model-research/findings/2026-05-10-64-regulatory-implementation-gap-analysis.md

Finding #64: Regulatory Implementation Gap Analysis

Date: 2026-05-10
Task: Identify regulatory implementation gaps in a FINRA PDT rule implementation document
Document: gargoyle's pdt-rule.md (96 lines) — Pattern Day Trader detection logic

Summary

GPT-5 excels at exhaustive regulatory cross-referencing; Opus identifies operational compliance gaps; Sonnet captures core compliance issues quickly.

Results

Model	Time	Output tokens	Reasoning tokens	Gaps found	Critical	High	Medium	Low
GPT-5	143s	2,989	9,024	20	2	8	9	1
Claude Opus 4	43s	1,353	(internal)	12	2	5	4	1
Claude Sonnet 4	16s	1,222	(internal)	9	1	4	4	0

Common Ground (all 3 identified)

1. PDT designation persistence/lifecycle (CRITICAL) — document doesn't address how PDT designation persists once triggered, or the 90-day restriction lifecycle
2. Equity calculation definition (HIGH) — "$25,000" threshold used without specifying valuation methodology
3. Customer notification/disclosure requirements (HIGH) — FINRA Rule 2270 day-trading disclosure requirements not addressed
4. Day trading buying power calculations (HIGH) — PDT accounts have 4:1 intraday buying power requirements not covered
5. Audit trail insufficiency (HIGH) — events lack sufficient detail for regulatory examinations

GPT-5 Unique Findings

Most significant:

1. Buy-to-cover shorts are day trades (CRITICAL) — document claims "sell actions only" but FINRA Rule 4210(f)(8)(B)(ii) explicitly includes "sale and purchase" (short sale then buy-to-cover) as day trades. This is a fundamental regulatory misread.

2. 6% test omission (MEDIUM) — FINRA requires ">6% of total trades" condition alongside 4+ day trades; document applies raw count only

3. User_id vs account_id conflation (HIGH) — PDT is per-account but implementation uses per-user language

Additional unique findings: options/short sales scope, lot-based counting overcounting, trade corrections handling, portfolio margin accounts, cross-reg margin calls, after-hours trade-date alignment, non-securities exclusion, options exercise/assignment, complex events, "reasonable belief" mechanism, terminology drift

Claude Opus Unique Findings

1. Continuous equity monitoring (HIGH) — only checks equity at decision time, but FINRA requires maintaining $25k at all times during trading day for PDT accounts

2. Pre-existing position handling (HIGH) — unclear if selling position opened 6+ days ago, then rebuying and selling same day, counts as day trade

3. Cross-broker PDT status (MEDIUM) — doesn't consider PDT designation from other brokers

Additional: cash account settlement violations, options exercise creates day trades, cash-secured puts/covered calls exemption

Claude Sonnet Unique Findings

1. Professional vs non-professional accounts (MEDIUM) — blanket PDT application without considering account type exemptions

2. Day trading margin call mechanics (HIGH) — no mechanism for issuing and tracking margin calls

Key Insight

Regulatory gap analysis requires domain expertise encoding. GPT-5's reasoning tokens enabled systematic rule-by-rule comparison with specific regulatory citations (4210(f)(8)(B)(ii), Rule 2270, Reg T). Opus found OPERATIONAL compliance gaps — things the system needs to DO at runtime. Sonnet provided fast screening for core issues.

Model Strengths

• GPT-5: Exhaustive regulatory cross-referencing. Best for compliance audits needing rule-by-rule coverage.
• Opus: Operational compliance gaps where implementation needs runtime behaviors. Best for "what the system needs to DO" gaps.
• Sonnet: Fast screening for core compliance issues. Good initial pass; catches major exemption categories.

Practical Implication

For regulatory compliance analysis of financial software:

1. Run GPT-5 first for exhaustive regulatory mapping
2. Then Opus to identify operational compliance requirements
3. Sonnet for fast initial screening

Union of findings (~28 distinct gaps) provides comprehensive compliance coverage.

Efficiency

• GPT-5: 150 tokens/gap, 143s (~7.2s/gap)
• Opus: 113 tokens/gap, 43s (~3.6s/gap)
• Sonnet: 136 tokens/gap, 16s (~1.8s/gap)