Incomplete agent autonomy framework)
AI System(s) Agents, Copilot
Description Add workspace-level autonomy policy; classify tools by risk level; persist approval decisions to audit log; add batch cancellation endpoint
Hypothesis Admins can enforce tool approval requirements org-wide; approval decisions are auditable; emergency stop available for batch operations
Metric Workspace-level policy enforced for all tool calls in affected orgs; approval decisions in auth.audit_log; batch cancellation API functional
Owner Platform engineering
Target Date TBD
Status Open
Evidence EG-016; apps/agentic/app/api/copilot/confirm/route.ts
Closed Date

Quarterly Review Process

Who Reviews

Role Responsibility
Compliance lead Chairs review; updates Vanta improvement evidence; signs off on closure criteria
Engineering lead Reports on open improvements; provides evidence for closed items
Security lead Reviews security-relevant improvements; confirms adversarial eval results

Review Inputs

  1. Export of all open improvement entries from this log.
  2. Post-Deployment Monitoring trend report for the quarter.
  3. AIUC-1 changelog delta (new requirements from quarterly AIUC-1 refresh).
  4. Vanta test status for approved-aims-improvement-policy-exists and related tests.

Review Outputs

  1. Updated status for each improvement entry.
  2. New entries for improvements identified during the review.
  3. Evidence package uploaded to Vanta for closed items.
  4. Meeting minutes stored in Vanta under the AIMS improvement evidence folder.

Cadence

Activity Timing
Quarterly review meeting First week of each quarter
Vanta evidence upload Within 5 business days of review meeting
AIUC-1 changelog review Within 2 weeks of each quarterly AIUC-1 publication

Closure Criteria

An improvement entry is closed when all of the following are true:

  1. The described change is merged to main (or the relevant production branch).
  2. Automated tests cover the changed behavior and pass in CI.
  3. The corresponding EG-### entry in gaps is updated to reflect the new status.
  4. Evidence is uploaded to Vanta (PR link, test results, or ADR).
  5. The compliance lead has reviewed and confirmed the closure.

Partial closures (e.g., foundation shipped, integration follow-up pending) are recorded as Substantially closed with an explicit list of residual items.


Gaps

Issue Description Impact
No formal AIUC-1 quarterly delta review process AIUC-1 refreshes quarterly; no process exists to triage new requirements New AIUC-1 requirements may not be captured in this log
Improvement meeting minutes not yet stored in Vanta Quarterly reviews not yet started Vanta test approved-aims-improvement-policy-exists NEEDS_ATTENTION
EG-021 No quarterly external eval to feed improvement inputs Improvement loop lacks external adversarial test data

Actions Required

  • Schedule first quarterly improvement review (target: Q3 2026, first week of July)
  • Upload this document and initial log entries to Vanta as evidence for approved-aims-improvement-policy-exists
  • Establish AIUC-1 quarterly delta review process (assign owner, add to engineering calendar)
  • Close AI-IMP-2026-001 through AI-IMP-2026-004 evidence in Vanta with PR/ADR references
  • Add improvement review meeting minutes template to Vanta evidence folder
  • Link quarterly review outputs to Post-Deployment Monitoring and Value Maintenance frameworks
  • Define target date for AI-IMP-2026-008 (EG-016) based on engineering roadmap
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment