AI decision audit trail
Chen Li-Wen — Golden Harvest Foods Co Ltd
Mapped to Taiwan FSC's "Guidelines for the Use of AI in the Financial Industry" (June 20, 2024), four-stage AI lifecycle: planning & design; data collection; model building & validation; deployment & monitoring. Generated 2026-10-01 21:44 UTC from case data current as of that time.
Not a compliance determination
FSC guidance requires firms using third-party generative AI to designate personnel who "objectively and professionally manage" the risk and to keep records of AI model design and decision-making to assess fairness. This document is that record for this case — it is evidence to support a human reviewer's decision, not a substitute for one.
Stage 1
Overview
This case draws on AI assistance in two places: identifying the beneficial owners of the subject entity and how they're connected, and assessing the evidence behind the client's source-of-wealth claims. In both cases, the AI does not make the final call — it produces a confidence assessment and a written explanation, and a human reviewer decides what happens next. Every AI response is captured in a fixed, consistent format specifically so it can be logged and reviewed here, rather than left as free-form text. AI model used: Claude, developed by Anthropic.
Stage 2
Data collection
Every document submitted for this case is kept exactly as received and is the only basis for any AI-derived finding that follows from it. 0 document(s) submitted to date.
No documents submitted yet beyond the seeded onboarding record.
Stage 3
Model building & validation
This system does not train or customize its own AI model — it uses a general-purpose model provided by a third party (Anthropic). Taiwan's FSC guidance is explicit that firms in this position must still actively manage that risk rather than accept the AI's output uncritically. To do that, every AI output passes through the checks below before anything is added to the case record:
A new name is only linked to an existing person or company when the AI positively confirms it's the same one.
Otherwise it's treated as new and kept separate. This is what lets 'Lin Wei-Chen' — a different ordering of the same name — correctly match the existing 'Wei-Chen L.' record instead of being logged as a different person.
A known ownership percentage can never be silently erased by a later, less specific document.
If a new document doesn't state a specific percentage for a relationship that's already on file, the existing figure is kept rather than cleared. This protects a confirmed stake from being lost just because a later document mentioned the relationship without repeating the number.
Indirect ownership through a chain of companies is calculated independently, not estimated by the AI.
The AI is only asked to report direct relationships a document actually states — for example, 'Company A owns 70% of Company B.' Any indirect stake reachable through a chain of companies is then calculated separately by the system itself, so the same ownership can't be counted twice.
Every confidence score comes with a written reason, and the system leans cautious.
The AI is never allowed to return a bare number — each score must include a short explanation. It's also instructed to weigh the two directions differently: evidence that contradicts a claim should lower confidence sharply, while evidence that supports it should raise confidence more gradually. Missing a real problem is treated as more costly than being overly cautious.
National ID numbers and dates of birth are never sent to the AI model.
When a submitted document contains one, it's compared locally, within our own systems, against the record already on file — the AI is only ever told whether that comparison matched, not the number itself. This keeps the raw identifier from ever reaching the third-party AI provider, while still letting the system use a positive match as supporting evidence when resolving a name variant.
Stage 4
Deployment & monitoring
Beneficial ownership — current state
Company Act — 10% threshold
0% confirmed
AML Act — 25% threshold
0% potential
| Relationship | Status | Rationale | Recorded |
|---|
How the ownership structure was uncovered
Source-of-wealth claims — how confidence changed over time
Dividend income from family manufacturing business, 2016–2022
NT$28,000,000 claimed
| When | Confidence | Status | Note |
|---|---|---|---|
| 2026-08-12 06:24 | 0% | pending | No evidence submitted yet. |
Sale of Taichung commercial property, 2020
NT$18,000,000 claimed
| When | Confidence | Status | Note |
|---|---|---|---|
| 2026-08-12 06:24 | 0% | pending | No evidence submitted yet. |
Aggregate check — claimed sources vs actual assets
Claimed sources total NT$46,000,000 against total assets of NT$60,000,000. A gap of NT$14,000,000 has no claimed source, independent of any single claim's status.