Validation and Quality
Validation and Quality
Fund Analyst Intelligence is designed to be reliable in production.
Reliability comes from validation before narrative.
Quality is treated as an operational metric, not a marketing claim.
This page explains how validation works, how quality is measured, and how human review fits into the production model.
Objectives
A production validation system must ensure:
- structured fund fields remain consistent over time
- contradictions are detected and surfaced explicitly
- missing or stale evidence becomes an exception
- material changes are prioritised and reviewable
- outputs remain comparable across funds and cycles
Validation philosophy
Deterministic gates first
Validation is rule-driven where possible.
The system should not “reason” about facts that can be checked.
Narrative is produced only after factual gates pass.
Human review is part of quality
Regulated workflows require accountability.
Review is not a failure of automation.
It is a control that makes automation safe.
Fail explicitly rather than guess
If evidence is missing or conflicting, the system should surface it.
It should not fabricate a clean output.
Exceptions are a feature, not an inconvenience.
Validation layers
Fund Analyst Intelligence validates across multiple layers.
Layer 1 — Completeness
Ensures required fields are present for the chosen scope.
Examples
- identifiers and domicile
- strategy and mandate fields
- fees and liquidity terms
- key people and operational providers
Outcome
- completeness score per fund
- missing-field exceptions
Layer 2 — Consistency
Checks logical coherence across fields and sources.
Examples
- redemption frequency consistent with notice periods
- fee descriptions consistent across documents
- share class terms aligned with fee schedules
- strategy claims consistent across quarters
Outcome
- contradiction exceptions
- “requires reviewer decision” flags for high-severity conflicts
Layer 3 — Freshness and recency
Ensures evidence meets policy expectations.
Examples
- factsheet older than allowed threshold
- DDQ out of date for required fields
- online source updated but internal record stale
Outcome
- stale-evidence exceptions
- follow-up prompts and escalation rules
Layer 4 — Change detection integrity
Ensures deltas are meaningful and correctly classified.
Examples
- identify true changes versus rewordings
- isolate material changes from minor wording differences
- detect hidden changes in operational terms
Outcome
- structured change log with category labels
- materiality scoring inputs
Layer 5 — Template and output validity
Ensures reports conform to a stable structure.
Examples
- required report sections present
- change summary included
- evidence links present for key statements
- approval stamp present before publication
Outcome
- report readiness checks
- publication gates
Quality signals
Quality in Fund Analyst Intelligence is monitored continuously.
Per-cycle quality signals
- completeness score for required fields
- number of contradictions detected
- proportion of key claims with evidence links
- count and severity of exceptions
- review time and time-to-close for follow-ups
Portfolio-level quality signals
- exception trends by category over time
- recurring issues by manager or strategy
- funds with repeated stale evidence
- funds with frequent material term changes
- adherence to cycle cadence and SLAs
These signals turn quality into a steering tool.
They also reduce dependence on individual memory and effort.
Managing exceptions
Exceptions are the mechanism that keeps quality honest.
They convert uncertainty into action.
Exception lifecycle
- detected by validation or change logic
- triaged and prioritised by severity and materiality
- reviewed and resolved, or marked as monitoring
- recorded with decision notes and ownership
- closed or escalated by SLA policy
A system without exception discipline does not scale.
Fund Analyst Intelligence is built around exception discipline.
Human review controls
Review is implemented as explicit workflow states.
Typical controls
- reviewer assignment and ownership
- decision states: accept, edit, reject, follow-up, monitor
- rationale capture for manual overrides
- approval gate before publication
Manual overrides are recorded.
They are not silent.
This protects governance and auditability.
Minimising variability in narrative outputs
Narrative quality is maintained by constraints.
- templates enforce stable structure
- narrative is restricted to validated facts and resolved exceptions
- change summaries are generated from structured deltas
- confidence and evidence requirements prevent unsupported statements
This reduces subjective variance across funds and cycles.
It also improves comparability for committees.
Continuous improvement loop
Production quality improves through iteration.
A typical improvement loop is:
- observe recurring exception patterns
- refine validation rules or source requirements
- tune materiality thresholds by category
- update templates for clarity and stability
- measure impact in the next cycle
This creates compounding operational value.
It converts experience into system capability.
Definition of done
Validation and quality are production-grade when:
- required fields meet completeness thresholds
- contradictions and gaps are surfaced as exceptions
- stale evidence is detected by policy
- outputs pass readiness checks before publication
- review decisions and overrides are recorded
- quality KPIs are tracked per cycle and portfolio-wide
This is how Fund Analyst Intelligence keeps outputs reliable.